Your Dose of Reg.exe, Week {16}
AI Stack roles debated. Coding revolution: Claude Code web, Google Opal, Ariana beta. OpenAI Agents projects released. NVIDIA Gr00t in LeRobot. Social media impacts LLMs.
Events
🇺🇸 Vibe AI Workflows Hackathon - October 29, San Francisco - Hackathon by Pipelex x Blackbox AI focused on building AI workflows to solve complex use-cases. 👉 Looking for participants and judges in the SF area. (🙏 Louis Choquel)
🇫🇷 dotAI Closing Gala x Dust - November 6, Paris - Closing celebration for dotAI featuring DJ, drinks, and cool vibes after a full day of AI talks and discussions. (🙏 Jules Belveze)
🇫🇷 Personalised Longevity Hackathon - November 22-23, Paris - 36-hour sprint focused on AI-driven personalised interventions using multi-omic data, digital biomarkers, and clinical-grade models. (🙏 Jeremie Kalfon)
We <3 Sota Blog
🔥 Our “The Case for the Return of Fine-Tuning“ article exploring the renaissance of fine-tuning with new tools and approaches reached the podium on Hacker News.
📈 Framework for AI competitive advantage - A comprehensive analysis on identifying, building, and sustaining defensible value in the age of AI.
🤔 The Deployment Paradox - Part 1 - Systems must be deployed before they’re ready since readiness only comes from real-world use, requiring strategic constraint design to make mistakes survivable.
🎤 Voice interface revolution - A personal note on how voice interfaces are evolving in the AI era. Discussed how one app completely changed perspective on voice interaction, moving beyond failed experiences with Siri.
Knowledge
🗣️ Discussion about the division of the AI Stack - A discussion between Robert Hommes, Pierre Chapuis and Gabriel Olympie explored the growing gap between application-layer engineers and those with deep ML fundamentals, and why understanding adjacent technical layers matters for building differentiated products.
The AI field has distinct layers (application vs model), and many post-LLM engineers only know the application side without deeper ML fundamentals.
AI is moving too fast for anyone to be truly “full stack” right now. Specialization is necessary but you should understand at least one layer above and below where you work.
Deeper technical understanding creates better products and enables novel approaches that surface-level developers might miss.
Balance is key: don’t be too purist about role boundaries, but maintain clarity about responsibilities and focus areas
👉 Gabriel Olympie’s synthesis and matrix of the roles in a data team
💸 Technical debt and knowledge creation - Analysis of paying down technical debt versus accepting ongoing maintenance costs. If you discover a better way to do things, the old way of doing it that is embedded in your code base is now “debt”:
you can either live with the debt, “paying interest” in the form of all the ways that it makes your code harder to work with;,
or you can “pay down” the debt by fixing all the code in light of your new knowledge, which takes up front resources which could have been spent on something else, but hopefully will make sense in the long term.
😎 Vibe coding survey released - Comprehensive survey on vibe coding published for developers and enthusiasts. (🙏 Gabriel Olympie)
Autonomous Agents
👻 Andrej Karpathy notes - Following its interview last week on Dwarkesh Patel, Karpathy has posted a few notes on X (🙏 Fabien Niel)
AGI in 10 years is bullish, just looks slow vs SF hype
LLMs are “ghosts” from internet text, not evolutionary “animals”, no single algorithm learns everything
RL is inefficient, prefers “agentic interaction” and new learning paradigms
Strip LLM memory to improve generalization - inability to memorize is a feature
Current agent tools overshoot capability - wants collaboration not autonomous code dumps, fears “slop” accumulation
Job automation is more complex than predicted
Teach physics early to “boot up brains”
🙏 Experiments with OpenAI Agents - Yusuf Eren open-sourced two projects with OpenAI Agents on Github:
OpenAI Agents Tracing - A comprehensive, self-hosted tracing and analytics platform for AI agents with real-time monitoring, cost tracking, and usage analytics.
OpenAI Agents Fullstack Starter - A production-ready full stack starter project showcasing OpenAI’s Agents framework with a real-time streaming UI. Built with Next.js, TypeScript, and WebSockets.
🎛️ Claude Skills usage patterns - Discussion emerged on how to effectively use Claude’s Skills feature. (🙏 Maxence Maireaux, Mathieu Breton)
Claude Skills = tutorials that teach LLMs specific tasks/tools, like the Matrix kung fu download
Common skills datasets will likely emerge soon
Biotech, Health, and Chemistry
🧬 Writing the human genome - The article explores the next revolution in biology focusing on writing life’s code rather than just reading it. Microbiologist Andrew Hessel argued for the Human Genome Project-write, suggesting we’re beginning to author living organisms. (🙏 Jeremie Kalfon)
Jeremie Kalfon (ENS/Pasteur)’s take on it: “Reading and writing DNA are both facets of the same coin and you can’t write good DNA without reading it and understanding it somehow. But writing a full genome is more of an academic exercise in tooling than anything else. If we could just edit the genome more easily, that would be a huge plus. If we could write parts of it (like plasmids, RNAs...) with high confidence in its effect, that would be amazing. If we could read it more efficiently, less destructively, spatially, with its base modifications, etc... that would be amazing. This article is trying to sell this as THE big thing, but it is not as clearly useful as the Human Genome Project and I don’t think this is where funding agencies want to spend their money.”
Computer Vision
📽️ Krea realtime video model released - Hugging Face released Krea realtime video model for video-to-video and text-to-video generation capabilities.
Infrastructure
💥 AWS us-east-1 outage - Major AWS outage impacted us-east-1 region, affecting multiple services. Community members shared HugOps support for those impacted. (🙏 Pierre Chapuis)
Language Models
🗑️ Social media exposure impacts LLM cognition - Research revealed that excessive social media exposure has similar effects on LLMs as on human brains, triggering cognitive decline. Study explained Grok’s bipolar behavior patterns. (🙏 Gabriel Olympie)
🛠️ Building GPT2 from Scratch - Julien Seveno has built GPT2 using only foundational papers, starting with basic components (FFN, LayerNorm, self-attention, positional embeddings using sinusoidal instead of learned embeddings) and training it on Karpathy’s character-level dataset. (🙏 Julien Seveno)
MLOps
⚙️ Jules Belveze’s deep dive “Instruction Fine-Tuning: Evaluation and Advanced Techniques for Efficient Training“ on why instruction fine-tuning changes LLM evaluation and optimization has been published on neptune.ai.
Programming
🗑️ Software quality crisis analysis - Article examined the great software quality collapse, citing examples like Apple Calculator leaking 32GB of RAM. Argued that companies spending $364 billion to avoid fixing fundamental problems isn’t sustainable. Predicted survivors will be those who remember how to engineer.
⚙️ Google Opal launched - Google released Opal, combining n8n and Lovable-style functionality for building AI mini-apps using natural language. Features no-code creation, rapid prototyping, and intuitive visual editing with node-based workflows.
🎞️ Ariana beta demo - Anicet Nougaret released the close beta of Ariana after 4-month development sprint. Platform offers Claude Code web on steroids with desktop app featuring file sync and network forwarding, enabling code collaboration on agent-produced code. (🙏 Anicet Nougaret @ Ariana.dev)
👾 Claude Code on the web - Anthropic launched Claude Code on the web in beta as research preview. Users can delegate multiple coding tasks that run on Anthropic-managed cloud infrastructure, ideal for bug backlogs and parallel fixes.
Robotic
🤖 NVIDIA Gr00t N1.5 in LeRobot - Collaboration between Hugging Face LeRobot team and NVIDIA Robotics made Gr00t N1.5 available. Model features cross-embodiment foundation capabilities for robots with multimodal inputs including vision, language, and proprioception.
New Members
🇫🇷 Amine Saboni (Pruna AI) - MLOps Engineer at Pruna AI, making inference faster, cheaper, and greener. Maintains the codecarbon package and is interested in the environmental footprint of the AI industry. Enjoys cooking and chilis.
🇫🇷 Chris Wallenwein (Mistral AI)
Thanks for the shoutout Kevin!