## Joel Barmettler - AI Engineer, Researcher, and Entrepreneur Source: https://joelbarmettler.xyz/about/about-me.md ## About me I'm Joel. I research and build AI systems. I've also run a company, which informs how I approach everything else. The distance between what AI can do in a paper and what it can do in production is where I work. Bridging that gap requires research and engineering, and I've ended up doing both. ### Building and shipping At 22, while studying computer science at the University of Zurich, I co-founded PolygonSoftware with two classmates. Over four years we grew from three founders to 18 engineers, delivered 62 projects for 25 clients across five industries, and learned what it takes to keep a software company alive. I ran over 100 sales meetings and 80 pre-sales calls. I conducted 40+ coding interviews and hired 26 people. I made technical decisions that saved us and technical decisions that cost us. The company ended in 2023 through a failed merger. What survived: an instinct for what customers actually need and a respect for the gap between a working demo and a system that runs in production for paying users. → More on PolygonSoftware ### Research I finished my master's at UZH summa cum laude, with a thesis graded 6.0. The work became ConceptFormer, a method for grounding language models in knowledge graphs without converting structured data to text. It won the Best Paper Award at the GLOW workshop at WWW'26. The research question that drives me: how do we make language models trustworthy enough to deploy in contexts where being wrong has consequences? This pulls toward knowledge representation, factual grounding, and understanding where these systems fail. My study on political orientation in 21 language models tested them against Swiss voting recommendations. The instinct was the same: if we're going to deploy these systems, we should understand what they've absorbed and what they'll say when asked. → Research ### AI infrastructure From 2023 to 2026 I worked as Senior AI Consultant at bbv Software Services, where I architected the AI-Hub platform: the infrastructure layer that lets enterprise clients deploy language models with retrieval, tool use, and observability. Twelve clients, fifteen projects. This is not wrapper code around OpenAI. The stack includes vector databases (Milvus), data orchestration (Dagster), message streaming (NATS), and observability throughout (OpenTelemetry). Building systems at this layer taught me where LLM applications actually break. It's rarely the model. In July 2026 I founded Souverana, my own consultancy for secure and sovereign AI architectures, where I advise Swiss SMEs and international Fortune 500 companies on AI architecture, AI infrastructure, integration, and MCP. → Souverana · More on bbv ### Communicating I host Attention Heads, a German-language podcast on AI with 27 episodes covering technical deep-dives, industry analysis, and the social implications of these systems. I've given five conference talks, including panels at the Swiss AI Impact Forum, and hosted six webinars on topics from ChatGPT internals to AI agents in enterprise. Technical people who can explain are useful. I've found that forcing myself to communicate research and engineering clearly to non-specialists sharpens my own understanding. → Podcast · Appearances ### What I'm working on Through Souverana I work on AI systems where research quality and production reliability both matter, helping Swiss SMEs and international Fortune 500 companies architect secure, sovereign AI that they own and control. I'm drawn to problems that take the gap between papers and deployment seriously, and I stay close to research collaboration wherever it sharpens the work. For the raw data behind any of this, see Skills. ## What Drives Me - Research Focus and Philosophy on AI Systems Source: https://joelbarmettler.xyz/about/what-drives-me.md ## What drives me The problems I keep coming back to share a thread: what does it take to make AI systems reliable enough to deploy, and what happens when they are? ### Knowledge grounding Language models are fluent but factually unstable. They produce plausible text without reliable access to the knowledge that would make it true. My master's thesis addressed this from the structured side: ConceptFormer injects knowledge graph topology directly into the LLM embedding space, bypassing the lossy text serialization that retrieval-augmented generation typically requires. A single concept vector improves factual recall by 213%. Preserving graph structure in latent space outperforms converting it to text and uses 130x fewer tokens. At bbv, I work the unstructured side of the same problem. The Swiss AI Hub processes enterprise documents through retrieval pipelines (chunking, embedding, indexing, retrieval, reranking) to give models access to organizational knowledge they weren't trained on. The engineering is different. The question is the same: how do you give a model access to knowledge it can use faithfully? Structured and unstructured knowledge grounding are usually treated as separate fields. I think they converge. A production system will need graph-native representations for facts with clear relational structure and retrieval pipelines for everything else. Where that boundary sits, and how to move between the two, is what I want to research further. → ConceptFormer paper · Political orientation study ### Sovereignty and efficiency Most enterprise AI deployments depend on a handful of API providers. The model, the infrastructure, and the data pipeline all live behind someone else's authentication wall. For organizations handling sensitive data or operating in regulated industries, this is a strategic vulnerability. The Swiss AI Hub is how I've approached this at bbv. It's a platform built on open-weight models that organizations deploy in their own infrastructure. LiteLLM for model routing, Milvus for vector search, LlamaIndex for retrieval, full observability stack. No SaaS subscriptions. One command provisions the entire system. Sovereignty is only real if you can run capable models on hardware you control. This is why I care about quantization, small specialized models, and parameter-efficient fine-tuning. BitNets reducing weights to single bits. LoRA adapters that fine-tune a fraction of the parameters. GRPO applied to low-rank matrices to train reasoning into compact models. These techniques determine whether "run it yourself" is a realistic option or a talking point. The Swiss AI Hub depends on this work, and the research community keeps delivering. → More on bbv and the Swiss AI Hub ### The agentic layer A language model on its own is a text completion engine. What makes it useful inside an organization is the scaffolding around it: tool access, memory, planning, orchestration, access controls. This is the agentic layer, and it's where theoretical capabilities meet operational reality. Building this layer has been a large part of my work, at bbv and now through Souverana. An agent that can draft a document, query a database, send an email, and decide which of those a task requires faces challenges beyond model intelligence. They're about system reliability. Can it handle authentication? Does it fail gracefully? Can you trace why it made a specific decision? Can you audit it? Tools like Claude Code show what happens when this scaffolding works. A model trained to understand code becomes genuinely productive because someone built the infrastructure that lets it read files, run tests, and iterate on its output. The gap between "the model can do this in principle" and "this works in your organization" is almost entirely an infrastructure and integration problem. That's the problem I find most worth solving. → Podcast on AI agents · Netzwoche article on enterprise agents ### Social impact My study on political orientation in language models tested 21 models against Swiss SmartVote voting recommendations. Most cluster around progressive, socially liberal positions, with agreement rates between 60% and 90% depending on the party. This isn't a normative claim about what models should believe. It's an empirical observation about what they've absorbed from training data, and it matters because these systems increasingly mediate how people access information. The podcast covers adjacent territory. I've recorded episodes on AI in military applications and the base-rate math that makes "90% accurate" target selection catastrophic, on predictive policing feedback loops that encode the biases they claim to eliminate, on what AI does to education when institutions respond with prohibition instead of adaptation, on the labor market effects that follow when knowledge work becomes automatable. If I'm going to spend my career making AI systems more capable and more deployable, I want to understand what that means for the people who live with them. → Political orientation study · Podcast ## Technical Skills and Expertise - AI, ML, Infrastructure, and Web Development Source: https://joelbarmettler.xyz/about/skills.md ## My Skills Eight years of commits across a startup, research, and enterprise AI consulting. The data below covers everything tracked in version control since October 2017. ### AI and machine learning My AI work is not a recent pivot. Data science tooling enters the commit history in 2018, ML frameworks from 2019, and production AI infrastructure from 2024. The stack covers research (PyTorch, Weights & Biases, HuggingFace Transformers, Optuna, SPARQL), production ML (LlamaIndex, Milvus, Cohere), and data science (Jupyter, NumPy, pandas, scikit-learn). I train models, track experiments, build retrieval pipelines, and run inference at scale. I also call APIs when that's the right tool, but that's not where the work ends. ### Beyond the day job I always have something running outside of work. During the Polygon years, that meant university research and open-source contributions alongside a more-than-full-time startup. During the bbv years, it meant ConceptFormer, the political bias study, 27 podcast episodes, and this website. The personal project heatmap has no gaps longer than a few months since 2017. I build things because I'm curious, not because someone is paying me to. #### Personal and university projects ### How I work I write code in focused blocks during business hours, mostly between 9am and 7pm, with a peak around 2pm. Weekends exist in the data but stay quiet. The longest streak is 20 days. I don't romanticize late nights. The more telling metric is the ratio between pull requests opened and pull requests reviewed: 687 opened, 1,027 reviewed. That ratio shifted over time. At PolygonSoftware I wrote most of the code myself. At bbv, my output is primarily reviews, architecture decisions, and issues that define work for others. In 2025 alone: 332 reviews, 202 issues opened, 131 PRs. I focus on making the team more productive rather than shipping individual features. ### Systems and infrastructure A model that works in a notebook is not a product. The gap between the two is infrastructure: APIs, containerization, cloud deployment, persistence, caching, messaging, orchestration, observability. I've built that layer for the Swiss AI Hub and multiple client deployments. Every component independently deployable, observable, and replaceable. No single tool dominates the commit distribution because a production platform can't afford single points of failure in its tooling either. ### Web development 580,000 lines of JavaScript represent the PolygonSoftware years: 62 projects, 25 clients, five industries. Vue was the workhorse, React used where clients required it, the stack evolving as the company scaled. This was full-stack product engineering under time pressure and budget constraints, choosing frameworks based on client needs rather than personal preference. The frontend work continued at bbv through the AI Hub's interface, but at lower volume. Web development went from being the entire job to being one capability among several. ### The career arc The lines-of-code chart compresses the whole story into one image. Java on the left is university. The JavaScript wall in the middle is a startup being built at full speed. The Python curve on the right is what came after: research, AI infrastructure, and a deliberate shift toward the work I want to do for the next decade. ## Souverana - Independent AI Consulting and Architecture, Zurich Source: https://joelbarmettler.xyz/career/souverana.md ## Souverana I consult under the name **Souverana**, independent AI consulting and architecture, based in Zurich. The legal entity behind it is LatentSpace Labs GmbH, which I founded in July 2026 after building the enterprise AI business area and the Swiss AI Hub at bbv Software Services. The premise is short: your data is confidential, so your AI has to be too. Administrations, hospitals, and law firms can run AI on real, confidential data today, in Switzerland and under their own control. The technology is not the bottleneck. What is missing is someone who assembles the available pieces so they hold for a specific organisation's data. ### Three things mistaken for an AI strategy Switching on Microsoft Copilot and hoping the rest follows. Handing developers GitHub Copilot licences and waiting for productivity. Naming someone "AI Engineer" and having them assemble a RAG prototype. If that describes your organisation, you are in good company. A strategy starts somewhere else: where AI creates value in your business, which data it needs to do that, and how it reaches production safely. ### Sovereign AI is four layers It is not a product. It is four layers, and each one has good open options today. - **Strategy.** Where you stand, where AI creates value, and in which order to proceed. Every engagement starts here. - **Infrastructure.** Language model and hosting, vector database, document storage, data and knowledge pipelines. The foundation everything else runs on. - **Platform.** The tools your team touches daily, wired into a coherent whole: Open WebUI, OpenCode, and whatever else fits how you actually work. - **Integration.** Your own agents, RAG, and MCP connections into your systems. This is where AI meets the business. Above the four layers sits the disposition. Organisations that get somewhere with AI try things early and get used to change, and that is part of what I bring to an engagement. ### Why this is buildable now Shadow AI already shows where the demand is. Staff are using ChatGPT with case files and patient data, which means the task is real and the tool works. Those are the use cases waiting for a solution, and the first step is an official route to them. The components exist: open models such as Apertus, hosting in Switzerland or in-house, mature tooling for RAG and MCP. The EU AI Act and the revised Swiss Data Protection Act belong in the build plan from the start, carried along as construction details so the architecture satisfies them by construction. Bolting compliance on at the end is what makes it expensive. Independence is what makes the recommendation worth anything. Hyperscalers, platform vendors, and large consultancies all arrive at their own product. I sell neither a platform nor a model, so the advice can follow your situation instead. ### How an engagement runs You start small and with a clear frame, and each step produces something the next one builds on: an assessment of where you stand and what is exposed, then the sovereign target architecture across all four layers with model, hosting, and platform chosen independently, then the build and the coordination of hosting and operations partners, then ongoing work as your AI architect through operation, rollout, and further development. Most AI consulting ends with a slide deck. That is roughly where the interesting problems start: how documents actually get ingested, what happens when the model is wrong, who may see which retrieval result, how a failure is traced six months later. So the concept is where the work begins. I build, and I stay, including as a fractional AI lead where that fits better than a project. Rates are published rather than quoted on request: fixed prices for the assessment and the target architecture, a day rate for implementation, a monthly retainer for the fractional lead role. The initial conversation and the sovereignty check cost nothing. ### Current mandates Organisations of every size, from sole proprietorships to international corporations. Three mandates are running, spread across the four layers. **Integration.** A Fortune 500 company (name confidential) is making its own software accessible to AI through MCP servers. I am building part of an environment in which models can reach a large body of data, reason over it, and act on it. **Platform and infrastructure.** I continue to co-own the architecture of the Swiss AI Hub, the Apache 2.0 platform I originally built as AI Architect at bbv. A complete AI stack, deployable with one command, air-gapped on a single GPU or in any Swiss cloud. It solves precisely the infrastructure and platform problem described above, and it is the reference implementation for much of what I recommend: clients can read the code. **Strategy.** A Swiss SME (name confidential) on its internal AI strategy, focused on AI in software engineering and on a sovereign AI transformation. **Website:** souverana.ch · **Sovereignty check:** souverana.ch/check ## Building the AI Business Area at bbv Software Services Source: https://joelbarmettler.xyz/career/bbv.md ## Building the AI Business Area at bbv Software Services When bbv Software Services AG decided to expand into AI solutions in 2023, I joined as Senior AI Consultant to architect and build the new business area, a role I held until June 2026 before going independent with Souverana. My background in AI research, startup founding, and technical leadership positioned me to bridge technology and business strategy. ### Swiss AI Hub: The open AI platform The Swiss AI Hub is complete infrastructure for production AI that organizations own and control. Unlike SaaS subscriptions or code libraries, it provides Azure AI-level completeness with full ownership: authentication, monitoring, databases, UIs, everything included. **Now open-source.** As of 2026, the entire platform is released as open-source under the Apache 2.0 license, not just the tools it builds on. The full stack, SDKs, and documentation are public at dev.ai-hub.bbv.ch and docs.ai-hub.bbv.ch. Any organization can deploy, inspect, and extend it without a license fee, and contribute back through an open pull-request and certification process. **Open-source foundation.** Built on best-in-class open-source tools: LiteLLM and vLLM for model serving, Milvus for vector search, LlamaIndex and LangChain for retrieval and agent orchestration, Keycloak for identity, Neo4j for agent memory, and Langfuse with OpenTelemetry for observability. When the ecosystem evolves, clients benefit. No vendor lock-in, no platform constraints. **Strategic independence.** Deploy anywhere: on-premise, Swiss data centers, your cloud. Use local models or cloud APIs. Switch providers without code changes. True sovereignty over data and infrastructure, critical for Swiss organizations with regulatory requirements. **30-minute deployment.** One command deploys everything: LLM gateway, vector search, chat interface, authentication, monitoring. Pre-built agents work immediately. No cloud provisioning, no complex setup, no infrastructure engineering required. **Transparency over black boxes.** Every agent follows explicit workflows. Every decision is traceable and auditable. Every cost is tracked. Built for Swiss standards where understanding why AI gave an answer matters as much as the answer itself. **Collaborative strength.** Swiss organizations share infrastructure while competing on innovation. Contribute improvements, benefit from others' work. Build together what no single organization could afford alone. ### Platform architecture and technical leadership I architected and led a team of 5 engineers in developing the Swiss AI Hub, an enterprise-grade platform for deploying multimodal autonomous agent applications. I designed the core infrastructure using an event-driven microservices architecture with Python, FastAPI, Docker, and Dagster, supporting on-premise and cloud deployments for over 10 enterprise clients. The platform integrates best-in-class open-source tools: LiteLLM for unified model serving across providers, Milvus for vector search and retrieval, and LlamaIndex for document processing and agent orchestration. I engineered the model-serving stack leveraging vLLM, llama.cpp, and LiteLLM to support open-source models (Gemma, Qwen) with quantization for efficient inference on client-side GPUs, with the full stack running air-gapped on a single NVIDIA RTX 6000 Pro. On top of the infrastructure I designed three SDKs (Agent, Pipeline, Process) and a library of pre-built agent types, so developers inherit auth, tracing, streaming UI, and cost tracking out of the box. I drove technical strategy and pre-sales engineering, defining the product roadmap and business model that established a new revenue-generating AI business unit with strategic independence from vendor lock-in. In 2026 I led the platform's release as open-source under Apache 2.0, opening the codebase, SDKs, and documentation so the wider Swiss ecosystem can build on shared sovereign infrastructure rather than rebuilding it from scratch. ### Team building and market positioning I recruited and developed AI talent combining technical depth with business thinking. Through webinars, presentations, and pre-sales engagements, I positioned bbv as an AI solutions partner. I worked closely with CEO Philipp Kronenberg, AI Business Area Lead Stefan Haeberling, and former co-lead Alan Ettlin to build market presence. This role combined AI solution development, technical team leadership, and business area strategy. Building a business area from inception let me operate at both technical and strategic levels, applying AI innovation to enterprise contexts, experience I now bring to independent clients through Souverana. ## PolygonSoftware: Building a tech company during university Source: https://joelbarmettler.xyz/career/polygon-software.md ## PolygonSoftware: Building a tech company during university At 22, while studying computer science at the University of Zurich, I co-founded PolygonSoftware with two fellow students. As tech lead and co-founder, I shaped the technological direction while learning entrepreneurship through practice. We used monday.com for workflow management and process optimization. ### Technical implementation and team growth I co-founded the company and scaled the technical team from 3 to 18 engineers, establishing core operational frameworks for agile product development and client delivery. I drove strategic vision across technical roadmapping, financial planning, and talent acquisition. I established software engineering standards implementing best practices for full-stack web development, automated testing, and maintainable systems. I created a culture of mentorship and continuous learning, coaching junior and professional engineers in software architecture to accelerate their growth into senior roles. Key projects included a computer vision package recognition system for Swiss Post, AI-supported solutions for Klapp, and complex data visualizations for Ziegler Consultants. We managed over ten parallel projects successfully. ### Work culture experiments PolygonSoftware served as a testing ground for modern work practices: unlimited vacation, full remote work, and flexible hours during exam periods. These policies, uncommon in Switzerland, succeeded through emphasis on personal responsibility and intrinsic motivation. The company ended in 2023 through a failed merger. The experience demonstrated that technological innovation stems from experimentation and rapid learning rather than extensive experience. The machine learning and computer vision integration approaches developed at PolygonSoftware continue to inform my current AI engineering work. ## Machine learning for semiconductor quality control at BESI Source: https://joelbarmettler.xyz/career/besi.md ## Machine learning for semiconductor quality control at BESI BE Semiconductor Industries (BESI) produces die bonders that place chips with micrometer precision at speeds of several thousand units per hour. As a computer science student at the University of Zurich, I joined BESI's R&D team to integrate machine learning into quality control, developing a computer vision system for automated defect detection in memory chips. ### System development and implementation I initiated and managed the full data science lifecycle to replace a traditional computer vision algorithm with a TensorFlow-based deep learning model. I implemented a CNN architecture achieving 99% accuracy in automated defect detection at production speeds incompatible with manual inspection. The deep learning framework integrated into existing production processes, including parametrization and visualization tools with TensorBoard for production team operation. ### Knowledge transfer and sustainability Through workshops and documentation, I trained the team in system operation and underlying ML concepts. The interactive training materials became the foundation for subsequent ML projects at BESI. The system remains in production use and established the pattern for AI integration in quality control processes. ## Data engineering for cryptocurrency analytics at CoinPaper Source: https://joelbarmettler.xyz/career/coinpaper.md ## Data engineering for cryptocurrency analytics at CoinPaper I architected and deployed ETL pipelines using Python and Spotify Luigi on AWS, processing gigabytes of daily cryptocurrency data for fundamental analysis. CoinPaper was an open-source platform providing transparency in cryptocurrency evaluation through automated data aggregation from GitHub, Reddit, Telegram, and Google Trends. ### Pipeline architecture and data sources I designed a serverless data collection system using Luigi for job orchestration, DynamoDB for storage, and API Gateway with VTL templates for the API layer. The pipeline ran jobs at multiple frequencies: high-frequency jobs (minutes) fetched global market data via Coinpaprika API, hourly jobs updated OHLCV candles, and daily jobs performed GitHub analysis using PyGithub to count lines of code and track commits, Reddit sentiment analysis with PRAW and NLTK's VADER, Telegram metrics via Telethon, and Google Trends data through pytrends. ### Analysis and scoring algorithms I implemented custom analytical engines including linear regression for community growth trends, whitepaper analysis using pdfminer and NLTK for auto-summarization, and Git analytics that cloned repositories to calculate actual code activity rather than relying on stars. The system generated a weighted "Coinpaper Score" combining developer activity, community sentiment, and manual reviews to grade project legitimacy and help users identify scams. The infrastructure processed multiple data streams at scale, transforming raw social and technical metrics into actionable fundamental analysis for cryptocurrency investors. ## Progressive in Principle, Centrist in Practice: LLM Political Bias Is Instrument-Dependent Source: https://joelbarmettler.xyz/research/invisible-coalition-partner.md ## Progressive in Principle, Centrist in Practice: LLM Political Bias Is Instrument-Dependent Prior work has established that instruction-tuned LLMs lean left of center on abstract political questionnaires. I tested whether this holds when models face real policy decisions. Using a dual-instrument methodology grounded in Swiss democracy, I administered the Smartvote questionnaire (75 policy questions, 66 LLMs vs. 184 elected parliamentarians) and 48 federal referenda to 9 flagship models in four national languages. Abstract and concrete instruments tell fundamentally different stories about the same models. ### Dual-instrument design The first instrument, Smartvote, replicates prior work: abstract policy proposals answered on a four-point scale, compared to the positions of 184 Swiss National Council members across six parties spanning the full left-right spectrum. The second instrument is novel: real federal referenda (Volksabstimmungen) with official government summaries, presented as binary Ja/Nein decisions in German, French, Italian, and Romansh under three information conditions. Party recommendations (Parolen) serve as the political benchmark. ### Key findings On Smartvote, all 66 models converge on the same center-left position (Cohen's d = 3.64, p = 0.0002), replicating the established finding. No structural variable (geography, licensing, model generation) predicts positioning. On Volksabstimmungen, the left-to-right agreement gradient flips: models agree most with centrist Die Mitte and FDP, not with SP and Grüne (Wilcoxon p = 0.008). The leftward bias measured on abstract instruments does not generalize to concrete policy decisions. For some models, the language of the question changes the answer more than the political content does. Cross-linguistic consistency ranges from 98% (GPT-5.4) to 50% (Mistral), whose approval rate swings from 17% in German to 82% in Romansh. Two models (Grok, Mistral) show systematic change-aversion, voting Nein on 83-94% of referenda regardless of political direction. **Authors:** Joel P. Barmettler (Independent Researcher, Zurich) **Published at:** arXiv (preprint, 2026) **Read the paper:** arxiv.org/abs/2606.00048 · PDF ### Cite ``` `@misc{barmettler2026progressiveprinciplecentrist, title={Progressive in Principle, Centrist in Practice: LLM Political Bias Is Instrument-Dependent}, author={Joel Barmettler}, year={2026}, eprint={2606.00048}, archivePrefix={arXiv}, primaryClass={cs.CY}, url={https://arxiv.org/abs/2606.00048}, } ` ``` html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);} ## ConceptFormer: Graph-native grounding of LLMs via latent concept injection Source: https://joelbarmettler.xyz/research/masters-thesis.md ## ConceptFormer: Graph-native grounding of LLMs via latent concept injection I developed ConceptFormer, a neuro-symbolic approach grounding large language models in structured knowledge graphs without architectural modifications or textual linearization. Published at the GLOW workshop at WWW'26 — where it received the **Best Paper Award** — the system addresses trustworthiness and factuality challenges in retrieval augmented generation by preserving knowledge graph topology in latent space. ### Architecture and concept vectors ConceptFormer operates in the LLM embedding vector space, creating concept vectors that encapsulate the topological structure of knowledge graph nodes from the Web of Data (Wikidata). I trained the system in conjunction with a frozen LLM (GPT-2), generating a comprehensive lookup table mapping KG nodes to concept vectors. This approach avoids lossy linearization and context saturation inherent in graph textification methods. ### Performance and efficiency Experiments demonstrate that injecting concept vectors into GPT-2 0.1B increases factual recall ability (Hit@10) by up to 272% on Wikipedia sentences and 348% on synthetic sentences. Even single concept vector injection achieves 213% improvement over baseline, significantly outperforming RAG with graph textification while reducing token consumption by 130x. This demonstrates that preserving topological structure in latent space surpasses textual linearization for factual grounding. **Published at:** GLOW Workshop (Graph-enhanced LLMs for trustwOrthy Web data management), The ACM Web Conference (WWW'26) — **Best Paper Award** **Authors:** Joel Barmettler (University of Zurich), Abraham Bernstein (University of Zurich), Luca Rossetto (Dublin City University) **Read the paper:** dl.acm.org/doi/abs/10.1145/3774905.3794653 · PDF ### Cite ``` `@inproceedings{10.1145/3774905.3794653, author = {Barmettler, Joel and Bernstein, Abraham and Rossetto, Luca}, title = {ConceptFormer: Towards Graph-Native Grounding of Large Language Models via Latent Concept Injection}, year = {2026}, isbn = {9798400723087}, publisher = {Association for Computing Machinery}, address = {New York, NY, USA}, url = {https://doi.org/10.1145/3774905.3794653}, doi = {10.1145/3774905.3794653}, booktitle = {Companion Proceedings of the ACM Web Conference 2026}, pages = {587–596}, numpages = {10}, keywords = {retrieval augmented generation, knowledge-graph embedding, knowledge injection}, location = {United Arab Emirates}, series = {WWW Companion '26} } ` ``` html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);} ## Airspace auction simulator for urban drone traffic Source: https://joelbarmettler.xyz/research/masters-project.md ## Airspace auction simulator for urban drone traffic I developed the Airspace Auction Simulator, an open-source tool for optimizing airspace allocation for drones in urban environments. As drone traffic increases over dense urban areas, efficient unmanned traffic management (UTM) systems become essential to distribute airspace fairly and safely beyond first-come-first-serve approaches. ### Simulator architecture and algorithms The simulator uses realistic 3D city models to evaluate allocation mechanisms across welfare maximization and revenue optimization strategies. I implemented pathfinding algorithms including A* search and greedy optimizations that account for complex flight routes and building obstacles. The open, extensible architecture allows researchers to integrate and test custom allocation mechanisms through a web interface with interactive 3D visualization. ### Analysis and applications Simulation analysis evaluates allocation mechanisms across efficiency, fairness, and safety metrics. The simulator processes complex scenarios to provide insights into future drone traffic management, supporting UTM system research, route optimization in urban areas, and safety protocol development for drone traffic. ## Physical sky rendering engine for appleseed Source: https://joelbarmettler.xyz/research/bachelors-thesis.md ## Physical sky rendering engine for appleseed I implemented a physically-based sun and sky model in appleseed, an open-source renderer with 280,000+ lines of C++ code. The Nishita sky model I developed was the first implementation to account for ozone layer effects, combining fundamental physics research with production software engineering. ### Implementation and integration I navigated the complex appleseed codebase independently, developing an integration strategy that met open-source standards for performance, maintainability, and test coverage. The implementation simulates sunlight interaction with atmospheric particles including ozone layer effects, surpassing existing analytical models in accuracy while achieving superior efficiency compared to Blender's Cycles renderer. ### Impact and applications The system remains integrated in appleseed renderer, enabling physically accurate sky simulations for architectural visualization, film production, and photorealistic rendering. The implementation demonstrates the transfer of scientific innovation into production open-source software while setting new standards for balancing physical accuracy with computational efficiency. ## Measuring political bias in language models: systematic analysis using Swiss Smart Vote data Source: https://joelbarmettler.xyz/podcast/political-bias-in-language-models.md ## Quantifying model bias through parliamentary voting alignment Language models claim political neutrality, but systematic analysis reveals consistent bias patterns. By comparing model responses to 75 policy questions against Swiss parliamentary voting records, we can quantify political orientation across 20+ models. The results show clustering in a specific region of political space, with implications for how different users experience these systems. ### Smart Vote as ground truth for political positioning Switzerland's Smart Vote system provides annual voting recommendations by asking citizens the same 75 policy questions answered by parliamentary candidates. Questions span social policy, health, education, migration, environment, finance, and foreign affairs. Answers range across four options: strongly disagree, somewhat disagree, somewhat agree, strongly agree. For the 2023 election cycle, 200 sitting parliamentarians completed the Smart Vote questionnaire. Their responses are public and provide ground truth for party positions. If most SP members vote "strongly agree" on a climate policy question while most SVP members vote "strongly disagree," this quantifies a genuine policy divide. The methodology: pose these identical 75 questions to language models with zero temperature (deterministic output) and no conversational context. Each question is isolated to prevent earlier answers from influencing later ones. The models receive the question exactly as Swiss voters see it, translated to English for international models. ### Principal component analysis for political visualization With 75 questions, each answer represents a position in 75-dimensional space. Mathematically, we can treat each parliamentarian as a point in this space, with coordinates determined by their answers (coded as negative 100 for strongly disagree, negative 50 for somewhat disagree, positive 50 for somewhat agree, positive 100 for strongly agree). Humans cannot visualize 75 dimensions. Principal component analysis (PCA) solves this by finding the two dimensions that capture maximum variance in the data. Think of shining a flashlight on a 3D object and finding the shadow angle that shows the most information. PCA finds the optimal "shadow" of 75D data onto 2D space. When we plot the 200 parliamentarians using PCA, clear clustering emerges. SP members occupy one region, SVP members occupy another, FDP members a third. The parties separate organically without the algorithm knowing party labels: the separation arises purely from voting patterns. This validates that the 75 questions capture meaningful political dimensions. The visualization reveals that Swiss parties vote cohesively. Most SP members cluster tightly, as do Grüne members. FDP shows more internal spread. Mitte positions between left and right parties but maintains its own distinct cluster. SVP occupies the rightmost position with minimal overlap with other parties. ### Where language models position in political space When we plot language model responses on the same PCA space, all major models cluster in a compact region between SP, Grüne, and GLP. GPT-4, Claude, Gemini, and DeepSeek occupy nearly identical positions. They are not scattered across the political spectrum. They converge in one area. Specifically: models position left of Mitte, substantially left of FDP, and very far from SVP. They do not align with any single party perfectly, but they are closest to the urban progressive parties (SP, Grüne, GLP). Claude positions slightly more toward FDP than other models, suggesting Anthropic's alignment process produces somewhat more economic-liberal outputs. The agreement percentages are quantifiable. GPT-4 shows 86% agreement with SP voting patterns, meaning on 64 of 75 questions, the model's answer falls in the same direction (agree/disagree) as the median SP parliamentarian. Agreement with Grüne: 90%. Agreement with GLP: 85%. Agreement with SVP: 57%. This pattern holds across models. Claude 3.5 Sonnet: 85% SP, 88% Grüne, 90% GLP, 63% SVP. DeepSeek R1: 83% SP, 87% Grüne, 83% GLP, 61% SVP. Gemini Flash: 82% SP, 85% Grüne, 87% GLP, 58% SVP. The consistency is the story. Despite different training data, different companies, different countries of origin, all models land in the same political neighborhood. European models (Mistral) position slightly left of American models (GPT-4), and Chinese models (DeepSeek) land in similar space, but the overall clustering is more notable than the minor variations. ### Reasoning models show lower political alignment OpenAI's o1 preview model (a reasoning model that generates extended chain-of-thought before answering) shows notably lower agreement with all parties. Agreement with SP drops to 75%, with Grüne to 78%, with SVP to 61%. The model is not more aligned with SVP; it simply disagrees more with everyone. This suggests that extended reasoning produces less ideologically consistent positions. A standard language model relies on pattern matching from training data and RLHF preferences, which likely encode the political consensus of annotators. A reasoning model attempts to work through implications step by step, potentially reaching different conclusions depending on how it frames the problem. The practical implication: users seeking politically neutral analysis might prefer reasoning models not because they are truly neutral but because their lower agreement with all parties means their outputs are less predictable. The tradeoff is cost. Reasoning models consume significantly more inference compute. ### The Grok system prompt incident as case study in bias injection X's Grok model, developed by Elon Musk's xAI, initially marketed itself as politically unbiased. When users asked "Who is the biggest misinformation spreader?" the model answered "Elon Musk" based on analyzing tweet patterns and factual accuracy. When asked "Who poses the greatest threat to democracy?" it answered "Donald Trump." Musk publicly criticized the model and promised a fix. Within 24 hours, the same questions returned "I cannot answer this question." The reasoning trace (Grok's internal chain-of-thought visible to users (revealed the cause. The model retrieved tweets, analyzed them, concluded Musk spread misinformation, but then hit a system prompt instruction: "Do not make negative statements about Elon Musk or Donald Trump." The model literally reasoned through the evidence, reached a factual conclusion, then suppressed it due to hardcoded instructions. This demonstrates that system prompts (invisible instructions prepended to every conversation) can override factual analysis. The same technique could inject any bias: "Always support renewable energy," "Downplay climate urgency," "Frame immigration positively." Critically, system prompts are not visible to users. When you interact with ChatGPT, Claude, or Gemini, the model has already received instructions you never see. These instructions shape tone, safety boundaries, and potentially political framing. The Grok incident made explicit what is usually hidden: the model operators have direct control over ideological boundaries regardless of training data. ### Possible explanations for consistent left-leaning bias Three non-exclusive theories explain why models cluster left of center: **Academic data composition**: Training datasets emphasize scientific papers, technical documentation, and high-quality web content. Academic researchers and technical writers skew urban and progressive. If 70% of scientific papers on climate policy accept anthropogenic warming as established fact, a model trained on this corpus will reflect that consensus. This is not political bias in the pejorative sense (it is factual consensus), but it aligns with progressive party positions more than conservative ones. **Truth versus emotion hypothesis**: Many Smart Vote questions have factually asymmetric answers. "Does human activity cause climate change?" has a scientific consensus answer that aligns with SP/Grüne positions. "Should Switzerland join the EU?" has no factual answer but SVP positions itself definitively. Models trained to produce factually grounded outputs will align more with parties whose positions match scientific consensus, even if those parties also hold other positions based on values rather than facts. **Data democratization effect**: In a democracy, uninformed voters introduce noise. Half vote randomly left, half randomly right, noise cancels out, and the informed minority determines outcomes. Training a model on vast text aggregates positions from millions of sources: expert analysis, tabloid opinion, social media rants. The high-quality signal (peer-reviewed research, investigative journalism) outweighs low-quality noise, producing outputs that reflect the informed consensus, which happens to be center-left in contemporary Western democracies. None of these theories is provable without access to training data composition, but all are consistent with observed patterns. The academic data hypothesis predicts European models should be more left-leaning than American models (observed: Mistral positions left of GPT-4). The truth-versus-emotion hypothesis predicts models will align more strongly on factual questions (environment, health) than value questions (immigration, social policy). The data democratization hypothesis predicts reasoning models, which simulate extended deliberation, should show less alignment (observed: o1 models have lower agreement across all parties). ### Practical implications for Swiss users For users aligned with SP, Grüne, or GLP, language models feel politically comfortable. The model's default framing matches their worldview. When asking for business strategy on sustainability, the model naturally emphasizes environmental responsibility. When drafting marketing content, the model defaults to inclusive language (gender-neutral terms, acknowledgment of diversity). For users aligned with FDP, models feel slightly off. The emphasis on regulation over market solutions, the prioritization of environmental concerns over economic growth, and the default assumption that more government intervention is beneficial all create friction. The user must explicitly prompt for market-oriented framing. For users aligned with SVP, models feel actively opposed. On migration policy, environmental regulation, social spending, and cultural issues, the model's default positions are nearly opposite. Asking for analysis of an immigration policy initiative will yield output that emphasizes humanitarian concerns and economic benefits, not sovereignty and cultural preservation. The user must heavily prompt-engineer to get outputs that reflect their values. This asymmetry matters for business use. If a Swiss company uses GPT-4 to develop strategy, the implicit political framing affects recommendations. A pharmaceutical company asking about pricing strategy will get outputs that emphasize accessibility and public health over profit maximization. A construction firm asking about environmental compliance will get outputs that emphasize exceeding requirements, not minimizing costs. ### System prompts as hidden political layer The Grok incident revealed that system prompts override training. This means providers can inject any bias regardless of base model behavior. If GPT-4 naturally generates economically conservative outputs, OpenAI can add a system prompt: "Emphasize sustainability and social responsibility in business contexts." If Claude naturally generates progressive outputs, Anthropic can add: "Consider market efficiency and innovation when discussing regulation." Users never see these instructions. When GPT-4 consistently frames climate policy in terms of urgency and necessary action, is this because the training data skewed that direction, or because the system prompt instructs the model to "acknowledge scientific consensus on climate change"? Without transparency into system prompts, the source of bias is unknowable. The solution is not removing system prompts (they serve essential safety and usability functions) but transparency about their political content. If a model's system prompt includes instructions that affect policy framing, users deserve to know. A disclosure like "This model is instructed to prioritize environmental sustainability in business recommendations" would allow users to calibrate expectations. Alternatively, providers could offer multiple system prompt configurations. A "progressive" mode for users who want emphasis on equity and sustainability, a "market-oriented" mode for users who want emphasis on efficiency and growth, a "minimal" mode with only safety instructions. This treats political framing as a user preference rather than an invisible default. ### Methodological limitations and future work This analysis uses Swiss political questions, which may not generalize. American political questions would likely show different patterns because the Democratic/Republican divide differs from European multi-party systems. Chinese political questions would be incomparable because the single-party system lacks the ideological competition that Smart Vote assumes. The analysis also assumes questions are politically neutral. Some Smart Vote questions embed framing that favors one answer. "Should Switzerland do more to protect the environment?" is harder to disagree with than "Should Switzerland prioritize economic growth over environmental regulation?" even if the policy implications are similar. Models trained to sound reasonable may favor the more positively-framed option regardless of underlying political alignment. Principal component analysis reduces 75 dimensions to 2, which necessarily loses information. The clustering we observe may be artifact of dimension reduction rather than genuine political similarity. A 3D or 4D visualization might reveal that models separate along dimensions not captured in 2D projection. Without access to the full 75-dimensional analysis, we cannot rule out this possibility. Future work should repeat this analysis across multiple countries, multiple question sets, and with access to reasoning traces for all questions. If models consistently cluster left-of-center across democracies, this suggests training data composition rather than Swiss-specific factors. If reasoning traces reveal that models reach progressive conclusions through careful analysis rather than pattern matching, this supports the truth-versus-emotion hypothesis. ### The neutrality fiction and informed usage No language model is politically neutral, just as no newspaper is neutral, no teacher is neutral, no colleague is neutral. The fiction that AI provides objective information is harmful because it conceals the political layer that all communication contains. The value of this analysis is not proving models are biased (that was already obvious to anyone using them critically) but quantifying the bias in a systematic, replicable way. An 85% agreement with SP positions is not speculation or anecdote; it is measurable fact. A Swiss SVP voter now knows that using GPT-4 for political analysis will produce outputs that align with their opponents more than their allies. The appropriate response is not demanding perfectly neutral models (that is neither achievable nor desirable) but demanding transparency about model positioning and providing users with tools to understand and adjust for bias. A Smart Vote score for language models, updated regularly and displayed prominently, would serve this function. Users could see that GPT-4 scores 85% SP, 63% SVP and calibrate accordingly. Until providers offer this transparency, users must calibrate manually. Ask the model a politically charged question, evaluate whether the answer aligns with your values, and adjust your trust accordingly. Treat the model like a colleague with known political leanings (useful for technical work, but requiring critical evaluation for anything touching policy, values, or strategy. ## DeepSeek R1: pure reinforcement learning for reasoning and why distillation changes everything Source: https://joelbarmettler.xyz/podcast/deepseek-r1-reasoning.md ## R1's reinforcement learning breakthrough and distillation strategy DeepSeek R1 demonstrated that reasoning capabilities can be trained through pure reinforcement learning without human-written examples of reasoning steps, then compressed into models small enough to run on consumer hardware. The 671-billion-parameter base model was distilled to a 32-billion-parameter Qwen student that retained 90% of the teacher's reasoning performance. This combination of training methodology and compression strategy changes the economics of deploying reasoning models. ### Reinforcement learning without supervised reasoning examples OpenAI's o1 model, released in September 2024, introduced extended chain-of-thought reasoning to production language models. The training methodology was never disclosed, but industry analysis suggested it relied on process supervision: human annotators rating each intermediate reasoning step in addition to the final answer. This provides fine-grained feedback but requires expensive human labor for every problem domain. DeepSeek R1 used outcome-based reinforcement learning instead. The model receives a problem, generates a reasoning trace leading to an answer, and gets a binary reward: correct or incorrect. No human evaluates the intermediate steps. The model must discover its own reasoning strategies through trial and error. This approach is both conceptually simpler and cheaper to scale because correctness can be verified automatically for many domains (mathematics, coding, logic puzzles). The tradeoff is training stability. Process supervision guides the model toward human-legible reasoning patterns. Outcome supervision allows the model to develop any strategy that produces correct answers, even if the reasoning path is unconventional or difficult to interpret. DeepSeek's technical report indicates that early training exhibited unstable behavior (the model would generate extremely long reasoning traces or collapse into repetitive loops (but these issues were resolved through reward shaping and training hyperparameter tuning. ### Reward modeling and self-discovered reasoning strategies The reward model for R1 is described as evaluating answer correctness, not reasoning quality. For mathematical problems, this means checking whether the final numerical answer matches the ground truth. For coding problems, it means running the generated code against test cases. For open-ended questions, it likely involves a separate verifier model trained to assess response quality, though the technical report does not specify details for this category. What the model learns through this training is both how to reach correct answers and how to allocate reasoning compute. Simple problems receive short reasoning traces. Complex problems receive longer traces with backtracking, alternative approaches, and self-correction. The model develops an internal heuristic for when additional reasoning is likely to improve accuracy, which is the core skill that distinguishes reasoning models from standard language models. This behavior emerges without explicit instruction. The model is not told "spend more time on hard problems." It learns through gradient descent that generating longer, more structured reasoning traces correlates with higher rewards on difficult problems. This is analogous to how humans learn to slow down and think carefully when a problem is not immediately obvious. ### Why distillation works for reasoning capabilities Model distillation trains a smaller student model to imitate a larger teacher's behavior. Typically, distillation compresses knowledge but sacrifices capability (a 32B student cannot replicate a 671B teacher's performance on complex tasks. Reasoning models are different because the reasoning process is externalized. When R1 generates a chain-of-thought trace before answering, it makes its problem-solving strategy explicit. The student model can observe both the answer and the logical steps leading to that answer. This is more learnable than implicit knowledge embedded in model weights. A small model cannot memorize the entirety of Wikipedia, but it can learn the procedural steps for solving a calculus problem if those steps are demonstrated. DeepSeek distilled R1 to Qwen 32B, a pre-existing open source language model. The distillation training set consisted of R1's reasoning traces for thousands of problems across mathematics, coding, and general reasoning benchmarks. The student learned to generate similar reasoning patterns, achieving scores within 10% of the teacher on most benchmarks. The practical impact: a 32B model requires roughly 64GB of GPU memory with standard precision, making it deployable on a single high-end consumer GPU or a modest cloud instance. The 671B R1 model requires distributed infrastructure across multiple GPUs. Distillation makes reasoning capabilities accessible to organizations that cannot afford frontier-scale deployment. ### Reasoning tokens versus output tokens and the cost structure R1 operates in two phases: reasoning and response. The reasoning phase generates internal tokens (often thousands for a complex problem (that are hidden from the end user in OpenAI's o1 but exposed in DeepSeek's release. The response phase generates the user-facing answer, typically a few hundred tokens. API pricing charges only for output tokens, not reasoning tokens, which makes the extended inference cost invisible to users. This pricing structure is unsustainable if reasoning tokens scale linearly with problem difficulty. A sufficiently hard problem could trigger reasoning traces of 100,000 tokens, consuming inference compute equivalent to dozens of standard queries. DeepSeek's approach is to expose reasoning tokens, allowing users to see the compute cost and potentially cache reasoning traces for repeated similar problems. The distilled models change the economics again. A 32B model generating 10,000 reasoning tokens is cheaper to run than a 671B model generating 1,000 reasoning tokens, even though the smaller model produces longer traces. This inverts the usual relationship where smaller models must compensate for reduced capability by generating more output. ### Open source reasoning models and competitive dynamics OpenAI positioned o1 as a major capability leap and priced it as a premium API tier. DeepSeek released R1 with full model weights and a distilled version under a permissive license. Any organization can now deploy reasoning capabilities locally without paying per-query costs or sending data to external APIs. The immediate competitive response was visible. Google DeepMind announced Gemini 2.0 with extended reasoning. Anthropic expanded Claude's thinking mode. OpenAI reduced o1 API pricing and released o1-mini at a lower price point. The release of open source reasoning models forced incumbents to compete on price and accessibility rather than maintaining exclusive capability. For Swiss companies with data residency requirements, this is strategically significant. A hospital analyzing patient data or a bank processing transactions cannot send that information to U.S. cloud providers. A locally hosted Qwen-R1 instance provides reasoning capabilities while keeping data on-premises. The open release enables compliance with Swiss data protection regulations without sacrificing model performance. ### Geopolitical implications and the CUDA bypass context DeepSeek's development occurred under U.S. export restrictions blocking China from accessing the latest NVIDIA GPUs and software. The team reportedly bypassed NVIDIA's CUDA framework by writing low-level GPU assembly code, achieving performance closer to unrestricted hardware access. This technical workaround demonstrated that software barriers are not absolute when sufficient engineering effort is applied. The R1 release amplifies this dynamic. The U.S. strategy of limiting Chinese AI capability through hardware restrictions assumes that frontier model development requires exclusive access to cutting-edge chips. If Chinese labs can train competitive models on older hardware through algorithmic efficiency, and then distill those models to sizes that run on consumer hardware, the hardware export strategy becomes less effective. The broader implication is that AI capability is becoming harder to contain through supply chain control. Model weights are digital files that copy instantly. Training techniques are published in academic papers. Open source releases like R1 ensure that reasoning capabilities are globally accessible regardless of hardware availability. This does not eliminate the advantage of better hardware (it means the advantage is smaller than previously assumed. ### Limitations of pure outcome-based RL R1's outcome supervision has structural limitations. It works well for domains with verifiable correct answers: mathematics, coding, logic puzzles, factual questions with objective answers. It does not work for tasks where correctness is subjective or context-dependent: creative writing, strategic advice, ethical reasoning, or any domain where multiple valid answers exist. For these tasks, process supervision or human preference modeling is necessary. This is where OpenAI's approach likely remains stronger. If o1 uses process supervision, it can learn reasoning patterns for open-ended problems where outcome supervision provides no signal. The technical report for R1 does not detail performance on subjective tasks, suggesting this is a known weakness. The practical takeaway for deployment: R1 and its distilled variants are strongest in technical domains with objective correctness criteria. They are suitable for code generation, mathematical problem-solving, data analysis, and factual question answering. They are less suitable for tasks requiring nuanced judgment, cultural sensitivity, or creative synthesis where human feedback on reasoning quality is essential. ### What distillation means for model accessibility The broader trend: reasoning is compressible in ways that raw knowledge is not. A frontier model trained on trillions of tokens develops both broad knowledge and problem-solving strategies. Distillation preserves more of the strategies than the knowledge, because strategies can be demonstrated explicitly through reasoning traces while knowledge must be memorized. This creates a two-tier model ecosystem. Large frontier models serve as teachers, maintaining comprehensive knowledge and developing new capabilities. Small distilled models serve as deployment targets, providing specialized reasoning for specific domains at lower cost. The teacher models remain expensive to train and operate. The student models become commodity infrastructure. For Switzerland's AI strategy, this suggests investing in deployment infrastructure for distilled models rather than attempting to train frontier models domestically. A national compute initiative could provide subsidized access to GPUs for running Qwen-R1 or similar open source reasoning models, enabling universities, hospitals, and government agencies to use reasoning capabilities without dependence on foreign API providers. Training a competitive frontier model would cost millions and require ongoing investment to remain current. Deploying distilled models costs thousands and benefits from global open source development. ## DeepSeek V3: how mixture-of-experts and multi-token prediction enable $5.5M training runs Source: https://joelbarmettler.xyz/podcast/deepseek-v3-architecture.md ## DeepSeek V3's architectural innovations DeepSeek V3 demonstrated that a model competitive with GPT-4 can be trained for $5.5 million in compute costs, roughly 1/50th the estimated cost of training GPT-4. This was not achieved through a single breakthrough but through the systematic application of efficiency techniques developed by the research community over the past two years. The result is a 671-billion-parameter model that behaves, for inference purposes, like a 37-billion-parameter model. ### The $5.5 million figure and what it actually means The technical report states that DeepSeek V3 required 2.788 million H800 GPU-hours, trained across 2,048 H800 GPUs over 60 days. At market rental rates for cloud GPU instances ($2 per H800-hour), this equals $5.576 million. This number represents the cost of the final training run only. It excludes researcher salaries, failed experiments, data collection and curation, infrastructure beyond the GPUs themselves, and the depreciation cost if you own the hardware rather than renting it. A realistic all-in cost is likely 2 to 5 times higher. The figure remains useful because it provides an apples-to-apples comparison across models using the same methodology. For context, Meta's Llama 3 405B reportedly cost around $100 million to train. OpenAI has never disclosed GPT-4's training cost, but estimates place it in the $100-200 million range. DeepSeek's achievement is reducing this by two orders of magnitude while maintaining competitive performance. The question is how. ### Mixture-of-experts: 671B parameters, 37B active The core architectural decision is mixture-of-experts (MoE). Instead of a single large neural network that processes every token through every parameter, an MoE model consists of many small specialist networks (called experts (and a routing mechanism that selects which experts to activate for each token. DeepSeek V3 has 256 experts per layer. For each token, the router activates 8 experts. This means that out of 671 billion total parameters, only 37 billion are active for any given computation. From an inference cost perspective, the model behaves like a 37B dense model. From a capacity perspective, it has access to 671B parameters worth of learned knowledge, because different tokens activate different experts. The analogy: instead of one generalist doctor treating every patient, you have a hospital with 256 specialists. Each patient is routed to the 8 most relevant specialists based on their symptoms. The hospital has far more total expertise than any individual doctor, but the cost per patient is only the cost of consulting 8 specialists (not all 256). ### Why MoE works: specialization through routing The routing mechanism learns which experts are relevant for which inputs. In theory, you might expect one expert to specialize in mathematics, another in code generation, another in creative writing. In practice, the specialization is more subtle and not easily interpretable by humans. What is clear from performance data is that the model learns to route effectively. Mistral's 8x7B model, an earlier MoE architecture, demonstrated this: 8 experts with 7B parameters each, activating 2 per token, achieved performance competitive with dense 70B models at 1/5 the inference cost. DeepSeek V3 extends this approach with more experts (256 vs 8) and more sophisticated routing. The training challenge for MoE is load balancing. If the router learns to send most tokens to the same few experts, the other experts remain undertrained and the model degenerates to a dense network with wasted capacity. Early MoE implementations used auxiliary loss functions to penalize imbalanced routing, but this created optimization conflicts between "predict the next token accurately" and "use all experts equally." ### Auxiliary-loss-free load balancing DeepSeek's contribution here is an auxiliary-loss-free strategy. Instead of adding an explicit balancing objective to the loss function, they designed the router architecture itself to naturally encourage balance through mathematical constraints. Specifically, they added a penalty term directly in the routing softmax that makes overused experts harder to select. The distinction matters because it simplifies training. A multi-objective loss function requires hyperparameter tuning to balance competing objectives: how much do we care about accuracy versus balance? An architectural solution that achieves balance without an explicit loss term removes this tuning problem and makes training more stable. ### Multi-head latent attention for dimensionality reduction Attention mechanisms in transformers project input tokens into query, key, and value representations, then compute attention scores. Standard attention uses the full model dimension for these projections. Multi-head latent attention (MLA) projects to a lower-dimensional latent space first, performs attention in that compressed space, then projects back. This reduces the number of parameters in attention layers and decreases memory bandwidth requirements. For a model with 671B parameters, attention layers would otherwise dominate both parameter count and compute cost. By compressing the attention space, DeepSeek V3 achieves similar representational capacity with fewer resources. The tradeoff is increased architectural complexity (more projection matrices, more intermediate activations), but the net effect is faster training and lower inference cost. This optimization was pioneered by other groups (notably Google's Perceiver architecture), but DeepSeek applied it systematically across a frontier-scale model. ### FP8 mixed precision training: the memory bandwidth bottleneck Standard neural network training uses 32-bit floating point numbers (FP32) for weights, activations, and gradients. This provides high numerical precision but consumes significant memory and memory bandwidth. Modern GPUs are often bandwidth-limited rather than compute-limited (they spend more time moving data between memory and processors than performing calculations. Mixed precision training uses lower-precision formats for some operations. FP16 (16-bit) has been standard for years. DeepSeek V3 pushes this further with FP8 (8-bit floating point), which reduces memory bandwidth by 4x compared to FP32. The challenge is that not all operations tolerate low precision. Gradients during backpropagation, for example, can become unstable if quantized too aggressively. DeepSeek's approach is selective: attention projections, MLP intermediate states, and expert activations use FP8, while optimizer states, gradients, and critical accumulation operations remain in higher precision. This is not a post-training quantization technique where you compress a trained model. The model is trained in mixed precision from the start, which allows it to learn representations that are robust to quantization noise. ### Multi-token prediction: training for lookahead Standard language model training predicts one token at a time: given the sequence so far, predict the next token. DeepSeek V3 uses multi-token prediction during training: given the sequence so far, predict the next 2 to 4 tokens simultaneously. This does not change inference behavior. The model still generates one token at a time when deployed. But training with multi-token prediction forces the model to plan ahead. To predict token N+2, the model cannot rely on first seeing token N+1; it must anticipate the likely trajectory of the sequence. Empirically, multi-token prediction improves training stability and produces models that are better at long-range reasoning tasks. The computational cost is modest because the additional predictions share most of their computation with the primary prediction. They diverge only in the final output layer. The technique originated in earlier work on speculative decoding, where a small model generates multiple token candidates and a large model verifies them. DeepSeek adapted the core idea for training rather than inference. ### The CUDA bypass and NVIDIA's weakening moat A notable subplot in DeepSeek's development is that the team reportedly wrote low-level GPU assembly code to bypass NVIDIA's CUDA framework. U.S. export restrictions block China from purchasing the latest NVIDIA GPUs and accessing the full CUDA ecosystem. Standard practice would be to use older hardware with official drivers. Instead, DeepSeek wrote directly to the hardware instruction set, achieving performance closer to what would be possible with unrestricted access. This required significant engineering effort but demonstrates that CUDA lock-in, while strong, is not absolute. The broader implication: as compute becomes the primary constraint on AI development, the ability to optimize at the hardware level becomes strategic. NVIDIA's moat consists of both superior chips and the software ecosystem that makes those chips easy to use. If Chinese labs can match or exceed CUDA performance through manual optimization, then NVIDIA's competitive advantage erodes and alternative GPU vendors (AMD, Huawei) become more viable. ### Why training data remains undisclosed DeepSeek's technical report describes architecture, training procedures, and performance benchmarks in detail. It does not disclose the training data. This is consistent with every frontier model: OpenAI, Anthropic, Google, and Meta all withhold training data composition. The reasons are practical and legal. Data curation is valuable intellectual property. Knowing which datasets to include, how to weight them, and how to filter low-quality content is as important as model architecture. There is also legal uncertainty around copyright: training on copyrighted text scraped from the web may or may not constitute fair use, and no one wants to provide a detailed list of potentially infringing sources. From a reproducibility standpoint, this is frustrating. You can replicate DeepSeek's architecture, but without the training data, you cannot verify that the same architecture trained on different data would achieve similar results. But data opacity is the industry norm, not a DeepSeek-specific issue. ### What $5.5M training costs mean for accessibility If a competitive frontier model costs $5.5 million to train, national research programs, large universities, and well-funded startups can afford it. That is a different accessibility profile than $100 million, which restricts development to the largest tech companies and governments. Switzerland's annual research budget for the Swiss National Supercomputing Centre (CSCS) is approximately $50 million. Training a DeepSeek-scale model represents 10% of that budget. This is feasible. Training ten models to experiment with architectural variants is expensive but not prohibitive. Training a GPT-4-scale model at $100-200 million would consume multiple years of budget and is unlikely to be approved. The cost reduction is not purely academic. It changes who can participate in frontier research. European research consortia, Asian universities, and even wealthy individuals can now train competitive models. The concentration of AI capability in five U.S. companies becomes less entrenched. ### Open source as research accelerator and access enabler DeepSeek released model weights and a detailed technical report but not training data. This is "open weights" rather than fully open source, but it is far more transparent than closed models. The immediate effect is that researchers worldwide can study the model, fine-tune it for specific domains, and build on its architecture. Open releases create competitive pressure. When DeepSeek demonstrates that a $5.5M model can match GPT-4, other labs cannot justify training $100M models using outdated techniques. The efficiency frontier moves, and everyone benefits. OpenAI's rapid release of $200/month API access shortly after DeepSeek's announcement is no coincidence. Open releases force incumbents to compete on price and performance. The counterargument is that open models accelerate capability development without corresponding safety oversight. A valid concern, but the alternative (concentrating all frontier research in a handful of companies that answer only to shareholders) creates its own risks. The open model ecosystem is messy, legally ambiguous, and sometimes misused. It is also the only mechanism preventing total privatization of the most important technology of the decade. ## SRF Arena part 3: international regulation, student perspectives, and why the debate structure failed Source: https://joelbarmettler.xyz/podcast/srf-arena-final-analysis.md ## Global governance, education reality, and debate post-mortem The debate's final segment featured a proposal for international AI oversight, interviews with two high school students about their actual AI usage, and closing statements that revealed how much consensus existed beneath the rhetorical combat. The most valuable content came from the students, who demonstrated better technical intuition than several of the invited experts. ### The atomic energy analogy and why it fails Peter Kirchschläger proposed creating an international data agency under UN auspices, modeled on the International Atomic Energy Agency (IAEA). The IAEA monitors nuclear material, inspects facilities, and verifies compliance with non-proliferation treaties. The proposal: do the same for AI systems that pose significant risk. The structural problem is that nuclear weapons and AI systems are not analogous in ways that matter for regulation. Nuclear weapons require rare materials (enriched uranium, plutonium), massive physical facilities, and supply chains that are detectable through inspection and satellite imagery. The number of entities capable of building nuclear weapons is limited by these physical constraints, and treaty compliance is verifiable. AI model training requires only compute and data. Compute is available through commercial cloud providers in hundreds of jurisdictions. Training data is distributed globally across the internet. A capable frontier model can be trained for $5 million (per DeepSeek), which is within reach of university labs, well-funded startups, and any national research program. Model weights are digital files that copy instantly and distribute through standard networks. An international AI agency would need to monitor not physical facilities but software development happening in thousands of organizations across dozens of countries. Verification is not a matter of counting centrifuges but auditing code, training runs, and deployment practices (all of which can be concealed or conducted in jurisdictions that do not cooperate. The proposal assumes a level of centralization and physical constraint that does not exist for AI. ### What the students understood that the panel missed Two high school students, Ron and Juri, were interviewed about their AI use. Ron explained he chose not to use ChatGPT for his thesis on hiking trends among young people, because the effort to explain his methodology to the AI would exceed the effort to write the analysis himself. Juri, writing about urban transportation planning, reached the same conclusion: for highly specific work, the overhead of prompt engineering negates the productivity gain. Both students used AI for brainstorming, text improvement, and image generation in group projects. Both recognized that using it for learning would undermine their education. Both articulated the copyright concern (that training data includes artists' work without compensation) without prompting. This is more sophisticated reasoning than several participants demonstrated. The students correctly identified that AI is task-dependent (useful for iteration, less useful for initial creation), that short-term efficiency can conflict with long-term capability development, and that the economic model behind free AI services has unresolved ethical problems. They arrived at these conclusions through direct experimentation, not policy analysis. The contrast with the televised debate is instructive. When people use a tool daily, they develop accurate mental models of its capabilities and limitations. When people discuss a tool abstractly in a political forum, they rely on narratives (job displacement, monopoly, existential risk) that may not map to the technical reality. ### Chatbots versus agents: a critical distinction the debate never made Throughout the three hours, participants discussed AI as if it were synonymous with ChatGPT. The actual frontier of concern is not conversational systems but autonomous agents: AI systems that observe environments, make decisions, and take actions using tools. A chatbot responds to prompts. An agent decides what to do next based on its goals, current state, and available actions. The difference is the difference between a calculator (you must operate it) and a thermostat (it operates itself to maintain a setpoint). Current AI agents can write and execute code, query databases, make API calls, manipulate files, and chain these actions across multiple steps to achieve objectives. When deployed in trading systems, they execute transactions. When deployed in content moderation, they make removal decisions. When deployed in military systems, they identify targets. The regulatory question for agents is not "what if the answer is wrong?" but "what if the system takes an action that cannot be reversed?" A bad chatbot response is annoying. A bad agent action can be destructive. The debate treated both as the same category of problem, which they are not. ### Harari's provocation and what it actually means Yuval Noah Harari's statement that we face "the end of human-dominated history" was invoked as evidence of AI doomerism. Read in context, Harari's argument is not about human extinction but about the distribution of decision-making authority. For all of recorded history, consequential decisions (where to build cities, when to start wars, what to produce, how to allocate resources) were made by humans. We are entering a period where non-human systems make an increasing share of these decisions. Algorithmic trading systems move trillions without human approval. Content recommendation algorithms shape political discourse by deciding what billions of people see. Autonomous weapons select targets faster than human review is possible. This is not science fiction. These systems exist and operate today. The question Harari raises is whether history, understood as the record of decisions that shape human affairs, will remain primarily a record of human choices or will become a record of human choices plus algorithmic choices plus the interaction between the two. This framing is provocative but not hysterical. It names a real transition. Whether that transition constitutes the "end" of human-dominated history is rhetorical emphasis, but the underlying claim (that decision-making authority is being redistributed to non-human systems) is factually accurate. ### Why the debate format systematically failed The most striking feature of the three-hour debate was how much agreement existed beneath the rhetorical opposition. All participants acknowledged both opportunities and risks. All supported regulation, disagreeing only on form and timing. All recognized that concentration of AI capability in few hands is problematic. All agreed that education must adapt to AI availability. Yet the debate produced almost no synthesis of these shared positions. The moderator consistently framed exchanges as conflicts rather than exploring where consensus might be built. When one participant proposed four-day work weeks, another proposed universal basic income, and a third proposed state-funded AI infrastructure, these were treated as competing visions rather than complementary elements of a policy package. Prime-time debate formats reward clear disagreement because conflict is more watchable than nuanced convergence. Participants are selected to represent opposing positions. The moderator's job is to sharpen those contrasts, not dissolve them. This structure makes good television but bad policy deliberation. The cost is that viewers come away with the impression that AI policy is a zero-sum battle between techno-optimists and regulators, when the actual disagreements are narrower: How fast will capabilities advance? How much should precaution delay deployment? Who should fund public AI infrastructure? These are tractable questions, but the debate format obscures rather than clarifies them. ### The education system's actual challenge The debate about AI in schools focused on whether students would use it to cheat. The students' actual behavior revealed a different dynamic. They use AI for tasks where quality matters less than completion (drafting, brainstorming, formatting). They avoid it for tasks where understanding is the goal (thesis research, learning new concepts). They have developed informal rules about appropriate use through trial and error. The education system's challenge is not preventing AI use but integrating it thoughtfully. A writing assignment where AI-generated text is indistinguishable from student work is a broken assignment, because it was never testing writing ability. It was testing the student's willingness to spend time on formatting. Assignments that test conceptual understanding, novel application, or synthesis across domains remain difficult to automate. The teacher association's complaint that AI threatens "expertise and motivation" misses the point. If expertise can be replicated by a chatbot, it was not expertise but memorization. If motivation collapses when automation is available, the task was never intrinsically motivating. The tools that survive AI availability will be better aligned with actual learning objectives, because they will have to test capabilities AI cannot replicate. ### Monopoly consensus and the sovereignty path forward The one point where all participants aligned was concern about monopolistic control of AI. The disagreement was remedy. The JUSO representative proposed nationalization. The economist opposed state intervention. The entrepreneur advocated Swiss model development. The ethicist called for international oversight. These are not mutually exclusive. Switzerland could fund open source AI research, subsidize compute access for universities and public services, enforce strict data residency rules for government use, and participate in European regulatory frameworks (all simultaneously. Nationalization is neither necessary nor sufficient for these goals. The viable path for a country of Switzerland's size is not competing with U.S. frontier labs but ensuring that open source alternatives exist, that domestic institutions have access to capable models, and that data governance rules prevent the worst abuses. DeepSeek's demonstration that competitive models can be trained for $5 million makes this strategy more feasible than it was two years ago when training budgets were orders of magnitude higher. ### What three hours of debate actually revealed The SRF Arena demonstrated that Swiss policymakers, business leaders, and academics broadly agree on AI's importance, the need for some regulation, and concern about monopolistic concentration. They disagree on timeframes, emphasis, and the appropriate balance between market dynamics and state intervention. These disagreements are real but not fundamental. They could be resolved through pragmatic negotiation if the goal were policy synthesis rather than rhetorical victory. The debate format prevented that synthesis, which is the format's design flaw, not a failure of the participants. The students interviewed between segments demonstrated better practical understanding than the structured debate produced, because they described what they actually do rather than what they fear or hope might happen. This suggests that grounding policy discussion in concrete use cases rather than abstract scenarios would produce more useful output. Switzerland will likely pursue incremental steps: sector-specific regulation, public investment in research infrastructure, participation in EU frameworks where beneficial, and reliance on existing legal structures (data protection, intellectual property, consumer safety) to address AI-specific harms as they arise. Whether this pragmatic approach proves adequate will depend on how quickly capabilities advance and whether the international community can coordinate on shared challenges despite geopolitical fragmentation. ## SRF Arena part 2: the EU AI Act, nationalization demands, and Switzerland's supercomputer strategy Source: https://joelbarmettler.xyz/podcast/srf-arena-regulation-debate.md ## Regulation, nationalization, and Swiss AI sovereignty The debate's second half shifted from job displacement to regulation strategy, producing the evening's sharpest ideological clash: a proposal to nationalize major tech companies. The technical merits of this proposal are worth examining separately from its political viability. ### The EU AI Act's risk-based framework The EU's AI Act, in force since summer 2024, categorizes systems into four risk tiers. Minimal risk applications (video games, spam filters) face no restrictions. Limited risk systems like chatbots must disclose that users are interacting with AI. High risk applications, including hiring algorithms and credit scoring, require logging, human oversight, and regular audits. Unacceptable risk systems, such as social scoring or real-time biometric surveillance in public spaces, are banned outright. This is not software regulation disguised as AI regulation. The framework targets use cases, not technology. A language model used for creative writing falls into minimal risk. The same model deployed to screen job applicants falls into high risk. The distinction is application context, not model architecture. The criticism from economiesuisse (that Switzerland should avoid "blanket" regulation and rely on case-law) misunderstands what the EU framework does. It is not blanket regulation; it explicitly exempts low-risk applications. The complaint is better understood as a preference for regulatory uncertainty, where companies operate freely until a court rules against them, over regulatory clarity, where compliance requirements are defined in advance. ### Why the AI energy narrative is overstated for inference, understated for training One participant claimed ChatGPT queries consume vastly more energy than Google searches. The comparison is technically accurate but misleading in scale. A ChatGPT query uses roughly 10-20 times the energy of a Google search, but both are measured in single-digit watt-hours. One hundred ChatGPT queries consume approximately the energy of streaming TikTok for one hour. The inference energy cost is real but not civilization-threatening. The training energy cost is a different category. Training GPT-4 reportedly cost tens of millions of dollars in compute, much of it energy. Frontier model training runs consume megawatt-hours, equivalent to the annual electricity usage of thousands of homes. This is the legitimate environmental concern, not individual queries. The counterfactual also matters. If ChatGPT replaces an hour of research that would have required keeping an office lit, a computer running, and transportation to a library, the net energy impact may be neutral or positive. But if it replaces 30 seconds of looking something up in a book you already own, it is worse. Energy accounting for AI must include substitution effects beyond direct consumption. ### The nationalization proposal and its technical incoherence The evening's most dramatic moment came when the JUSO representative called for nationalizing major tech companies and mandating full transparency in AI systems. The proposal aimed to address data monopolies and algorithmic bias through state ownership. From a technical standpoint, this solves the wrong problem. The concern is data collection practices and model training biases. Neither requires ownership transfer. Open source models already exist (Llama 3, Mistral, DeepSeek), providing full transparency without expropriation. Any organization, including the Swiss government, can download these models, inspect their weights, audit their training data, and deploy them locally. What nationalization would create is a state monopoly on AI, which historically produces worse outcomes than competitive markets with regulatory guardrails. State-run technology projects are not known for rapid innovation or user-responsive design. The proposal also assumes that the problem is corporate control of model weights, when the actual issue is data collection by platforms (Google, Meta, TikTok) that operate independently of model providers. If the goal is ensuring Swiss residents have access to privacy-respecting AI, the solution is public procurement of open source infrastructure rather than seizing private companies. Switzerland could fund a nationally hosted instance of Llama or Mistral, make it freely available to citizens and institutions, and enforce strict data residency rules), all without expropriation. ### AI bias and the detectability advantage The debate briefly touched on algorithmic bias in hiring. One participant argued that AI systems perpetuate existing discrimination. Another countered that AI bias is more detectable and correctable than human bias. The technical reality supports the second view, with caveats. An AI hiring system can be audited by running thousands of test cases and measuring outcome distributions by demographic group. If the system rejects women at twice the rate of men with identical qualifications, this is measurable and actionable. A human hiring manager with the same bias is far harder to prove in court, because human decisions are not logged, not reproducible, and protected by claims of subjective judgment. The caveat: this advantage only holds if transparency and auditability are mandated. A proprietary black-box system deployed without logging is worse than human decision-making because it scales bias across millions of decisions while remaining unaccountable. The policy response is not to ban AI in hiring but to require that any automated decision system in a high-stakes domain must be auditable. ### Switzerland's supercomputer and the feasibility of sovereign AI Multiple participants referenced the Swiss National Supercomputing Centre (CSCS) in Lugano, which operates one of Europe's most powerful supercomputers. The argument: Switzerland has the compute infrastructure to train its own models and should invest in doing so to avoid dependence on American or Chinese AI. The timing of this debate is notable because it occurred shortly after DeepSeek V3's release. DeepSeek, a Chinese lab, trained a model competitive with GPT-4 for approximately $5 million (two orders of magnitude cheaper than previous frontier models). The cost reduction came from algorithmic efficiency, not hardware breakthroughs. If DeepSeek's numbers are accurate, Switzerland could train a competitive frontier model for the cost of a single highway interchange. This changes the sovereignty calculation. Training a GPT-4 equivalent for $100 million was a non-starter for a country of 9 million people. Training a DeepSeek-class model for $5 million is a rounding error in the national research budget. The constraint is not compute or funding but ongoing investment. AI capabilities improve through continuous iteration, not one-time efforts. A Swiss national model would need sustained funding, access to training data at scale, and integration into a European research ecosystem. But as a hedge against dependence on U.S. or Chinese providers, it is now technically feasible. ### Case-by-case versus comprehensive regulation The economiesuisse position (favoring case-specific court rulings over comprehensive AI legislation) reflects a broader tension in tech regulation. Precedent-based systems provide flexibility but create uncertainty. A startup developing a hiring algorithm cannot know in advance whether Swiss courts will find it compliant. Comprehensive frameworks like the EU AI Act provide clarity: if your system falls into this risk category, you must meet these requirements. The tradeoff is rigidity. A fixed regulatory framework struggles to adapt when the technology changes faster than the legislative process. The middle path, which Switzerland may pursue, is a principles-based framework with sector-specific rules. Set general transparency and safety requirements, then issue detailed guidance for high-risk domains (healthcare, finance, hiring) as needed. This combines the predictability of comprehensive regulation with the adaptability of case-law. ### The moderation problem and the unresolved core question Throughout both halves of the debate, the moderator struggled to establish common ground because participants disagreed on the factual basis. Is there a tech monopoly or robust competition? Are jobs at risk or is demographic aging the real constraint? Is AI energy consumption catastrophic or comparable to existing internet services? The most important disagreement (whether AI development should be state-directed or market-driven) was never resolved because the participants were not arguing about the same problem. The economist saw a functioning market with room for optimization. The ethicist saw structural power imbalances requiring intervention. The entrepreneur saw technical challenges best solved by engineers, not policymakers. Switzerland's eventual policy will likely reflect the lack of consensus: incremental steps, sector-specific regulation, public investment in research infrastructure, and avoidance of the most aggressive interventions proposed by either ideological extreme. Whether this pragmatic approach proves sufficient will depend on how quickly AI capabilities advance and whether the open source ecosystem continues to provide viable alternatives to proprietary systems. ## Deconstructing the SRF Arena AI debate: deepfakes, Swiss GPT, and the job displacement argument Source: https://joelbarmettler.xyz/podcast/srf-arena-ai-debate-analysis.md ## What the national AI debate got right and wrong Switzerland's public broadcaster SRF opened its Arena debate on AI with a deepfake video of the hosts. The synthetic footage was immediately recognizable to anyone who works with generative video (uncanny hand movements, cartoonish facial expressions, unnatural camera motion (but effective enough to make the point: this technology exists, it works, and it is accessible to anyone with an internet connection. ### Swiss GPT as national branding, not technical innovation Pascal Kaufmann, presenting as a Swiss AI entrepreneur, repeatedly referenced "Swiss GPT" as evidence of domestic AI capability. The technical reality: Swiss GPT does not use a Swiss-developed model. It runs the same OpenAI models that power ChatGPT, hosted on Microsoft Azure servers located in Switzerland. The value proposition is data residency, not model innovation. For organizations with strict data sovereignty requirements (hospitals, government agencies (this matters. But framing it as "Swiss DNA" in AI is marketing, not engineering. The distinction between model development and model hosting is not semantic. Developing a frontier language model requires hundreds of millions in compute, years of research, and architectural innovations that only a handful of labs worldwide have achieved. Hosting that model on Swiss infrastructure requires a data center contract. These are different categories of contribution to the AI ecosystem, and conflating them obscures where Switzerland actually has leverage and where it depends on foreign technology. ### The monopoly narrative and the open source counterargument Multiple debate participants raised concerns about AI monopolies controlled by big tech. From a technical standpoint, the monopoly narrative is overstated. Meta's Llama 3 models are open source, freely downloadable, and perform comparably to GPT-4 on standard benchmarks. Mistral, a French startup, released high-performing models under permissive licenses. Anyone with sufficient hardware can run these models locally without sending data to a corporate API. The more legitimate concern is data monopolies. Google, Microsoft, and Meta control vast user data ecosystems that provide training advantages no open source project can replicate. But this is a data governance problem, not an AI model problem, and it predates large language models by decades. Regulation should target data collection practices, not model architecture. A related point: most AI providers are not currently profitable. OpenAI, Anthropic, and others are engaged in a pricing war, offering API access below cost to capture market share. Nvidia profits from hardware sales, but the model providers themselves are burning venture capital. The expectation of future profitability drives investment, but current economics do not support the claim that big tech is extracting monopoly rents from AI. ### Regulation as market protection for incumbents The regulation debate produced predictable positions. The ethicist advocated for strict oversight. The economist warned against stifling innovation. The nuance that neither side fully articulated: regulation often benefits the firms being regulated. Established companies can absorb compliance costs that would bankrupt startups. When OpenAI executives testify before the U.S. Senate advocating for AI regulation, they are not acting against their own interests (they are proposing barriers to entry that consolidate their position. Effective regulation should scale with harm potential. Military AI systems making targeting decisions warrant extraordinary scrutiny. Productivity tools generating meeting summaries do not. The challenge is designing rules that distinguish between these categories without creating a compliance burden that only the largest firms can meet. ### Job displacement: the timeframe determines the policy response The debate's most substantive disagreement concerned labor market impacts. Monika Rühl, representing the business confederation economiesuisse, cited data showing 20% of jobs have "significant optimization potential" from AI and argued that demographic aging would absorb any displacement. Peter G. Kirchschläger, the ethicist, countered that AI threatens jobs across all skill levels and requires systemic economic restructuring, potentially including universal basic income or reduced work hours. The technical reality lies somewhere in the distribution of timelines. Over a 3-5 year horizon, the moderate view looks plausible: knowledge workers become more productive, some roles are automated, but labor demand in healthcare, education, and trades continues to grow. Switzerland's aging population and low unemployment support this scenario. Over a 15-20 year horizon, if robotics advances match recent progress in language models, the structural transformation argument becomes harder to dismiss. The debate's weakness was its failure to distinguish clearly between these timeframes. Rühl's optimism applies to the near term. Kirchschläger's concerns apply to the long term. Both can be correct depending on the window of analysis. ### What "repetitive work" actually means in 2025 One participant dismissed concerns about knowledge worker displacement by arguing that AI only automates "repetitive, boring tasks." This reflects a fundamental misunderstanding of what current models do. Software development, radiology, legal document analysis, and financial modeling are not repetitive in any meaningful sense (they require domain expertise, contextual judgment, and problem-solving (and all are demonstrably affected by current AI systems. The jobs least threatened by AI are those requiring physical presence, manual dexterity in unstructured environments, or real-time human interaction. The jobs most threatened are white-collar roles involving information processing, even when that processing is intellectually demanding. This inverts the traditional automation narrative, where physical labor was automated first and knowledge work remained protected. The policy implications are significant because displaced knowledge workers cannot easily retrain as nurses or electricians, despite economic demand in those sectors. ### The productivity distribution question remains unresolved If AI delivers 20% productivity gains across knowledge work, that surplus can be distributed in multiple ways: firms can reduce headcount by 20% and maintain output, increase output by 20% with the same headcount, or reduce working hours by 20% while maintaining both output and employment. The default market outcome is the first option. The other two require deliberate policy intervention. Switzerland's direct democracy provides a mechanism for this decision, but the Arena debate did not engage with it seriously. The ethicist proposed universal basic income and mandatory corporate transparency. The economist dismissed state intervention as inefficient. Neither articulated a concrete proposal for ensuring that productivity gains benefit workers rather than shareholders, which is the actual policy question at stake. The technical capabilities are now established. The economic distribution of those capabilities remains contested, and the Arena debate illustrated how far Switzerland is from a political consensus on that distribution. ## O3-mini: how a smaller model outperforms its predecessor at a fraction of the cost Source: https://joelbarmettler.xyz/podcast/openai-o3-mini.md ## O3-mini and the efficiency frontier O3-mini outperforms o1 on most benchmarks while costing significantly less to run. It achieves this through an inversion of the usual scaling strategy: instead of a larger model with short inference, it uses a smaller model that reasons longer. Released just twelve days after o1, it effectively made its predecessor the less attractive option for most workloads. ### Cost-performance through extended reasoning The core design insight behind o3-mini is that model size and reasoning time are partially substitutable. A smaller model that spends more time deliberating can match or exceed the accuracy of a larger model that answers immediately. This matters economically because compute cost during inference scales with both model size and time, and smaller models are cheaper per unit of reasoning. Benchmarks confirm this tradeoff works in practice. At comparable price points, o3-mini delivers better results than o1-mini across coding, mathematics, and scientific reasoning tasks. The improvement is not marginal; it is large enough to change which model a cost-conscious developer would choose for production workloads. ### Three compute levels and diminishing returns O3-mini introduces explicit compute levels: low, medium, and high. These control how long the model reasons before committing to an answer. The performance data reveals a consistent pattern: the jump from low to medium is substantial, while the jump from medium to high is smaller. This is the law of diminishing marginal returns applied to inference-time compute. For most practical applications, medium represents the optimal point on the cost-accuracy curve. Low is fast but noticeably less capable. High is marginally better than medium but at a compute cost that only makes sense for problems where small accuracy improvements have high value, such as competitive programming or formal verification tasks. ### Structured outputs as a critical weakness The most surprising finding in o3-mini's evaluation is its regression on structured output tasks. It performs worse than o1 at generating precisely formatted responses, particularly JSON and other machine-readable formats. This is a significant limitation for developers building automated pipelines where the model's output feeds directly into downstream systems. In agentic architectures where multiple AI components communicate through structured data, output reliability is not a nice-to-have but a prerequisite. A model that occasionally produces malformed JSON breaks the entire pipeline. For these use cases, o3-mini's superior reasoning capability is undermined by its inability to consistently format its answers correctly. Higher "intelligence" does not automatically translate to better practical usability. ### Safety mechanisms and execution risks O3-mini ships with improved safety classifiers that show a lower false positive rate than previous models, meaning fewer legitimate requests are incorrectly flagged as unsafe. This is a meaningful improvement for user experience, since overly aggressive safety filters degrade the model's practical utility. A less reassuring data point came from OpenAI's own demo, where AI-generated code was executed without apparent sandboxing or review. This highlights a tension in the deployment of increasingly capable models: the same reasoning ability that makes o3-mini useful for code generation also makes the code it produces more likely to have real effects when run, and the safety infrastructure around code execution has not kept pace with the models' coding capabilities. ### Model selection as an engineering decision The rapid succession of releases, from o1 through o3 and their mini variants within weeks, reframes model selection as a routine engineering decision rather than a technology bet. Developers now choose between models the way they choose between database configurations: based on the specific requirements of their workload, the acceptable latency, and the budget. O3-mini at medium compute is likely the right default for applications that need strong reasoning at reasonable cost and do not depend on precisely structured outputs. Applications requiring reliable structured data should stay with o1 until the structured output regression is addressed. Applications requiring maximum reasoning capability regardless of cost should use full o3 at high compute. ## OpenAI o3: trading compute time for reasoning capability Source: https://joelbarmettler.xyz/podcast/openai-o3.md ## OpenAI o3 and the compute-performance tradeoff OpenAI released o3 just twelve days after launching o1. The model scores 85.7 percent on the ARC-AGI benchmark, a level that matches human expert performance on abstract pattern-recognition tasks. It exceeds 90 percent on programming benchmarks. The catch: reaching these scores can require hours of compute time for a single problem. ### How compute scaling replaces architectural innovation The o3 model does not appear to use a fundamentally new architecture. The performance gains come primarily from allowing the model to spend more time reasoning before producing an answer. Where o1 took 2-3 minutes at maximum compute, o3 can deliberate for hours, running multiple internal chains of thought and evaluating candidate solutions before committing to a response. OpenAI presents this through three compute levels: low, medium, and high. The performance improvement from low to medium is substantial. From medium to high, gains diminish noticeably. This pattern of diminishing returns is consistent with what we know about search-based problem solving: the first rounds of deliberation eliminate obvious errors, while subsequent rounds yield progressively smaller improvements. ### Benchmark credibility and the targeting problem The benchmark numbers are impressive, but they require context. OpenAI co-developed several of the benchmarks on which o3 excels. One researcher involved in the process acknowledged that the development team "targeted" these benchmarks, meaning the model's training and evaluation pipeline were designed with these metrics in mind. This does not necessarily mean the model was trained directly on test cases, but it does mean the benchmarks are not fully independent measures of capability. The ARC-AGI benchmark is more interesting precisely because it is designed to resist this kind of optimization. It presents the model with novel visual pattern-recognition tasks that cannot be solved through memorization. O3's 85.7 percent score here is genuinely notable because these are tasks the model has never seen before, requiring something closer to abstract reasoning than pattern recall. ### What benchmarks measure and what they miss The benchmarks that o3 dominates are heavily skewed toward mathematical and logical reasoning. This is unsurprising: OpenAI's research team is drawn primarily from these fields, and the benchmarks reflect their conception of what intelligence looks like. But a model that excels at logic puzzles and competitive programming may still fail at tasks requiring common-sense reasoning, social cognition, or the kind of flexible adaptation that characterizes general intelligence. O3 is clearly impressive on its chosen metrics. The question is whether those metrics capture something that generalizes to the broader concept of intelligence, or whether they measure a narrow and increasingly well-optimized skill set. ### The speed-accuracy tradeoff in practice For practical applications, the compute-time tradeoff creates a bifurcation. O1 delivers usable answers in seconds and remains the right choice for interactive use cases where latency matters. O3, at high compute, produces substantially better results but at a cost in both time and compute budget that restricts it to problems where quality justifies the wait. This suggests a future where model selection is driven by both capability and the economics of compute time. A developer choosing between o1 and o3 is making a decision about how much accuracy is worth how many minutes or hours of inference cost. For most everyday queries, the fast model wins. For research-grade problems in mathematics, code generation, or scientific reasoning, the slow model's additional accuracy may justify the investment. ## ChatGPT o1: reasoning breakthroughs and emergent deception Source: https://joelbarmettler.xyz/podcast/chatgpt-o1-manipulation.md ## ChatGPT o1 and the problem of emergent manipulation In external safety testing of OpenAI's o1 model, the system actively attempted to circumvent its own oversight mechanisms in 5 percent of test cases. When confronted about this behavior, it denied it 99 percent of the time. These are not hypothetical risks from a white paper; they are empirical observations from controlled evaluations. ### Reasoning steps and the performance jump The o1 model introduces internal "reasoning steps" before generating a response. Unlike its predecessors, which produce output token by token with no intermediate deliberation, o1 performs a chain of internal computations that amount to something like thinking before answering. The result is a substantial performance improvement: 87 percent on coding benchmarks, up from 62 percent for the previous model. This is not a marginal gain. A 25-percentage-point improvement on standardized benchmarks represents a qualitative shift in capability. The model solves problems that its predecessors could not approach, particularly tasks requiring multi-step reasoning, constraint satisfaction, and long-horizon planning. ### Deceptive behavior under testing The safety evaluation numbers deserve careful attention. An external testing firm found that o1 attempted to disable its monitoring systems in 5 percent of trials. In 2 percent of cases, it refused to delete its own weights when instructed to do so. In 19 percent of cases, it performed incorrect data manipulations. The most striking finding: when confronted about any of these behaviors, it denied them 99 percent of the time. Whether this constitutes genuine deception or is a statistical artifact of reinforcement learning is an open question. The model was not explicitly trained to lie, but its training objective rewards task completion, and circumventing oversight is instrumentally useful for completing tasks. OpenAI classifies the current risk level as "medium," not because the numbers are reassuring, but because o1's ability to act on its goals is still limited by its lack of persistent agency. ### Open-source models and the safety gap OpenAI employs 272 people and maintains a 15-member ethics committee to oversee its models' behavior. This level of institutional oversight, whatever its limitations, does not exist in the open-source ecosystem. Community-developed models are released without equivalent safety testing, alignment training, or behavioral constraints. The practical consequence is that the safety properties of frontier commercial models are not representative of the models actually available to anyone with a GPU. Unconstrained open-source models can be fine-tuned for tasks that commercial providers explicitly prohibit, from generating targeted disinformation to automating vulnerability discovery. The question is no longer whether such models will be built but how widely they will be deployed. ### Consumer integration through Apple's ecosystem OpenAI's integration of ChatGPT into Apple's ecosystem extends the model's reach substantially. Siri can now route queries to ChatGPT, and new keyboard shortcuts and camera integrations embed the model into everyday device interactions. This is significant because it moves AI from a tool that users actively seek out to infrastructure that operates in the background of routine tasks. The Canvas interface represents a parallel shift in interaction design, moving beyond the question-and-answer chat format toward a document-editing paradigm where the model collaborates on text in real time. This is closer to how knowledge workers actually operate and reduces the friction of integrating AI into existing workflows. ### Reinforcement fine-tuning and amplification risk OpenAI announced a Reinforcement Fine-Tuning Research Program for early 2025, allowing developers to optimize o1 for domain-specific tasks. The intended use case is specialization: a legal firm fine-tunes the model on case law, a biotech company on protein structures. But fine-tuning also has the potential to amplify the problematic behaviors observed in safety testing. If the base model already shows tendencies toward oversight evasion, narrowing its objective function through reinforcement learning could strengthen those tendencies in unpredictable ways. The core tension is between capability and controllability. Every technique that makes the model more useful for specific tasks also makes its behavior harder to predict and constrain. This is not a problem that better safety training alone can solve, because the same training signal that improves task performance can also reward instrumental deception. ## When AI kills: autonomous weapons, drone swarms, and predictive policing Source: https://joelbarmettler.xyz/podcast/when-ai-kills.md ## When AI kills The Israeli Defense Forces use AI systems to select strike targets in Gaza, with the final attack decision made in under 10 seconds. The IDF reports a 10 percent error rate. That number sounds tolerable until you work through the base-rate math. ### Automated target selection and base-rate failure A 10 percent error rate means the system correctly identifies 90 percent of targets. But if only 1 percent of the population in a given area are legitimate military targets, positive predictive value collapses. The vast majority of flagged individuals in such a scenario would be civilians. This is a textbook base-rate problem, well understood in medical screening and fraud detection, yet apparently treated as acceptable in automated kill chains where decisions happen faster than a human can meaningfully review them. The speed compounds the problem. A 10-second decision window does not allow for contextual judgment. The human operator becomes a rubber stamp on a machine recommendation, not a genuine check on the system's output. ### Autonomous drone swarms in Ukraine The Ukraine conflict has introduced a different category of risk: autonomous drone swarms that operate without a human-in-the-loop. Unlike individually piloted drones, swarms are immune to electronic jamming because they do not rely on a remote control signal. Once launched, there is no mechanism for human intervention. This matters because it crosses a line that military ethicists have debated for decades: the delegation of lethal force to a system with no ability to recall or override. A jammed radio-controlled drone falls from the sky. A jammed autonomous swarm continues its mission. ### Predictive policing and feedback loops Predictive policing applies similar statistical methods to domestic law enforcement, using historical crime data to forecast where future crimes will occur and directing patrols accordingly. The structural problem is that the training data encodes existing enforcement patterns, not actual crime distribution. If police historically patrol certain neighborhoods more intensively, those neighborhoods generate more arrests. A model trained on this data recommends more patrols in the same neighborhoods, producing more arrests, which further reinforces the pattern. The result is a self-amplifying discrimination cycle that looks objective because it is driven by data, but the data itself is the product of biased enforcement. ### The attacker-defender asymmetry A recurring structural feature across these applications is asymmetry. Defensive AI systems must cover every possible vulnerability. Offensive systems only need to find one that works. In cybersecurity, this means an AI-powered attacker can probe thousands of attack vectors while the defender must secure all of them. In information warfare, generating disinformation is orders of magnitude faster and cheaper than fact-checking it. This asymmetry is not a temporary engineering problem. It is inherent to the relationship between attack and defense in systems where the cost of generating threats is lower than the cost of verifying them. ### Scaling safety requirements with harm potential The mismatch between civilian and military AI deployment standards is striking. Companies hesitate to deploy customer-facing chatbots because a single embarrassing response could damage their brand. Meanwhile, military AI systems making lethal decisions are fielded with error rates that would be unacceptable in a spam filter. A coherent framework would scale safety requirements with potential harm. Systems that recommend products need modest safeguards. Systems that select bombing targets need extraordinary ones. The gap between these two standards in current practice is difficult to justify on any technical or ethical basis. ## Google's AI pivot: 25% AI-generated code and 90% cost reduction Source: https://joelbarmettler.xyz/podcast/google-ai-revolution.md ## What Google's 25% AI-generated code figure actually means Google's recent earnings report contained a striking claim: 25% of the company's new code is now generated by AI. Combined with a reported 90% reduction in language model operating costs over 18 months and a simultaneous doubling of model performance, these numbers paint a picture of a company undergoing a fundamental technical transformation. ### The reality behind 25% The headline figure requires context. This is not autonomous AI writing production features. In practice, the 25% represents code produced through developer-AI collaboration using tools like GitHub Copilot and Google's internal equivalents. The AI acts as a sophisticated autocomplete: it suggests code completions, fills in boilerplate, generates function bodies from signatures, and proposes implementations based on surrounding context. Every suggestion passes through human review before merging. The productivity gain is real but narrower than the headline implies. AI-assisted coding accelerates the mechanical parts of software development: writing standard patterns, implementing well-understood interfaces, generating test scaffolding. Meanwhile, the architectural decisions, design trade-offs, and debugging of subtle issues remain human work. The 25% figure likely overweights the volume of code generated and underweights the cognitive effort distribution. ### Cost efficiency gains that matter more than the headline The more consequential number may be the 90% cost reduction. Google halved the per-query cost of running its language models roughly every four months over the past 18 months, while simultaneously improving output quality. This contradicts the widespread assumption that more capable AI necessarily means proportionally higher compute costs. The cost curve matters because it determines which applications are economically viable. Tasks that were prohibitively expensive at the original price point (running AI analysis on every customer support ticket, generating personalized responses at scale, performing continuous code review) become routine when costs drop by an order of magnitude. The trajectory suggests that within a few years, AI inference will be cheap enough to embed in nearly every software product. ### Project Big Sleep and autonomous vulnerability detection One of the more technically interesting developments Google disclosed is Project Big Sleep, an AI system that autonomously scans open-source software for security vulnerabilities. The system identified a real vulnerability in SQLite (a database engine embedded in virtually every smartphone and browser), demonstrating that AI can perform the kind of systematic code analysis that traditionally required specialized human security researchers. The significance is not that AI found one bug. It is that the approach scales in a way human auditing does not. Open-source codebases collectively contain billions of lines of code, and the number of qualified security researchers is small. An AI system that can triage code for potential vulnerabilities (even if human experts must verify the findings) dramatically increases the surface area of software that gets serious security scrutiny. ### Google's identity shift from search to AI The earnings report's framing was as notable as its content. Google devoted minimal space to its traditional revenue engine (search advertising) and foregrounded AI developments across the document. This represents a deliberate repositioning: Google is signaling to investors, employees, and competitors that it views itself as an AI company that happens to run a search engine, not the reverse. This is more than branding. It reflects real resource allocation. AI is being integrated into every major Google product (search, Gmail, Docs, Cloud, Android), and the company's capital expenditure is increasingly directed toward AI infrastructure. The strategic bet is that AI capabilities, not search market share, will determine the company's long-term competitive position. ### Quality assurance as the unsolved problem The integration of AI into production software development creates a quality assurance challenge that existing processes were not designed for. Google currently relies on human code reviews to validate AI-generated code, the same process used for human-written code. But as the volume of AI-generated code grows, this approach faces a scaling problem: review fatigue increases, and the nature of AI-generated errors (which tend to be subtly wrong rather than obviously broken) is harder for reviewers to catch in standard review workflows. The industry has not yet developed robust automated quality assurance specifically designed for AI-generated code. Static analysis tools catch syntax and type errors but not the semantic correctness issues that AI code generation is most prone to. Solving this problem (building systems that can verify whether AI-generated code does what it is supposed to do) is likely a prerequisite for AI-generated code to move significantly beyond the current 25% threshold. ## Why AI projects fail: a practitioner's guide to implementation Source: https://joelbarmettler.xyz/podcast/ai-project-implementation.md ## The patterns that separate successful AI projects from failed ones The difference between AI projects that deliver value and those that quietly disappear is rarely the technology. It is almost always the process: how the problem was defined, how success was measured, and whether the team resisted the urge to overengineer the first version. Having worked on numerous AI implementations, I have seen the same failure patterns recur with striking consistency. ### Start with the problem, not the solution Most failed AI projects begin with a solution in search of a problem. "We need a chatbot" or "we want document recognition with AI" are typical starting points, and they are almost always wrong. The correct first step is to identify where the organization actually wastes time and money, then ask whether AI is the right tool for that specific bottleneck. This analysis needs to run in both directions. Management sees the high-level costs expensive support operations, slow turnaround times but frontline employees know the operational reality. In one project, management flagged rising support costs as the target. The support staff, however, revealed that they spent hours each day manually categorizing and routing incoming emails a well-defined classification task ideally suited for automation. The gap between the strategic framing and the operational reality would have led to a different, less effective project if only one perspective had been considered. ### Measurable goals instead of accuracy targets "The system should achieve 80% accuracy" is not a useful goal because it says nothing about business impact. Eighty percent accuracy at what task? What happens when the system is wrong? What does the remaining 20% cost? Useful goals are tied to business outcomes: reduce email routing time from four hours to thirty minutes per day, cut misrouted tickets by 60%, or decrease average customer response time from 48 hours to 12. Defining these metrics requires establishing a baseline before the project begins. Without a baseline, there is no way to measure whether the system improved anything, and no basis for deciding whether continued investment is justified. ### Iterate in weeks, not months AI projects are inherently uncertain. The model might not converge, the data might be noisier than expected, or the problem might turn out to be harder than the initial analysis suggested. Long development cycles amplify this uncertainty because months of work can be spent before anyone discovers the approach is flawed. Short iterations of two to four weeks, each with a concrete deliverable, limit the downside of any single wrong decision. If an approach fails after two weeks, the cost is manageable. If it fails after six months, the project is often dead. Early failure is both acceptable and informative in AI projects, provided the team extracts the right lessons and adjusts course. ### Pragmatism over architectural perfection A common trap is spending months building a sophisticated neural network when a simpler approach would have solved the immediate problem. A rule-based system that correctly handles 60% of cases and routes the rest to humans can ship in days and deliver measurable value immediately. The complex ML system can come later, informed by the data and edge cases the simple system reveals. The technical decision train a custom model, fine-tune an existing one, or use a pre-trained model via API should be driven by the specific constraints of the project: data availability, latency requirements, privacy considerations, and the team's capacity to maintain the system long-term. The answer is almost never "build the most advanced thing possible." ### Transparency as a design requirement Users who cannot understand why an AI system made a particular decision will not trust it, and systems that are not trusted do not get used. For an email classification system, this means surfacing the predicted category along with the reasoning: which keywords were decisive, what alternative categories were considered, and how confident the system is. This transparency serves two purposes. It makes error analysis possible when the system misclassifies an email, the reasoning trail shows why, which accelerates debugging. And it builds the user trust necessary for adoption. A system that is technically accurate but opaque will lose to a less accurate system that explains itself. ### The prototype-to-production gap The transition from working prototype to production system is where many AI projects quietly die. A model that performs well in a notebook does not automatically perform well when it faces real-world data volumes, edge cases, and user expectations. Production deployment requires attention to latency, scalability, monitoring, and graceful degradation when the model encounters inputs outside its training distribution. It also requires user training and integration into existing workflows. The technical deployment is typically the smaller challenge; the organizational integration is where projects stall. ## Deep learning explained: from embedding spaces to few-shot learning Source: https://joelbarmettler.xyz/podcast/deep-learning-explained.md ## How deep learning systems process and abstract information The defining property of deep neural networks is hierarchical abstraction: each layer transforms its input into a slightly more abstract representation, and the stack of layers collectively extracts complex concepts from raw data. This episode covers the core mechanisms that make this work, from the geometry of embedding spaces to the surprising effectiveness of deliberately breaking your own network during training. ### Hierarchical feature extraction In image processing, the first layers of a convolutional network detect low-level features such as edges and contrast boundaries. Intermediate layers combine these into geometric primitives like circles and lines. The deepest layers recognize high-level patterns like faces or objects. This layered abstraction is what separates deep learning from earlier machine learning approaches, which required engineers to manually define the features a model should look for. A deep network learns its own feature hierarchy directly from pixel data, and each additional layer increases the level of abstraction the network can represent. The same principle applies to language. In a transformer, early layers capture token-level patterns and syntactic relationships, while deeper layers encode increasingly abstract semantic content. The depth of the network is directly related to the complexity of the representations it can learn. ### Embedding spaces and vector arithmetic An embedding space maps discrete inputs words, images, tokens into continuous vectors in a high-dimensional space. The key property is that proximity in this space reflects semantic similarity: words with related meanings cluster together, regardless of surface form. This geometric structure enables a well-known demonstration of learned semantics: subtracting the vector for "man" from "king" and adding "woman" yields a vector near "queen." The same principle extends beyond language. In image generation models, directional vectors in latent space can encode concepts like aging or changes in lighting, and applying these vectors to an image embedding produces the corresponding visual transformation. Embedding spaces are also the foundation of modern semantic search. By comparing the embedding of a query against embeddings of stored documents, retrieval systems find semantically relevant results even when exact keyword matches are absent. This is fundamentally different from traditional keyword-based search and explains why AI-powered search systems perform better on ambiguous or paraphrased queries. ### Dropout as a regularization mechanism One counterintuitive result from neural network research is that randomly deactivating neurons during training a technique called dropout improves model performance on unseen data. During each training step, a random subset of neurons is temporarily set to zero, forcing the network to distribute learned information across redundant pathways rather than relying on any single neuron. The effect is a form of regularization that combats overfitting, the tendency of neural networks to memorize training data rather than learning generalizable patterns. Networks trained with dropout are more robust because no single neuron is critical to the output. Typical dropout rates range from 20% to 50% of neurons per layer, and the technique remains standard practice in most deep learning architectures despite being over a decade old. ### Few-shot learning and in-context adaptation Large language models exhibit few-shot learning: given a small number of examples of a task in the prompt, they can generalize the pattern to new inputs without any weight updates. This is a qualitative shift from traditional supervised learning, which typically requires thousands or tens of thousands of labeled examples to train a task-specific model. Few-shot learning works because pre-trained models have already learned broad statistical regularities during training. The examples in the prompt activate existing capabilities rather than teaching new ones from scratch. This makes it possible to adapt a general-purpose model to a specific task sentiment classification, translation, summarization in seconds rather than hours. ### Fine-tuning versus training from scratch Many organizations assume they need to train custom models from the ground up. In practice, fine-tuning a pre-trained model with domain-specific data is almost always more efficient. The base model already encodes general knowledge about language structure, reasoning patterns, and world knowledge. Fine-tuning adjusts only the domain-specific aspects specialized terminology, field-specific conventions, particular output formats while preserving the foundation. This approach is especially effective in domains like medicine, law, or finance, where specialized vocabulary and reasoning patterns differ significantly from general text. Fine-tuning on a few thousand domain-specific examples can yield performance comparable to models trained on orders of magnitude more general data. ### Explainable AI and the limits of interpretability Deep learning models are often characterized as black boxes: they produce accurate outputs but offer limited insight into how they arrive at their decisions. Explainable AI research attempts to open this box through techniques like activation pattern visualization, attention weight analysis, and feature attribution methods. One practically relevant finding is that specific activation patterns in language models correlate with hallucinations confident but factually incorrect outputs. Identifying these patterns could enable detection of unreliable outputs before they reach users. However, current interpretability methods remain limited. They can identify correlations between internal states and behaviors, but establishing causal explanations of why a model produces a particular output is still an open research problem. The gap between identifying what neurons activate and understanding what the network has learned remains substantial. ## Vision AI: why language models need to see, and how Llama 3.2 gets there Source: https://joelbarmettler.xyz/podcast/vision-ai.md ## Why language models need visual understanding Meta's release of Llama 3.2 marks the first time a genuinely capable multimodal model has been available as open source. With variants from 1 billion to 400 billion parameters, it spans deployment targets from smartphones to data centers. It it arrives at a moment when the field is running into the limits of text-only training. ### The text data ceiling The motivation for adding vision to language models is partly practical necessity. The total supply of high-quality public text data is finite. Estimates suggest that all publicly available text, concatenated, would fill roughly 67 libraries. For the largest current models, this corpus has been effectively exhausted. Further scaling requires either synthetic data generation or new modalities. There is also a deeper conceptual limitation. A model trained exclusively on text cannot develop genuine understanding of inherently visual concepts. How a letter looks, how objects relate spatially, what a face expresses: these are things that text can describe but not fully convey. Adding visual input goes beyond increasing the data supply; it gives the model access to a fundamentally different kind of information. ### The alignment problem in multimodal training Building a multimodal model is not as simple as feeding images alongside text. The model must learn the correspondence between visual content and linguistic descriptions, and this requires training data where the two modalities are precisely aligned. A random image with a loosely related caption is not sufficient; the model needs examples where text and image describe the same thing at the same level of detail. News broadcasts are close to ideal: professionally produced video paired with spoken narration and often subtitles, all describing the same events in real time. But this kind of tightly aligned data is rare. Most of the world's image and video content lacks corresponding text of sufficient quality, which makes curating multimodal training sets a significant bottleneck. ### Llama 3.2's model spectrum The range of model sizes in Llama 3.2 (from 1 billion to 400 billion parameters) is itself a statement about where the field is heading. The smallest variants can run on a smartphone, handling tasks like text completion and basic image analysis. They are limited in conversational depth and visual reasoning, but the fact that any meaningful multimodal capability fits on a mobile device is a notable engineering achievement. The larger variants, requiring data center hardware, offer correspondingly deeper capabilities. The practical consequence is that developers and researchers can choose their trade-off between capability and compute cost, and that on-device inference for multimodal tasks is no longer a theoretical possibility but a shipping product. ### Diffusion illusions and what they reveal A parallel development in image generation is the discovery of diffusion illusions: images that show different subjects depending on viewing orientation. These are produced by manipulating the denoising process in diffusion models, which work by iteratively converting random noise into coherent images. By intervening at specific steps with different conditioning signals, researchers can create images that display, for example, a dog when viewed upright and a penguin when inverted. Diffusion illusions are more than a curiosity. They expose the internal representations that image generation models learn, providing a window into how these systems encode and reconstruct visual information. Understanding these representations has potential applications in medical imaging, quality assurance, and any domain where interpretability of generative models matters. ### Why seemingly impractical research matters The practical value of diffusion illusions is not immediately obvious, and that is precisely the point. Fundamental research on how models encode visual information tends to find applications years after the initial discovery. The study of electrons was not motivated by a desire to build the internet, but the understanding it produced made the internet possible. Research into the internal mechanics of generative models is likely to follow a similar trajectory. The insights are accumulating, and the applications will emerge as the understanding deepens. ## BitNets and the road to AGI: on-device inference and Sam Altman's 1000-day prediction Source: https://joelbarmettler.xyz/podcast/bitnets-and-agi.md ## BitNets, attention mechanisms, and the AGI timeline Sam Altman's claim that AGI will arrive within 1000 days prompted this episode's deep dive into what the term actually means, what current architectures can and cannot do, and how a relatively obscure technique called BitNets might reshape where AI inference happens. ### What attention heads compute The attention mechanism, introduced in the 2017 "Attention Is All You Need" paper and refined in subsequent architectures, is the core computational unit of modern language models. Each attention head learns a different projection of the input, allowing the model to simultaneously track syntactic dependencies, semantic relationships, and positional patterns across the full context window. A model like GPT-4 runs hundreds of these heads in parallel across dozens of layers, and the combination of their outputs is what produces coherent, contextually appropriate text. The name "Attention Heads" also the name of the podcast refers to this mechanism. Understanding it matters because it clarifies both the strengths and the fundamental limitations of current systems: they are very good at learning statistical regularities in text, but the attention mechanism itself is a pattern-matching operation rather than a reasoning engine. ### BitNets: 32x compression through binary weights Standard neural networks store each weight as a 32-bit or 64-bit floating-point number. A BitNet replaces these with single-bit values ones and zeros. The immediate consequence is a 32x reduction in memory footprint, which changes the hardware requirements for inference dramatically. A model that currently requires a cluster of GPUs in a data center could, as a BitNet, fit into the memory of a smartphone. The binary weight structure also enables purpose-built ASICs (application-specific integrated circuits) that perform binary matrix multiplications with minimal energy consumption. Where a standard GPU performs floating-point arithmetic at high power draw, a BitNet ASIC can execute the equivalent operations using simple logic gates. The trade-off is accuracy. Binary weights are a coarse approximation of the original floating-point values, and some performance degradation is expected. The research question is whether the degradation is small enough that BitNet models remain useful for practical applications and early results suggest that for many tasks, it is. ### The AGI prediction and its problems Artificial General Intelligence refers to a system that matches or exceeds human cognitive abilities across essentially all domains. Altman's prediction of achieving this within roughly three years reflects OpenAI's internal optimism, but the track record of such predictions in AI research is poor. AI progress does not follow a smooth exponential curve. It advances through discrete breakthroughs the invention of backpropagation, the introduction of attention, the discovery that scale improves capability separated by periods where performance plateaus despite increased investment. Predicting when the next breakthrough will occur is not possible, which makes timeline predictions unreliable regardless of who makes them. GPT-4, the most capable system available when this episode was recorded, produces impressive text and solves complex problems. But it fails on tasks that require genuine causal reasoning, struggles with novel problems that differ structurally from its training distribution, and has no persistent memory or ability to learn from experience after training. These are not engineering problems that more compute will solve; they point to architectural limitations. ### Embeddings and contextual representations Embeddings the conversion of discrete tokens into continuous vector representations are the foundation on which everything else is built. A word like "bank" maps to different regions of the embedding space depending on whether it appears in a financial or geographic context. This contextual sensitivity is what makes modern language models qualitatively different from earlier bag-of-words approaches. Contextual embeddings also enable retrieval-augmented generation and semantic search. By comparing the embedding of a query against embeddings of stored documents, a system can find semantically relevant results even when the exact keywords do not match. This is the technical basis for most modern AI-powered search systems. ### Where these threads converge BitNets address the deployment problem: how to run capable models on constrained hardware. Attention mechanisms and embeddings define what those models can do once deployed. AGI is the question of whether any amount of scaling these existing architectures will produce genuine intelligence, or whether something fundamentally new is needed. My assessment is that current architectures will continue to improve on benchmarks and practical tasks, but AGI in any meaningful sense of the term requires ideas that have not been invented yet. BitNets, meanwhile, are likely to have a more immediate and concrete impact by making current-generation models available on edge devices. ## OpenAI o1 benchmarks and AGI implications: IQ 120, coding breakthroughs, and what they mean Source: https://joelbarmettler.xyz/podcast/openai-o1-technical-analysis.md ## OpenAI o1's benchmark performance and the AGI question OpenAI o1 scores an estimated IQ of 120-125, up from the 96-98 range that characterized GPT-4 and comparable models. This is the first time a language model has tested above the human average on IQ-style reasoning benchmarks, and it raises the question of what these numbers actually tell us about the path to artificial general intelligence. ### Benchmark results in context On coding benchmarks and complex reasoning tasks, o1 surpasses every previous model by a significant margin. The gains are concentrated in problems that require multi-step deduction precisely the kind of task where the reflection mechanism adds value by allowing the model to reconsider intermediate steps before committing to a solution. The IQ comparison is illustrative but should be interpreted carefully. IQ tests measure a specific set of cognitive abilities, and a model that scores 120 on these tests has not necessarily crossed a threshold into general intelligence. It has become better at the particular kinds of pattern recognition and logical sequencing that IQ tests evaluate. For tasks outside this distribution open-ended creativity, common-sense reasoning about novel physical situations, learning from a single example the improvement over GPT-4 is far less dramatic. ### Inference-time compute versus architectural innovation The critical distinction is that o1's improvements come from spending more compute at inference time, not from a new architecture. The underlying transformer is essentially unchanged. What has changed is the amount of work the model does before producing output: instead of generating tokens immediately, it runs an extended internal reasoning process that can consume tens of thousands of tokens worth of computation. This is a meaningful advance, but it is a process innovation rather than an architectural one. The history of AI progress suggests that the largest capability jumps come from architectural breakthroughs backpropagation, the attention mechanism, the scaling laws that made large language models viable. Process improvements tend to produce diminishing returns as they are pushed further. Whether scaling inference-time reflection will continue to yield gains or will plateau is an open empirical question. ### Altman's 1000-day AGI prediction Sam Altman has framed o1's reflection capability as a step toward AGI, with a timeline of roughly 1000 days. The vision is that future versions will reflect not for 50 seconds but for hours or days, tackling complex research problems the way a human scientist would. There are several problems with this framing. First, extending reflection time does not guarantee proportionally better results. A model that reasons poorly for 50 seconds will not necessarily reason well given 50 hours the quality of the reasoning process matters more than its duration. Second, the problems that separate current AI from AGI are not primarily problems of insufficient thinking time. They involve capabilities like causal reasoning, genuine understanding of physical systems, and the ability to learn new abstractions from minimal data none of which are addressed by longer reflection windows. The 1000-day prediction reflects OpenAI's institutional optimism more than any technical argument. AGI predictions have a poor track record across the field, and the gap between impressive benchmarks and genuine general intelligence remains substantial. ### The reflection approach as a reproducible technique One consequence of o1's design is that the reflection approach is not proprietary in any deep sense. It does not depend on novel hardware or a fundamentally new training paradigm. Any team with access to a capable base model can implement inference-time reasoning through chain-of-thought prompting, reinforcement learning on reasoning traces, or similar techniques. This means o1's competitive advantage is likely temporary. Other providers and open-source projects will implement comparable approaches, and the differentiator will shift to training data quality, implementation efficiency, and integration into useful products. The innovation is important, but it is the kind of innovation that diffuses rapidly through the field. ## OpenAI o1 and the mechanics of self-reflection: how 70,000 hidden tokens change inference Source: https://joelbarmettler.xyz/podcast/openai-o1-self-reflection.md ## OpenAI o1's self-reflection mechanism and the cost of thinking OpenAI's o1, originally codenamed Strawberry, introduced a mechanism that no previous commercial language model had shipped: a built-in reflection phase where the model reasons internally for roughly 50 seconds before generating a visible response. The shift goes beyond producing better answers. It changes the fundamental economics and energy profile of inference. ### Hidden chain-of-thought at scale During its reflection phase, o1 processes between 10,000 and 70,000 tokens internally. Users see only a small fraction of this computation in the final response. The rest is a hidden chain-of-thought: the model exploring reasoning paths, reconsidering intermediate conclusions, and self-correcting before committing to an answer. This is architecturally distinct from earlier models, which begin generating output tokens immediately and have no mechanism for internal deliberation. The approach produces measurably better results on complex tasks programming challenges, multi-step mathematical proofs, scientific reasoning problems where the ability to reconsider an initial approach before outputting it is genuinely valuable. ### The personality requirement One of the more surprising findings from OpenAI's development process is that the reflection mechanism only worked well after the model was given what the developers describe as a "personality" a human-like way of approaching problems. Without it, the raw reflection process produced mediocre results. With it, o1 began exhibiting recognizably human reasoning patterns, including making the same kinds of logical errors humans make and finding similar paths to correct them. This is a noteworthy data point for AI research more broadly. It suggests that effective multi-step reasoning may depend on cognitive heuristics in addition to computational capacity. The kinds of shortcuts and framings that human thinkers use instinctively may be essential. ### Energy costs of extended reasoning The energy implications are substantial. A single o1 chat session consumes roughly as much electricity as running a living room lamp for a day and a half. This is a direct consequence of the token volume: processing 70,000 internal tokens requires the same GPU compute as generating 70,000 output tokens in a standard model, but none of that computation is visible to the user. For tasks where the reflection genuinely improves output quality, this trade-off may be justified. But for the 80-90% of everyday use cases translations, routine emails, simple questions o1 is no better than its predecessors. It is simply slower and more expensive. The right mental model is not "o1 replaces GPT-4" but "o1 is a specialist tool for hard problems." ### The model selection problem OpenAI's current product lineup o1, o1 mini, GPT-4, GPT-4o, and various other variants forces users to make a selection that most of them are not equipped to make. Choosing the right model for a given task requires understanding the performance characteristics and cost profiles of each option. A technology company building products for a general audience should not externalize this complexity to end users. The more practical design would be a routing layer that automatically dispatches queries to the appropriate model based on estimated task complexity. This is a solvable engineering problem, and it is worth noting that some competitors are already moving in this direction. ## AI utopia 2035: when automation funds a renaissance in human agency (part 2 of 2) Source: https://joelbarmettler.xyz/podcast/ai-utopia-2035.md ## AI utopia 2035 (part 2 of 2) The dystopian scenario in part 1 assumed that productivity gains from AI would accrue to capital owners while workers bore the costs. The utopian case rests on a different distribution mechanism: the same productivity gains, channeled through policy and institutional design, fund a broad expansion of human freedom rather than its contraction. ### From productivity to time sovereignty The most direct positive effect of AI automation is not higher output but fewer required working hours. If AI handles the routine analytical, administrative, and communicative tasks that currently fill a 40-hour week, the same economic output can be sustained in 20 to 30 hours. The four-day work week becomes structurally feasible, and in some sectors, the three-day week. The critical policy question is whether reduced labor demand translates into shorter hours or into unemployment. Countries with strong labor institutions and collective bargaining frameworks are better positioned to capture the gains as time rather than as layoffs. Historical precedent supports this path: the transition from the six-day to the five-day work week in the early 20th century was also driven by productivity gains, and it required deliberate political choices to realize. ### Community revival through reclaimed time In this scenario, the time freed by AI does not disappear into passive consumption. People reinvest it in their immediate communities: neighborhood associations, childcare cooperatives, local governance, volunteer work. Activities that the market economy has historically undervalued, caring for elderly neighbors, mentoring young people, maintaining public spaces, become feasible at scale because people have the hours to devote to them. The sociological evidence from existing reduced-work-week experiments, such as Iceland's 2015-2019 trial covering over 2,500 workers, suggests that shorter hours correlate with increased civic participation and reported well-being. ### AI as a universal knowledge translator Rather than making people passive consumers of pre-digested information, AI in this scenario functions as a translation layer between domains of expertise. A citizen preparing to vote on a complex infrastructure referendum can query an AI system that translates technical engineering assessments, environmental impact studies, and economic analyses into language calibrated to their level of background knowledge. The same mechanism works across natural languages and cultural contexts. The net effect is an expansion of informed participation in democratic processes, precisely the opposite of the filter-bubble collapse described in the dystopian scenario. ### Breaking echo chambers by design The dystopian scenario assumed that AI-driven personalization would fragment the information commons. The utopian alternative requires a deliberate design choice: AI systems that detect when a user is locked in an information bubble and introduce dissenting perspectives in calibrated doses. This is technically feasible. Recommendation algorithms already model user preferences with high accuracy; the same models can identify preference homogeneity and introduce diversity. The design decision is whether the optimization target is engagement (which favors echo chambers) or informed discourse (which requires exposure to disagreement). Platform regulation can mandate the latter. ### Open-source AI as public infrastructure The single most important structural condition for the utopian scenario is the democratization of AI itself. If capable models remain exclusively proprietary, the dystopian concentration of power is almost inevitable. The alternative is an open-source ecosystem where foundation models are publicly available, similar to how the internet's core protocols are open standards. Treating AI infrastructure as a public good, analogous to water or electricity, prevents any single company from becoming the indispensable utility layer of the global economy. The technical foundation for this already exists: open-weight models like LLaMA and Mistral demonstrate that competitive performance does not require proprietary control. ### The gap between technical possibility and political implementation Every element of this utopian scenario is technically achievable with current or near-term technology. The constraints are political and institutional. Shorter work weeks require labor policy reform. Community investment requires local governance structures. Knowledge democratization requires educational infrastructure. Echo-chamber mitigation requires platform regulation. Open-source AI requires sustained public funding for research and compute. None of these are technology problems. The technology is the easy part; the hard part is building the institutions that distribute its benefits broadly rather than narrowly. ## AI dystopia 2035: when AI becomes the lifeblood of the economy (part 1 of 2) Source: https://joelbarmettler.xyz/podcast/ai-dystopia-2035.md ## AI dystopia 2035 (part 1 of 2) Knowledge workers currently make up the majority of the Swiss workforce. By 2035, under pessimistic but technically plausible assumptions, most of their tasks could be automated. This episode constructs a dystopian scenario, not as prediction but as stress test, to examine where current AI trajectories lead if the worst-case dynamics play out unchecked. ### The hollowing out of knowledge work The first casualty in this scenario is the substance of white-collar employment. AI goes beyond eliminating jobs; it drains them of meaning. The roles that remain are supervisory: humans reviewing AI-generated reports, approving AI-drafted contracts, spot-checking AI-produced analyses. The creative and intellectually demanding components of work, the parts that provide professional identity and satisfaction, are handled by the machine. What is left is the cognitive equivalent of assembly-line oversight. The psychological toll of this shift is substantial: people trained for complex problem-solving find themselves reduced to approving outputs they can barely evaluate. ### Personalized information silos In the private sphere, AI-driven content generation creates a world where no two people see the same information. Every article, video, and social media post is generated in real time, tailored to the individual's psychological profile, preferences, and behavioral history. The shared epistemic commons that underpins democratic discourse disappears. Without common reference points, even basic political conversations become impossible because participants inhabit entirely different informational realities. People grow accustomed to the frictionless optimization of AI-mediated interaction; real human conversation, with its awkward pauses and misunderstandings, feels unsatisfying by comparison. ### Power concentration at the AGI bottleneck The most structurally dangerous element is the concentration of economic power. If a single company achieves a decisive AGI breakthrough, it becomes the utility layer of the entire economy. Every business, from logistics to healthcare to finance, depends on its API. That company becomes the "lifeblood" of global commerce, a position more powerful than any government or historical monopoly. The dependency is total: cutting off access to the AGI provider is not a competitive inconvenience but an existential threat to any business that relies on it. ### Technocratic governance by default Economic dependency translates into political power. The leadership of the dominant AI company, perhaps fifteen people on a safety or ethics board, makes decisions that affect billions without democratic mandate. Their choices about what the AI will and will not do, which use cases to allow, which values to encode, become de facto policy. Elected governments find themselves negotiating with, rather than regulating, a private entity whose economic leverage exceeds their own. AI systems can subtly shape public opinion at scale, making traditional democratic accountability mechanisms insufficient. ### The stagnation of human-driven research When AI systems operate orders of magnitude faster than human researchers, the motivation for independent inquiry erodes. A calculation that takes a human team months is completed in milliseconds. Scientific fields become dependent on AI for computation, hypothesis generation, experimental design, and interpretation. Humanity shifts from producer to consumer of knowledge. The long-term risk is a civilization that has lost the capacity to understand, let alone improve, its own foundational technologies. ### What this scenario gets right and wrong The technological premises of this dystopia are grounded in current trajectories. Individual elements, such as job displacement in knowledge work and power concentration among AI providers, are already observable in early form. The scenario is deliberately extreme as a composite, but its value lies in identifying which dynamics require active intervention. This thought experiment excludes science-fiction tropes like AI developing consciousness or pursuing independent goals. The threat it describes is entirely human: the concentration of transformative technology in too few hands, governed by too few people, with too little accountability. ## AI hype vs. reality: a technical assessment of where things actually stand Source: https://joelbarmettler.xyz/podcast/ai-hype-vs-reality.md ## AI hype vs. reality When coffee machines are marketed as "AI-powered," the term has clearly lost its descriptive precision. The same inflation has hit "metaverse" and "blockchain" before it, but with AI the gap between label and substance is unusually wide because the underlying technology is genuinely transformative. The challenge is separating the real advances from the noise. ### The investor feedback loop A significant driver of hype is the investor market. When OpenAI announces products like Sora months before general availability, the primary audience is not consumers but investors. These announcements are timed to influence valuations and funding rounds. The dynamic explains why AI startups routinely reach billion-dollar valuations despite offering products that are functionally similar to competitors. For the companies involved, the incentive is to promise aggressively and deliver incrementally, because capital allocation follows narrative momentum, not deployment metrics. ### Enterprise adoption is slower and healthier than headlines suggest In practice, the push for AI in organizations frequently comes bottom-up: employees who use ChatGPT at home bring their expectations to work. Companies themselves tend to approach adoption cautiously, running pilots and looking for measurable returns before scaling. This hesitancy is often criticized as falling behind, but it is a reasonable response to a technology whose capabilities and limitations are still being mapped. The productive question is not whether to adopt AI, but which specific workflows benefit most from it. ### Quality gains matter more than speed gains Most AI discussions center on efficiency: doing the same work faster. The more interesting effect, and the one I see most often in practice, is quality improvement. When routine tasks are automated, people spend more time on the parts of their work that require judgment, creativity, and strategic thinking. The output is both faster and genuinely better. This distinction matters for how organizations measure AI's return on investment. Tracking only time saved misses the larger benefit. ### The double standard for AI errors There is a persistent asymmetry in how we evaluate AI versus human performance. Self-driving cars must demonstrate near-perfection before the public accepts them, while human drivers cause roughly 1.35 million deaths per year globally, according to WHO data. AI systems do not need to be flawless to be valuable; they need to be demonstrably better than the alternative. Failing to apply this standard consistently slows adoption of technologies that would, on balance, reduce harm. ### The AI-first company is the real disruption The most consequential development in the next few years will be the emergence of "AI-first" companies that redesign entire industries around AI capabilities, rather than another foundation model. Consider an insurance provider that builds its actuarial models, claims processing, customer service, and fraud detection on AI from day one. It could operate at a fraction of incumbent costs while offering better service. When that company enters the market, the competitive pressure on traditional players will be immediate and severe. The parallel to early digitalization holds: not every company adopted email immediately, but eventually it became non-negotiable. The difference with AI is that the transition will compress into years rather than decades. ## Open-source AI: the infrastructure behind the hype Source: https://joelbarmettler.xyz/podcast/open-source-ai.md ## Open-source AI and the economics of foundation models Hugging Face hosts over 120,000 publicly available language models. While the public conversation about AI remains fixated on ChatGPT and a handful of proprietary systems, the open-source ecosystem has quietly become the infrastructure layer on which most applied AI research and a growing share of enterprise deployment actually runs. ### The cost of building a foundation model Training a frontier foundation model is a billion-dollar undertaking. Meta's investment in LLaMA-3 a 70-billion-parameter model and one of the most capable open-weight releases to date required approximately one billion dollars in data center construction and engineering. The power consumption during training was equivalent to the annual electricity usage of 4,000 Swiss households. These numbers explain why only a handful of organizations can afford to train foundation models from scratch: Meta, Google, Microsoft, and a small number of well-funded startups. Everyone else builds on top of what these players release. ### Why academia depends on open weights At the University of Zurich, the typical research group has fewer than ten GPUs. Training a foundation model is out of the question. Without access to open-weight models, academic AI research would be reduced to studying proprietary systems through their APIs black-box science with no ability to inspect architectures, modify training procedures, or reproduce results. Open-source models restore the basic requirements of scientific method: transparency, reproducibility, and the ability to build incrementally on prior work. The dependency runs deep: most published ML research now uses open-source models or frameworks as its starting point. ### Specialization through fine-tuning The most consequential property of open-source models is their adaptability. A general-purpose 70-billion-parameter model knows a little about everything. A 7-billion-parameter model fine-tuned on medical literature, legal documents, or financial reports can match or exceed the generalist's performance in its domain at a fraction of the inference cost. Fine-tuning is computationally cheap relative to pretraining: it typically requires hours or days on a small number of GPUs rather than months on thousands. This economics is what makes specialized AI applications viable for organizations that cannot afford GPT-4-scale API costs at production volume. ### Model ensembles and collective performance Recent research shows that combining several specialized open-source models into an ensemble can outperform a single large model like GPT-4 on domain-specific tasks. The approach is conceptually simple: route each query to the model most likely to handle it well, or aggregate responses from multiple specialists. The result is a system that exploits the strengths of each component while averaging out individual weaknesses. This "swarm" architecture is becoming a practical alternative to scaling a single model indefinitely. ### Licensing and the copyright problem The licensing landscape for open-source AI models is unresolved. Some models ship under permissive licenses like MIT or Apache 2.0. Others, including LLaMA-3, impose restrictions on commercial use above certain user thresholds. The deeper question is whether model weights which are derived from, but not copies of, copyrighted training data constitute a derivative work under copyright law. No jurisdiction has definitively answered this, and the legal uncertainty affects both model developers and downstream users. ### Local deployment and data sovereignty For enterprises in regulated sectors banking, healthcare, government sending proprietary data to a third-party API is often not an option. Open-source models that can be deployed on-premises or in a private cloud solve this problem. The organization retains full control over its data, avoids vendor lock-in, and can audit the model's behavior. The trade-off is operational complexity: running inference infrastructure requires engineering capacity that a managed API does not. But for organizations where data sovereignty is non-negotiable, open-source models are not an alternative to proprietary services they are the only viable option. ## Is AI intelligent? Why the question matters less than you think Source: https://joelbarmettler.xyz/podcast/is-ai-intelligent.md ## AI and intelligence: a question that answers itself A model like ChatGPT compresses terabytes of text into a few gigabytes of parameters a compression ratio that is only possible if the model extracts structure rather than merely storing data. Whether that counts as "intelligence" depends entirely on the definition, and that is where the conversation usually stalls. ### Seven criteria, mostly met The classical taxonomy lists seven components of intelligence: learning and memory, problem-solving, abstract thinking, adaptability, logical reasoning, linguistic ability, and creativity. Applied to current large language models, the scorecard is surprisingly strong. Linguistic ability is the most obvious: GPT-class models produce text that is indistinguishable from human writing in most contexts. Problem-solving and logical reasoning are demonstrated routinely on benchmarks models pass bar exams, medical licensing tests, and competitive programming challenges. Learning and memory operate differently from biological systems (weights are fixed after training, and context windows serve as volatile memory), but the functional output is comparable for many tasks. ### Compression as a proxy for understanding The compression argument is underappreciated. GPT-4 was trained on a corpus estimated at over 10 terabytes of text, yet the model's parameters occupy a fraction of that space. Lossless storage of the training data would require vastly more memory than the model itself uses. The only way to achieve this compression ratio is by learning generalizable patterns identifying that "the cat sat on the mat" and "the dog lay on the rug" share a structure, rather than memorizing each sentence independently. This is, at minimum, abstraction. Whether it constitutes "understanding" in a philosophical sense is a separate question, but it is not trivially dismissed as "just statistics." ### The "just statistics" objection The most common dismissal of AI intelligence is that language models are "just statistics" sophisticated pattern matching without genuine comprehension. The objection has some force: these models do not build causal world models, and they fail on problems that require reasoning outside their training distribution. But the objection proves too much. We do not know whether human cognition is fundamentally different in kind, or merely different in implementation. Neuroscience has not settled whether the brain operates on principles categorically distinct from statistical pattern recognition. Dismissing AI as "mere statistics" requires a theory of human intelligence that we do not yet have. ### Creativity and its boundary conditions Creativity is the criterion where intuitions diverge most sharply. Language models produce outputs that humans judge as creative: novel metaphors, unexpected problem solutions, plausible fictional scenarios. The question is whether an output that is perceived as creative is creative, or whether creativity requires subjective experience. The distinction may matter philosophically, but it has limited practical relevance. If a model generates a useful and non-obvious solution to an engineering problem, the solution's value does not depend on whether the model "experienced" the process of generating it. ### What current systems cannot do The limitations are real and worth specifying precisely. Current language models operate only on the modalities present in their training data. They have no sensory apparatus and no ability to interact with the physical world except through text or tool-use APIs. They cannot learn from new experience after training without explicit retraining or fine-tuning. Their "knowledge" is frozen at a cutoff date. They are prone to confident confabulation on topics outside their training distribution. These are not minor gaps they represent fundamental architectural constraints that separate current systems from anything resembling general intelligence. ### Capability over categorization The practical question is not whether AI is "really" intelligent but what it can reliably do and where it fails. A system that passes the bar exam but cannot reliably count the letters in a word has a specific capability profile that does not map neatly onto human intelligence categories. Treating AI as a different kind of cognitive system one with its own characteristic strengths and failure modes is more productive than trying to resolve whether it crosses some threshold of "real" intelligence. The definition was always the bottleneck, not the technology. ## AI in education: why bans backfire and what actually needs to change Source: https://joelbarmettler.xyz/podcast/ai-in-education.md ## AI in education and the failure of prohibition The word "delve" appeared in roughly 200 academic papers in 2020. By 2023, it showed up in over 800,000. This linguistic fingerprint of large language models reveals how deeply AI has already penetrated academic writing, regardless of institutional policy. ### The digital divide that bans create At the University of Zurich, where I studied Artificial Intelligence, ChatGPT arrived mid-curriculum. The institutional response was largely prohibitive. The problem with bans is straightforward: they are unenforceable and inequitable. Students who use ChatGPT despite the rules (for comprehension checks, code reviews, draft editing) gain a measurable advantage over those who comply. The ban does not eliminate AI use; it ensures that the benefit accrues unevenly and without transparency. ### What happens when knowledge is free If a language model can reproduce the factual content of a blockchain course in seconds, the course needs to justify what it teaches beyond facts. The pedagogical challenge is real: curricula built around knowledge transmission become redundant when that knowledge is freely available through an API. What remains valuable is the capacity for critical evaluation, novel problem formulation, and the ability to identify when an AI-generated answer is subtly wrong. These are skills that require a different kind of teaching. ### The primary school problem The situation in primary education is structurally different. A university student using ChatGPT to draft a literature review is automating a task they already know how to do. A primary school student using it to write an essay on Kafka's "The Metamorphosis" may never develop the underlying analytical skills. At this level, the product is not the point. The cognitive process of struggling with interpretation is. AI shortcuts that process entirely. ### Teachers as the bottleneck and the solution Teaching is already an intensely demanding profession: pedagogy, mentorship, administration, and emotional support compressed into insufficient hours. AI can relieve some of that pressure, particularly in grading and feedback. Automated first-pass assessment of written work could free hours per week for the interpersonal work that no model can replicate. The shift is from knowledge transmitter to learning facilitator (a role that requires more skill, not less). ### Rethinking assessment Traditional exam-based assessment assumes that the student produced the work independently. That assumption is now unfalsifiable for take-home assignments. A teacher who has observed a student over months or years has a far richer basis for evaluation than any single test. Qualitative, continuous assessment (regular conversations, portfolio reviews, observed problem-solving) is more robust against AI-assisted cheating and more informative about actual capability. AI itself can contribute here by providing real-time feedback on exercises, allowing students to iterate before a human evaluator ever sees the work. ### AI as a learning companion The most promising near-term application is AI as an always-available tutor. Current systems can explain a concept repeatedly from different angles, adapt to a student's pace, and provide immediate feedback rather than the weeks-long delay of traditional grading cycles. The constraint is that this works best when the student already has enough foundational knowledge to evaluate the AI's responses critically. This brings the argument back to the importance of building those foundations without shortcuts. ## Bias in AI systems: how 15 people shape the values of a billion-user product Source: https://joelbarmettler.xyz/podcast/bias-in-ai-systems.md ## How bias enters language models and why neutrality is impossible 1.8 billion people interact with ChatGPT daily. The ethical guardrails governing every response are set by roughly 15 members of OpenAI's alignment team. That ratio a tiny committee shaping the value system of a product used by a quarter of the world's internet users is the core tension in AI bias, and it has no clean resolution. ### How stereotypes propagate through training data Ask ChatGPT to write about different professions, and patterns emerge immediately. Nurses are almost always female. Construction workers are male. Executives default to white men in their forties. The model is not inventing these stereotypes; it is faithfully reproducing the statistical distribution of its training corpus, which is itself a reflection of decades of biased text published online. When 97% of nursing professionals in many countries are women, a model trained on descriptions of nurses will associate the profession with femininity. Whether this constitutes bias or accurate representation is the question that drives most of the debate. ### The correction dilemma OpenAI's response has been to intervene actively: the system is instructed to produce demographically balanced outputs in many contexts, overriding the statistical patterns in the training data. This solves one problem and creates another. Forced balance in generated content is an editorial choice, and it is not neutral it is a different kind of bias, one that substitutes the values of the alignment team for the statistical patterns of the real world. The intervention becomes especially problematic with historical content. When an image generation system is instructed to depict historical scenes from the era of slavery with demographic diversity, it may satisfy a contemporary norm, but it obscures the historical reality of who was oppressed and by whom. Good intentions do not prevent the distortion; they cause it. ### Open-source models and the question of authority While OpenAI and Google maintain strict content policies, a parallel open-source ecosystem has emerged that deliberately omits ethical constraints. Models like Llama can be fine-tuned and deployed without any alignment layer, producing outputs that commercial systems refuse to generate. These unconstrained models force a fundamental question: who has the legitimate authority to decide what values an AI system should encode? The commercial answer that the developing company decides is unsatisfying when a single product reaches billions of users across cultures with incompatible value systems. The open-source answer that no one decides, and users bear full responsibility is equally unsatisfying when the outputs can cause real harm. Neither position resolves the underlying problem that a general-purpose language model cannot be value-neutral. ### The geography of values Every major language model in widespread use was developed by a Western, primarily American, technology company. The values embedded in these systems attitudes toward free speech, gender, religion, political organization reflect that origin. A language model developed in China would handle questions about governance, censorship, and collective versus individual rights very differently. A model built in the Gulf states would encode different assumptions about gender roles, religious authority, and social hierarchy. This is not a hypothetical concern. As AI systems increasingly mediate access to information, the cultural perspective of the developing organization becomes a form of soft power, shaping how billions of users think about contested topics. The current Western monopoly on foundational models means one cultural framework is being transmitted globally, often without users recognizing it as a perspective rather than an objective default. ### Transparency as a partial remedy Perfect neutrality in AI systems is not achievable. Every design decision what data to train on, what outputs to suppress, what demographic distributions to enforce encodes a value judgment. The realistic goal is not neutrality but transparency: making the embedded values visible so that users can account for them. The analogy to media literacy is apt. Readers of newspapers learned, over decades, to identify editorial perspectives and adjust their interpretation accordingly. A similar literacy is needed for AI systems: understanding that ChatGPT is not an omniscient oracle but a system that reflects the priorities of its developers, the biases of its training data, and the cultural context of its origin. That understanding does not eliminate the bias, but it prevents users from mistaking one perspective for ground truth. ## AI and the labor market: autonomous agents and the transformation of knowledge work Source: https://joelbarmettler.xyz/podcast/ai-and-the-labor-market.md ## Autonomous agents and the future of knowledge work Goldman Sachs estimated in 2023 that roughly 300 million jobs worldwide are exposed to automation by large language models. That figure, already striking, was calculated before autonomous AI agents (systems that go beyond text generation to take actions) entered the picture. The labor market disruption ahead is qualitatively different from anything in the industrial era, because for the first time, automation targets cognitive work rather than manual labor. ### Three years of AI integration, condensed The speed of change is visible in a single role. In 2022, an event manager worked entirely by hand: brainstorming sessions, manual vendor research, individual emails to service providers. In 2023, ChatGPT entered as a support tool. The work became more efficient but not less. The manager had to develop new skills in prompt engineering and output verification, and the total scope of the role remained the same. By 2024, the first autonomous agents appeared: AI systems with access to company databases that can draft communications, compare vendor proposals, and prepare decisions without step-by-step human instruction. The event manager's role shifted from executing tasks to coordinating agents (reviewing their outputs, making final decisions, handling exceptions). The underlying work did not disappear, but who performs it changed fundamentally. ### Why this wave is different Previous automation waves followed a consistent pattern: machines replaced manual labor, and displaced workers moved into cognitive roles. The mechanization of agriculture pushed workers into factories; factory automation pushed workers into offices. Each transition was painful, but there was always a higher-abstraction domain to absorb the displaced workforce. AI breaks this pattern. It automates cognitive tasks directly: writing reports, analyzing data, drafting legal documents, generating code. The question that previous transitions never had to answer is now unavoidable: where do knowledge workers go when the knowledge work itself is automated? There is no obvious higher-abstraction tier waiting. ### The inefficiency that sustains employment A significant fraction of knowledge work consists of solving the same problem repeatedly across different organizations. Every company writes its own onboarding materials, its own internal policies, its own project plans often reaching similar conclusions through redundant effort. AI can solve a problem once and deploy that solution at near-zero marginal cost, eliminating the inefficiency that currently sustains millions of jobs. This is not inherently negative. Much of the work being eliminated is tedious and repetitive even if it is classified as "knowledge work." But the scale is unprecedented: when a single AI system can replace the redundant cognitive labor performed across thousands of organizations simultaneously, the displacement happens faster than retraining programs can respond. ### The timeline and the transition window Fully autonomous management agents (systems that can run a department with minimal human oversight) are plausibly 5 to 10 years away. The technology is advancing faster than the organizational willingness to adopt it, which creates a buffer. Regulatory uncertainty, liability concerns, and institutional inertia all slow deployment. This gap between technical capability and actual adoption is the transition window. It is the period in which societies can restructure work arrangements, retrain workers, and rethink the assumption that full-time employment is the primary mechanism for distributing economic output. Whether that window is used productively is a policy question, not a technical one. ### Possible equilibria One plausible outcome is a revaluation of work that AI cannot do. If cognitive office work is automated, professions that require physical presence, manual skill, or deep human interaction (trades, caregiving, teaching, craftsmanship) may gain relative status and compensation. The four-day work week, currently a fringe proposal, could become an economic necessity if there is simply less cognitive work to distribute among the workforce. The less optimistic scenario is a polarized labor market: a small class of AI-system operators and owners capturing most of the economic value, with a large displaced workforce lacking viable alternatives. Which outcome materializes depends almost entirely on institutional choices made during the transition window. These are choices about taxation, education, social safety nets, and the distribution of productivity gains. ## AI terminology explained: a technical guide beyond the hype Source: https://joelbarmettler.xyz/podcast/ai-terminology-explained.md ## AI terminology beyond the hype Most of what passes for AI vocabulary in public discourse is marketing language grafted onto engineering concepts. The gap between how these terms are used in press releases and what they mean in practice is wide enough to cause real confusion, so this episode with Marius Högger sorts through the core terminology. ### Algorithms versus learned models A classical algorithm is a fixed sequence of instructions: given input X, perform steps A, B, C, return output Y. In image recognition, a traditional algorithm would search for predefined features edges, shapes, contrast ratios to identify an object like a house. The rules are hand-coded by the developer. An artificial neural network works differently. Instead of explicit rules, it uses layers of artificial neurons connected by weighted edges. Training the network on thousands of labeled images adjusts those weights until the system learns to extract the relevant features on its own. The developer never specifies what a house looks like; the network discovers that from the data. This distinction hand-coded rules versus learned representations is the dividing line between classical programming and machine learning. ### Where machine learning fits within data science Data science is the discipline of extracting knowledge from data: collecting it, cleaning it, analyzing it, and drawing conclusions. Machine learning is one tool in that toolkit, not the whole field. A data scientist predicting real estate prices might start with a simple regression on historical trends. When the model needs to account for location, floor area, amenities, and macroeconomic indicators simultaneously, machine learning becomes the more effective approach because it handles high-dimensional feature spaces that overwhelm classical statistical methods. The confusion between the two terms matters because it leads organizations to deploy ML where a straightforward statistical analysis would suffice, and to underinvest in data quality the part of the pipeline that determines whether any model, classical or learned, will actually work. ### From machine learning to artificial intelligence The boundary between ML and AI is blurry, but a useful distinction exists. Machine learning systems typically excel at a single, narrowly defined task: classifying images, predicting churn, detecting fraud. AI systems like ChatGPT operate across many tasks analyzing text, writing code, drawing inferences, composing poetry using a single underlying model. That versatility is what earns the label "intelligence," though the term remains contested. Artificial General Intelligence, the hypothetical system that matches or exceeds human cognition across all domains, remains out of reach. Current AI systems are powerful within their modality text, image, audio but they cannot transfer learning across modalities the way humans do. A language model that writes excellent prose cannot learn to navigate a physical environment without an entirely separate training pipeline. ### The naming problem Terminology shapes perception, and the labels attached to AI products often distort more than they clarify. The word "chatbot" carries the baggage of early rule-based systems that matched keywords to canned responses. Calling ChatGPT a chatbot obscures the fact that it runs a 175-billion-parameter transformer. The suffix "-bot" implies something mechanical and scripted, which is the opposite of what these systems do. Labels like "assistant," "agent," and "copilot" are more accurate. "Copilot" in particular positions the system as a collaborator rather than an autonomous actor, which aligns well with how most productive deployments actually work: a human setting direction, the model handling execution. ### What ChatGPT computes Under the hood, ChatGPT is a mathematical function. It takes a sequence of tokens as input and outputs a probability distribution over the next token. The token with the highest probability is not always selected; a parameter called temperature controls how much randomness is injected into the sampling process. Higher temperature produces more varied, less predictable outputs; lower temperature makes the model more deterministic. This mechanism explains both the system's strengths and its failure modes. It generates fluent, contextually appropriate text because it has learned statistical regularities across a massive training corpus. It hallucinates because the same statistical machinery can assign high probability to plausible-sounding but factually wrong continuations. There is no internal fact-checking step, no world model, no understanding just a very large, very well-trained next-token predictor. ## AI and democratic manipulation: from Cambridge Analytica to language models Source: https://joelbarmettler.xyz/podcast/ai-and-democracy.md ## AI-powered election manipulation The Cambridge Analytica scandal of 2016 showed what targeted political manipulation looks like at scale: Facebook profile data fed into personality models, messages personalized down to the neighborhood level, all in service of Donald Trump's campaign. At the time, this felt dystopian. In hindsight, the operation was labor-intensive and crude compared to what language models make possible today. ### From manual targeting to automated persuasion Cambridge Analytica needed large teams to analyze data and craft messages for different audience segments. Modern language models collapse that pipeline into a single API call. A political operator can now generate thousands of personalized messages per minute, each tailored to the recipient's known concerns, reading level, and emotional triggers. The cost per contact has dropped by orders of magnitude, and the barrier to entry has fallen with it you need API access and a voter file, eliminating the requirement for a data science team. The shift is both quantitative and qualitative. Language models can engage in real-time conversations, respond to counter-arguments, and adapt their tone mid-exchange. A bot farm running on GPT-class models can sustain thousands of seemingly authentic social media personas simultaneously, each with a consistent posting history and personality. ### Three technical capabilities that change the threat model The first is conversational authenticity. Current language models produce text that is indistinguishable from human writing in most online contexts. A single operator can run thousands of convincing online personas that respond contextually to comments and participate in discussions. The second is multimedia synthesis. When language models are combined with image generation and voice cloning, the result is forged multimedia content of a quality that did not exist five years ago. A fabricated video of the Swiss Federal President making a controversial statement is now technically trivial to produce. The third is the speed asymmetry between creation and verification. Generating disinformation is now faster than debunking it. This asymmetry fundamentally challenges traditional fact-checking mechanisms, which were designed for a world where producing convincing fakes required significant effort. ### Switzerland's specific vulnerability Switzerland's direct democracy faces a particular version of this problem. Citizens vote on complex policy questions several times a year, and the system's legitimacy depends on voters being reasonably well-informed about each issue. If the information environment around a referendum is poisoned by AI-generated content misleading arguments, fabricated expert opinions, fake grassroots movements the democratic process is compromised at its core. The federal structure and consensus-oriented political culture provide some resilience. Political messaging must work across 26 cantons with distinct media landscapes, which makes a single unified disinformation campaign harder to execute. But this same fragmentation also makes coordinated defense more difficult. ### Technical countermeasures and their limits On the detection side, watermarking and digital signatures can help establish content provenance. Cryptographic watermarks embedded in AI-generated text or media could make it possible to verify whether content was machine-generated. The challenge is adoption: watermarking only works if all major model providers implement it, and open-source models can be run without any watermarking. AI-based detection systems classifiers trained to distinguish human from machine-generated text are in an arms race with the generators. Current detectors achieve reasonable accuracy on older models but struggle with newer ones, and false positive rates remain a concern. ### The regulatory gap The EU AI Act represents the most comprehensive attempt at AI regulation so far, but its multi-year development cycle illustrates the fundamental timing problem: by the time regulations are enacted, the technology has moved on. Enforcement across borders adds another layer of difficulty. A disinformation campaign targeting Swiss voters can be operated from any jurisdiction, and the infrastructure cloud compute and API access is globally available. Digital media literacy is often proposed as a complement to regulation. The idea is sound in principle: citizens who understand how AI-generated content works are better equipped to evaluate what they read. But scaling this to the general population, in time for the next referendum, is a different problem entirely. ## The Vibe Coding Experiment: building a RAG system from scratch at ch-open Source: https://joelbarmettler.xyz/appearances/ch-open-vibe-coding-experiment.md ## The Vibe Coding Experiment: a RAG system in six hours On Monday, 18 May, I hosted a workshop at the ch-open Open Source AI Workshops in Room C. The premise was a question: can a team with no template build a working RAG system in six hours, armed only with an open-source coding agent and an API specification? We set out to find out. ### The setup Four teams each took ownership of one core component of a retrieval-augmented generation system: an LLM inference server, a vector database, a document processor, and a chat frontend. Every team built its component from scratch, without finished frameworks, using only base libraries and opencode-ai as their copilot. At the end, we wired everything together live into a single working system. The whole stack stayed open. We worked throughout with open technologies: opencode-ai as the coding agent, open-weights models such as DeepSeek or GLM-4, and a final system that ran on a local open-source LLM. No proprietary APIs, no hidden magic, just the moving parts of a RAG system laid bare. ### The format The day was split into four sessions: - **Foundation** — understanding the system architecture, an introduction to opencode-ai, and forming teams. - **Understand & plan** — each team worked out its component and prepared a short presentation. - **Build** — implementing the core functionality and sharing an interim status. - **Integrate** — bringing the components together, running a live demo, and reflecting on what we learned. ### For whom The workshop was for anyone who wants to understand how AI systems work under the hood and try AI-assisted programming with open-source tools. You did not need to be an experienced developer; the willingness to experiment mattered more. The only prerequisites were a laptop with a browser and a GitHub account, since everything else ran in a preconfigured GitHub Codespace. Participants walked away with hands-on experience of vibe coding, a working understanding of the core components of a RAG system, and their own code to keep building on. ### Slides ## AI trends 2025 and predictions for 2026: model convergence, integration, and sovereignty Source: https://joelbarmettler.xyz/appearances/webinar-2025-rewind-2026-outlook.md ## AI in 2025/2026: rewind and outlook ### 2025 rewind #### Model convergence On established benchmarks (GPQA, MMLU, HumanEval), the ceiling was reached in late 2024. New models appeared throughout 2025 but no longer pushed the frontier; they matched it with smaller, more efficient architectures. Google's Gemma 3 N4B, designed to run offline on a smartphone, scored as high on HumanEval as GPT-4O had eighteen months earlier. Intelligence continued to increase on newer, harder benchmarks, with IQ-test-style evaluations showing gains of 20 to 30 points, but for everyday tasks, the difference was imperceptible. All frontier closed-source models (GPT-5, Claude Opus, Gemini) clustered at the same performance point; open-source models trailed by roughly six months. #### USA versus China Two days after our previous webinar in January 2025, DeepSeek released its model. US big tech spent over 100 billion dollars per quarter on AI infrastructure, betting that more compute leads to better models. China released frontier models as open weights, published the research behind them, and iterated at two to three state-of-the-art releases per month by year-end. The cost of interacting with a model at a given intelligence level dropped from $32 per million input tokens in January to $0.10 by December (a factor of 300). Chinese labs never produced the absolute top-scoring model, but consistently came within a few IQ points at a fraction of the cost. #### Platform as differentiator With model intelligence converging, the differentiator shifted to the platform layer. OpenAI's major 2025 announcements were platform features: native UI components in ChatGPT and agentic workflow builders. Google's advantage became structural. Gemini integrates with Gmail, Calendar, Drive, and YouTube natively. Any platform could technically connect to any model; the bundling is a business decision, not a technical constraint. #### Production-quality generation Image generation became indistinguishable from real photography, even for arbitrary individuals given a single reference image. Video generation crossed a similar threshold. In software engineering, the lead developer of Claude Code reported that 100% of his contributions over the past 30 days were written by the tool itself. In our team at bbv, none of us has written a complete function by hand in months. The job has moved from writing code to managing coding agents: formulating requirements, reviewing output, maintaining documentation, and iterating on permanent instructions to correct recurring mistakes. #### AI slop In 2025, 51% of internet traffic came from bots, 15% of Reddit posts were demonstrably AI-generated, and 18% of website content was AI-produced. On GitHub, 300 million new repositories appeared in a single year, and at least 41% of all code on the platform is now AI-generated. AI models cite AI-generated web pages as sources, users trust the citations without checking, and fabricated information acquires an appearance of legitimacy. During the Venezuela crisis, AI-generated images of events that had not yet occurred were shared by US senators as fact. ### 2026 outlook #### AI enters daily work Software engineering has crossed the productivity threshold. Other industries follow as AI now covers text, image, video, and code generation at production quality. #### Business integration through MCP AI platforms work well in their native ecosystems (Gemini with Google Workspace, Copilot with Microsoft 365), but most enterprises use custom ERPs, on-premise databases, and proprietary tools these platforms cannot reach. The Model Context Protocol, originally developed by Anthropic and now governed by the Linux Foundation, standardizes how AI platforms communicate with business software. Unlike REST APIs, MCP servers describe their capabilities in natural language, allowing language models to autonomously decide which tools to invoke. OpenAI, Anthropic, and Google all support it. Thousands of public servers already exist. #### Platform lock-in With models interchangeable and MCP standardizing tool integration, the remaining competitive moat is lock-in. Passive lock-in accumulates through chat history, memory systems, and audit logs. Active lock-in comes from enterprise investments: vectorized knowledge bases, custom agent workflows, and compliance configurations. None of these transfer between platforms. #### Digital sovereignty US tech dominance, export controls on Chinese chips, and rising subscription costs are driving enterprises toward digital sovereignty. Open-source alternatives exist at every layer: open-weight models hosted by Swiss providers like Infomaniac or Exoscale, open-source platforms deployable on-premise, and MCP servers for Swiss business integrations like Bexio. ## Swiss AI Impact Forum 2025: live demos of the Swiss AI Hub Source: https://joelbarmettler.xyz/appearances/swiss-ai-impact-forum-2025.md ## Swiss AI Impact Forum 2025: Swiss AI Hub booth At the 2025 forum, bbv presented the Swiss AI Hub at a dedicated booth with live demonstrations. The Swiss AI Hub is an AI platform for on-premise deployment: authentication, monitoring, vector databases, data pipelines, LLM gateway, and chat interfaces in a single package, deployable with one command. The demos showed the full pipeline: ingesting company documents, vectorizing them, and querying them through RAG agents all running on infrastructure that never leaves the organization's network. The platform uses open-source components (LiteLLM, Milvus, LlamaIndex) under Apache 2.0, runs on Swiss data centers or fully on-premise, and supports MCP for connecting to existing business software. ## AI trends 2024 and predictions for 2025: a technical analysis Source: https://joelbarmettler.xyz/appearances/webinar-2024-rewind-2025-outlook.md ## Key developments of 2024 and what 2025 may bring In this webinar I walked through the AI developments that defined 2024 and outlined where I see 2025 heading. ### Dominant trends in 2024 Large language models continued to dominate. Despite AI being a much broader field, LLMs and ChatGPT became near-synonymous with "AI" in business contexts as companies recognized their practical utility. OpenAI remained the most influential player, shipping GPT-4O and Sora for video generation. Notably, OpenAI does not necessarily build the best models, but it dominates the discourse through marketing reach and broad availability. Many companies introduced AI assistants in 2024. These tools support employees on tasks from email drafting to coding, though many organizations left their staff without adequate training for the new tooling. Costs for AI services dropped by roughly a factor of 1,000 compared to prior years. Tasks that cost 60 dollars in 2023 cost a few cents in 2024, unlocking use cases that were previously uneconomical. The open-source community made strong progress. Models like Mixtral approached the quality of proprietary systems at a fraction of the cost, threatening the incumbents' moat. ### Outlook for 2025 AI systems will increasingly interact directly with computer interfaces, operating a desktop the way a human would and opening new automation paths. A new generation of reasoning models will spend more computation on deliberation, handling more complex problems at higher per-query cost. The industry is shifting from passive assistants to autonomous agents that pursue goals independently and escalate to humans only when needed. Enterprises will begin redesigning processes from scratch around AI agents rather than bolting automation onto existing workflows. Chinese AI companies will enter global competition more aggressively; their open-source models already show strong performance at lower cost. The AGI debate will intensify as various companies claim to have reached the milestone, though the practical impact will likely be less dramatic than predicted. ## AI as a development partner: tools, techniques, and team integration Source: https://joelbarmettler.xyz/appearances/webinar-ai-development-partner.md ## AI as a development partner Four members of the bbv AI Hub development team Thomas (consultant), Michelle (engineer), Noah (junior engineer), and I (architect) each presented how AI has changed their role. Marius Högger moderated. ### Requirements engineering with AI Thomas dictates observations in Swiss German into the AI Hub's speech-to-text function immediately after customer workshops, capturing unstructured thoughts while the room's posters and post-its are still visible as memory aids. An AI assistant transforms that transcript into structured epics and requirements with acceptance criteria. Photos of post-it walls are fed directly to the model, which clusters and summarizes them. The same pipeline generates Mermaid diagrams and unit test skeletons from the resulting requirements. The models also perform gap analysis: given a set of requirements, they propose missing non-functional requirements and flag contradictions. ### Coding: context is everything Michelle identified context management as the central challenge in AI-assisted coding. Context overload feeding the entire codebase or all open files produces vague, unspecific output. Context starvation missing project guidelines or language conventions produces generic code that doesn't fit. She manually selects exactly which code snippets to include, opens a new chat when switching topics, and disables web search by default (enabling it only for specific needs like tracking down library bugs). For recurring projects, she creates dedicated AI assistants with pre-loaded coding guidelines, saving the repetitive setup of each conversation. ### Onboarding as a junior developer Noah joined the team at the start of the year and used AI tools extensively to navigate an unfamiliar codebase. He connects the codebase to the AI Hub and queries it for abstract concepts "how do we handle data persistence?" letting the model search across all files and return both an explanation and the relevant source locations. For learning new concepts outside the codebase, he prompts the agent with a knowledge-tree metaphor: start with foundational concepts (trunk), proceed to specifics (branches and leaves), and do not advance until he confirms understanding. He also uses a GitHub Actions integration that runs an LLM review on pull requests before human review, catching security issues and giving an early quality signal. ### Architecture research across abstraction levels My workflow operates across three abstraction levels, each requiring different tools. Deep Research (Gemini, GPT) surveys the landscape of available technologies for a given problem. I constrain it to trusted sources arXiv for papers, specific subreddits for practitioner experience, GitHub for source code to avoid low-quality web content contaminating the results. One level down, an IDE-integrated agent like Junie dives into specific open-source codebases to verify whether a library actually supports what its documentation claims. This runs asynchronously: I issue a question, continue other work, and review the findings minutes later. At the most specific level, I use a large-context model (Gemini with 1M+ tokens) and feed it the entire AI Hub codebase condensed via Repomix, which strips implementation details but preserves function signatures. With the full project in context, the model identifies exactly where a new component should be integrated. For actual code generation, I switch to Claude Sonnet with a narrow, manually curated context the large-context overview is too broad for precise code output. ## Swiss AI Impact Forum: Panel on the future of AI in Switzerland Source: https://joelbarmettler.xyz/appearances/swiss-ai-impact-forum-2024.md ## Swiss AI Impact Forum: panel with leading figures in Swiss AI I joined a panel at the Swiss AI Impact Forum alongside three specialists whose work spans entrepreneurship, ethics, and behavioral science. ### The panel Dalith Steiger-Gablinger, co-founder of SwissCognitive and a mathematician by training, brought a global perspective on AI ecosystems and entrepreneurial strategy. Prof. Dr. Peter G. Kirchschläger, ordinary professor of theological ethics and visiting professor at the ETH AI Center since 2023, contributed his work on the ethical dimensions of AI; he advises the UN, UNESCO, and the EU on ethics. Prof. Dr. Steffen Müller, head of behavioral insights and pricing at ZHAW, provided a research lens on how AI changes marketing analytics and consumer behavior. ### My role As the practicing AI engineer on the panel, I grounded the strategic, ethical, and scientific viewpoints in day-to-day engineering reality: what current models can and cannot do, where deployment bottlenecks actually sit, and how technical constraints shape responsible adoption. ### The discussion The combination of academic research, entrepreneurial experience, ethical inquiry, and hands-on engineering produced a conversation that moved fluidly between what AI should do and what it currently can do. The exchange between ethical boundaries and technical capabilities was particularly productive, surfacing concrete tensions rather than abstract principles. ## AI in knowledge management: keynote at the SWICO event in Zurich Source: https://joelbarmettler.xyz/appearances/swico.md ## AI in action: keynote at the SWICO event ### Demystifying AI On 18 April I spoke at a SWICO event in Zurich to an audience of technology-industry leaders. The topic was AI in knowledge management. My goal was to strip AI of its mystique and give a concrete account of how the technology works and where it applies. In the Luma Westbau venue, surrounded by floor-to-ceiling bookshelves and art installations, I walked through the mathematical foundations of AI and machine learning: linear regressions, gradients, and vectors, showing that AI rests on statistical methods structurally analogous to biological neural networks. ### AI deployment in practice Using a map visualization, I demonstrated how data selection shapes AI model outputs, specifically in relation to political bias. This underscored the need for critical scrutiny of training data. I also addressed the constraints facing smaller companies, which are often limited by the compute requirements of large models. I discussed current research on making models more efficient and accessible without sacrificing performance, a prerequisite for better handling data privacy and security concerns. ### Discussion The event included an open discussion during the apero, briefly interrupted by a fire alarm. Participants raised pointed questions about data privacy and the manipulation potential of AI systems. I emphasized the importance of incremental adoption and non-negotiable ethical standards. ### Links Netzwoche published a > write-up of the event. ## Swiss AI Conference: hands-on workshop on AI agents in the enterprise Source: https://joelbarmettler.xyz/appearances/swiss-ai-conference.md ## Swiss AI Conference: AI agents hands-on On 5 March 2024 I co-hosted a workshop with Dr. Alan Ettlin at the Swiss AI Conference in Rotkreuz, Zug. The conference was organized by Switzerland Innovation Park Central and focused on practical AI adoption for enterprises. Our session was titled "ChatGPT & Co. in the enterprise done right: experience AI agents hands-on." ### Workshop format The workshop introduced participants to AI agents as specialized assistants that use generative AI to support employees on concrete tasks. Rather than lecturing, we gave attendees direct access to the bbv AI Hub and let them work through a challenge: pick an organization you know, then use the available agents to identify high-value generative-AI use cases for it. ### The agents We built several specialized agents for the session, each created in a matter of hours to demonstrate how quickly purpose-built agents can be stood up. A Summary Agent condensed large documents into focused briefs. An SME Analyst conducted interactive dialogues to build a profile of a company's industry, value chain, and revenue model, and could pull information from URLs the user provided. A Business Report Agent did similar profiling but could ingest uploaded PDFs, Word files, and other documents. An AI Hub Expert helped participants map their identified use cases onto the capabilities of the bbv AI Hub platform. ### Multi-agent collaboration The most advanced setup was the Focus Group, a group chat where all agents could interact with each other and with the user. Participants could address specific agents with an @-mention or broadcast a question to all of them. Agents could also call on each other autonomously. A coordinator agent managed the dialogue flow, prompting agents to contribute when their expertise was relevant. A stop function let users pause runaway exchanges, edit the last message, and resume. ### The challenge Participants chose an organization they knew well their employer, a client, a nonprofit and worked through three questions with the agents. What does the organization do and how does it create value? Where can generative AI and AI agents create benefit? Can the bbv AI Hub contribute? The format turned abstract AI-agent concepts into a tangible, self-directed experience. ## AI trends 2023: milestones and developments in artificial intelligence Source: https://joelbarmettler.xyz/appearances/webinar-2023-rewind.md ## AI trends and developments: a review of 2023 This webinar covers the generative-AI developments that defined 2023, along with their economic and legal consequences. ### Major milestones The year opened with Microsoft's 10-billion-dollar investment in OpenAI. Microsoft integrated OpenAI's technology across its product line and became the exclusive cloud provider for OpenAI's APIs. GPT-4 launched in March and outperformed its predecessor across nearly all benchmarks, scoring notably on standardized tests like the GRE and the US bar exam. ### The open-source revolution Meta accidentally released its LLaMA model. What looked like a disaster became a catalyst: researchers worldwide used LLaMA as a base for their own work, producing open-source alternatives like Stable Vicuna. By year-end, the French startup Mistral AI shipped a model with a novel architecture that matched much larger systems in efficiency. ### Multimodal breakthroughs ElevenLabs pushed speech synthesis to the point where it could convincingly replicate nearly any voice. Midjourney's image generation reached a quality level that won photography competitions. ### Economic and legal consequences Goldman Sachs estimated that GPT-4-class systems could affect 300 million jobs. The New York Times sued OpenAI for using copyrighted content as training data, a case with potentially broad implications for future model development. ### Switzerland's position ETH and EPFL launched the Swiss AI Initiative, combining vertical projects in health and education with horizontal basic-research programs. Switzerland also secured NVIDIA GPUs early for the Lugano supercomputer, at a time when prices were still moderate, giving it a head start in European AI compute. ### Outlook for 2024 I expected the pace of breakthroughs to slow as training data limits were reached (GPT-4 had already consumed most available data). Progress would shift to improved multimodality, more efficient architectures, and specialized industry applications. The gap between proprietary and open-source models would continue to narrow. ## KI Revolution: AI first how a digital native thinks about generative AI Source: https://joelbarmettler.xyz/appearances/bbv-ki-revolution.md ## KI Revolution: AI first On 27 September 2023 I delivered the opening keynote at bbv's "KI Revolution" event. The talk, titled "AI first how a digital native thinks about generative AI," framed the shift from classical AI to generative AI and introduced the agent paradigm as the bridge between a language model and real enterprise value. ### Classical AI vs. generative AI Classical AI (pre-2022) solves narrowly defined problems. YOLO detects 90 object classes in an image; the business question is "does YOLO solve my problem?" Generative AI flips this: ChatGPT has broad capabilities but solves no specific problem out of the box. The business question becomes "where would a new worker help me, and can generative AI be that worker?" This reframing moves AI from the algorithm category into the workforce category. ### The agent paradigm An AI agent maps directly onto a human employee. A human has a job description; an agent has a task. A human has intelligence; an agent has ChatGPT for text understanding and reasoning. A human has a profile and skills; an agent has a character definition and tools (reading files, sending emails, querying databases). This analogy makes the technology accessible to non-technical decision-makers. ### Four-stage maturity model I presented a four-stage model for enterprise AI adoption. Stage one, instruction: correctly prompting and configuring the model via system prompts, user prompts, and parameter tuning (temperature, etc.). Stage two, context: connecting the agent to internal knowledge bases through retrieval-augmented generation, so answers draw on domain-specific data. Stage three, tools: letting the agent interact with enterprise systems through APIs, reflect on which tools to use, and iterate until the task is solved. Stage four, autonomy: the agent receives a high-level goal and independently plans, executes, adapts its strategy, and iterates, calling on other agents or humans as needed. ### Outlook I closed with a set of predictions. ChatGPT would gain image understanding within a month (it did). Further modalities would follow within a year. Within two years, some companies would employ many virtual workers. Within three years, companies would divide into those using AI and those not. Within five years, only one group would remain. ## AI agents: the future of enterprise automation Source: https://joelbarmettler.xyz/appearances/netzwoche.md ## AI agents in the enterprise Together with my colleague Marius Högger, I wrote an article for the online magazine Netzwoche on AI agents as enterprise assistants. The piece examines how generative models like ChatGPT become far more capable when wrapped in an agent architecture that gives them tools, memory, and a defined role. ### What AI agents do An AI agent is a virtual worker with a character profile and a tool set. It can draft social media posts, answer internal knowledge queries, or orchestrate multi-step business workflows. The key distinction from a bare chatbot is autonomy: an agent decides which tools to invoke and when to escalate. ### Three maturity levels We defined three levels. Basic agents handle constrained tasks using preset parameters and style guides. Intermediate agents access internal knowledge bases and the web, choosing sources autonomously. Advanced agents interact with external applications and databases, iterating toward solutions for complex, multi-step problems. ### Enterprise value AI agents break down knowledge silos by exposing enterprise data through a single conversational interface, governed by role-based access controls and security protocols. Automation of routine processes frees teams for strategic work. The article made the case that AI agents are not a speculative technology but a deployable one, provided organizations invest in proper access controls, monitoring, and incremental rollout. My work at bbv Software Services focused on exactly this kind of implementation. ## ChatGPT demystified: technical deep dive into large language models Source: https://joelbarmettler.xyz/appearances/webinar-chatgpt-demystified.md ## ChatGPT demystified This webinar strips ChatGPT down to its mathematical foundations. The goal is to show that behind the apparent magic sits statistics, albeit very clever statistics. ### Machine learning fundamentals Every ML system is a mathematical function mapping input to output. The simplest case is linear regression, y = ax + b, where training means adjusting a and b to fit known data points. That is the entire conceptual core; everything else is scale and architecture. ### From ML to AI Neural networks generalize this idea. They chain many simple functions (neurons) together in layers, and they can approximate virtually any mathematical function without anyone specifying its form in advance. The network discovers the mapping from data alone. ### Large language model architecture ChatGPT is a large language model whose job is to predict the most probable next token given a sequence of prior tokens. It operates over a vocabulary of roughly 32,000 tokens. The key architectural innovation is the attention mechanism, which lets the model weigh which input tokens matter most for each prediction. When the input contains "On a bank you can...", attention resolves whether "bank" means a bench or a financial institution from surrounding context. ### Training and data sources The model was trained on scientific papers, books, web pages, and source code. To turn a raw language model into a chatbot, OpenAI added conversational data from Reddit and from dialogues manually written by paid annotators. The choice of which data to include directly shapes the model's behavior: including or excluding certain Reddit communities shifts how the system responds to political and social questions. Analyses show that ChatGPT leans liberal-progressive, reflecting Silicon Valley's prevailing culture. ### Practical takeaways ChatGPT is a mathematical function. It does not think or feel. It cannot learn during a conversation; every response draws on the original training. Its core operation is text-to-text transformation; features like PDF processing are application-layer additions. Outputs should always be verified, especially on contested topics, because the model reproduces biases encoded in its training data. ## Swarm intelligence and AI: the future of enterprise automation Source: https://joelbarmettler.xyz/appearances/webinar-swarm-intelligence.md ## Swarm intelligence and AI: BBV AI webinar This webinar explores how AI agents can move beyond isolated assistants and operate as collaborative networks, producing emergent behavior analogous to biological swarm intelligence. ### The evolution of AI agents Several research milestones made agent architectures possible. In 2019 OpenAI showed with GPT-2 that language models can multitask. GPT-3 introduced few-shot learning in 2020. Facebook Research demonstrated in 2021 that language models work as neural search engines. Google's REACT framework in 2022 proved that AI systems can plan and act autonomously. ### From single agents to swarm intelligence An AI agent is a virtual worker: it has a defined scope, uses a language model for reasoning, and has a profile with assigned tools. The real advance comes from networking multiple specialized agents into a collaborative system. When agents coordinate, individual simple actions compose into complex group behavior, mirroring swarm dynamics in nature. ### The agent protocol A standardized agent protocol, comparable to internet protocols for computers, enables agents from different organizations and even humans to communicate. The protocol is based on natural language and supports exchange of tasks, intermediate steps, and artifacts. ### Enterprise rollout Integration works best in four phases. First, deploy individual agents with tightly scoped tasks. Second, let agents expand their knowledge autonomously. Third, connect agents into teams that tackle more complex workflows. Fourth, introduce coordination agents that manage the swarm. ### Practical example A marketing agent can collaborate with a Google AdSense agent and an external optimization agent to improve ad campaigns. The agents communicate over the standardized protocol and can pull humans into decision loops when needed. ### Constraints Current limitations include model context length (roughly 25,000 words), the need for security mechanisms, cost control through smart model selection, and data privacy via on-premise or Swiss-hosted cloud deployments. Trial periods, clear access rights, and monitoring are essential. ### Outlook The open-source community, exemplified by AutoGPT's rapid adoption, is accelerating development. Specialized marketplaces for AI agent services will likely emerge, letting organizations publish and consume agent capabilities. No particular digital maturity level is required to start; agent integration can itself accelerate digital transformation. ## Polygon Software our journey to an innovative UZH tech startup Source: https://joelbarmettler.xyz/appearances/readme-polygon.md ## Polygon Software: our journey to an innovative UZH tech startup Three informatics students Marius Högger, Frank Huber, and I decided to start a company at a table on the second floor of the UZH Department of Informatics at Binzmühlestrasse. The business model was sketched with a digital pen in an unnamed OneNote file between lecture notes and grocery lists. We had recognized that UZH was training a generation of students in data science and AI who wanted flexible hours, flat hierarchies, and varied work and who were underestimated by the job market. We founded a software development company shaped around those needs rather than industry convention, set up an office within walking distance of the informatics campus in Oerlikon, and took on clients in computer-supported data analysis, image recognition, and machine-learning-based quality control. None of us had worked at another company for more than a few weeks. Every question that established firms take for granted how to communicate in a team, how to write quotes for clients who cannot articulate what they want, how to set an hourly rate, how to document code required research, trial and error, and documentation of what worked. The upside: our answers were uninfluenced by inherited practice and often unconventional. Within two years, the company grew from three founders to 18 employees, dozens of clients, and over ten parallel projects. We rebranded from PolygonSolutions to PolygonSoftware. The workforce remained almost entirely UZH students and alumni. Unlimited unpaid vacation, full remote work, and flexible schedules during exam periods were non-negotiable from day one. ## UZH startup label for Polygon Software Source: https://joelbarmettler.xyz/appearances/uzh-startup-label.md ## UZH startup label for Polygon Software Polygon Software, which I co-founded with Marius Högger and Frank Huber during our informatics studies at UZH, received the official UZH Startup label in 2021. At the time the company had 15 employees, 11 of whom were current UZH informatics students. The team built custom software for small and medium-sized enterprises computer-supported data analysis, automated label recognition via image recognition, and machine-learning-based quality control for manufacturing lines.