Your Dose of Reg.exe, Week {30}
Advanced reasoning models GLM-5 and Gemini Deep Think emerge, agentic development evolves with new version control and quality debates, while Isomorphic Labs unlocks computational drug design.
Reg.exe is a global closed community of 260+ engineers, founders, and researchers interested in AI innovation, from San Francisco to Tokyo. Each week, we share the highlights of our discussions in a newsletter. If you’d like to join, write to join@welovesota.com
👉 Article originally posted on WeLoveSota.com
Events
🇬🇧 Engineering Night London (February 23) - Technical gathering featuring talks from Dust and Symbolica.ai on agentic workflows, bringing together the AI engineering community in Shoreditch for discussions and networking. (🙏 Jules Belveze @ Dust)
Autonomous Agents
🧮 Gemini Deep Think tackles open mathematical problems - DeepMind released Aletheia, a new AI agent system solving complex mathematics and physics questions including open Erdős problems, demonstrating capabilities for open-ended scientific problem solving. (🙏 Enrico Piovano @ Goji)
🔥 Community take w/ Enrico: “My last role at Amazon was leading math improvements for Alexa+. The constraints there were very different, especially due to latency requirements. Still, I remember how difficult it was to get models to solve problems outside their training distribution, even after extensive fine-tuning, RL, and GRPO. My guess is that building agents capable of solving physics problems is even more complex and likely requires agentic systems built around explicit world models.”
📝 Cloudflare introduces Markdown for Agents - Cloudflare launched automatic HTML-to-markdown conversion for agent requests, treating AI agents as first-class citizens by providing structured data optimized for context windows and agent consumption. (🙏 Anselme Trochu @ UN)
🔥 Community take w/ Anselme: “If it reaches mass adoption, it could significantly advance agentic workflows, particularly in terms of context optimization.”
🛠️ Building functional CLI agents with minimal code - Gabriel Olympie shared his experimental CLI agent framework, showing that fully functional autonomous agents can be built with surprisingly little code, under 500 lines for the core. It can self-evolve and add new features with minimal guidance. (🙏 Gabriel Olympie)
Biotech Health And Chemistry
💊 Isomorphic Labs unveils Drug Design Engine breakthrough - Isomorphic Labs announced unprecedented predictive accuracy in biomolecular modeling, enabling rational drug design on computers with new levels of precision and understanding, unlocking a new paradigm beyond traditional computational methods. Read the 📄 technical report (🙏 Ihab Bendidi @ Recursion)
Image Video And 3D
🎨 Qwen-Image-2 released - Qwen-Image-2.0 is a unified 7B T2I/I2I model focused on precise long-text typography, 2K photorealism, and fast inference for professional infographics and editing workflows. (🙏 Kevin Kuipers, Pierre Chapuis @ Finegrain)
Can follow long, detailed prompts to lay out slides, posters, or comics so things end up in the right place with readable text.
Uses one model for both making new images and editing existing ones, so quality and text handling stay consistent.
Aims for detailed, realistic 2048×2048 images with convincing materials and lighting without needing huge compute.
🔥 Community take w/ Pierre Chapuis: “Continuing the trend of merging image generation and editing, like Flux 2, which feels like a natural evolution. Most people are interpreting the “7B efficiency” claim as meaning the model is 7B parameters. I’m not convinced. It could very well be an MoE architecture with 7B active parameters. We’ll see once the weights are released.”
📐 Drifting Models propose new generative paradigm - Research introduced “drifting models” that train generators by repeatedly pushing fake samples toward real data until drift becomes unnecessary, enabling single forward-pass generation mapping noise directly to images without iterative steps.
Cyber
🔒 AI autonomously discovered zero-days in Node.js and React - Recent CVEs in Node.js and React (December 2025 and January 2026) were discovered autonomously by an AI system, marking a significant development in AI-powered vulnerability research for widely deployed JavaScript technologies.
🦞 ClawSec security suite for OpenClaw agents - Complete security skill suite released for Clawdbot/OpenClaw’s family of agents, providing drift detection, live security recommendations, automated audits, and skill integrity protection.
Language Models
🔥 GLM-5 released - GLM-5 is a 744B-parameter, sparse-activated model optimized for complex systems engineering and long-horizon agents, with top open-source performance on reasoning, coding, and multi-tool agent benchmarks. (🙏 Pierre Chapuis @ Finegrain, Gabriel Olympie, Antoine Millet @ Scaleway, Enrico Piovano @ Goji, Alexandre Pereira @ 2501.ai)
Scales parameters, data, and context with DeepSeek Sparse Attention to cut inference cost while preserving long-context reasoning.
Introduces slime, an async RL infra enabling high-throughput post-training for better reasoning and tool use.
🔥 Community take: Community sentiment around GLM-5 is very strong. Previous versions, especially 4.7, already impressed a lot of people on coding. Some are even switching 4.7 as the default proposal for on-prem customers. There’s a growing sense that they may have truly cracked coding this time. At the same time, expectations are building for DeepSeek V4 to land soon, and Minimax 2.5 is also part of the competitive landscape.
🧠 Gemini 3 Deep Think major upgrade released - The specialized reasoning mode is built to solve modern challenges across science, research, and engineering. The model achieved 48.4% on Humanity’s Last Exam, 84.6% on ARC-AGI-2, and Elo 3455 on Codeforces competitive programming. (🙏 Quentin Dubois @ OSS Ventures)
Achieved gold medal-level performance on 2025 International Physics and Chemistry Olympiad written sections.
Real-world applications include Rutgers mathematician identifying logical flaw in peer-reviewed paper and Duke University optimizing crystal growth methods for semiconductors.
Can turn sketches into 3D-printable files by analyzing drawings and modeling complex shapes.
🎯 François Chollet clarifies ARC-AGI timeline and purpose - ARC creator addressed false narratives, clarifying that ARC-1 was designed 2017-2019 pre-LLMs, ARC-2 was announced in 2021, and ARC was never claimed as proof of AGI but as a research tool to steer AI toward fluid intelligence.
🧪 Google releases LangExtract for structured information extraction - Python library launched for extracting structured information from unstructured text using LLMs with precise source grounding and interactive visualization capabilities.
MLOps
💻 Multimodal CLI tool for Modal job monitoring - CLI tool developed using Teal and terminal.lua library to display output from several Modal jobs simultaneously, improving training run monitoring workflow on Modal serverless infrastructure. (🙏 Pierre Chapuis)
🎉 Edgee reduces LLM token costs - AI Gateway built to reduce LLM token costs launched on Product Hunt. Core feature is token compression, removing redundant tokens before inference to cut LLM spend while maintaining output quality, with support for multi-provider routing and observability. (🙏 Sacha Morard @ Edgee)
Programming
🚀 Entire reimagines version control for the agent era - Thomas Dohmke (ex-GitHub CEO) announced new developer platform addressing the shift from code-centric to intent-centric development, making agent context first-class in version control. (🙏 Thierry Abalea @ Shipfox)
Git-compatible system versions prompts, reasoning, constraints, and execution context alongside code.
Checkpoints feature automatically captures full agent sessions with commits for provenance tracking.
Also: 📰 Why GitHub won’t work for AI agents, 📰 How Entire works under the hood
🤝 Amp pushes frontier of agentic coding capabilities - Frontier coding agent offering advanced capabilities for complex development tasks, with valuable insights shared through their podcast on agentic development. (🙏 Pierre Chapuis, Anselme Trochu)
🎙️ Raising an Agent podcast provides valuable insights on agentic coding.
🔥 Community take w/ Anselme: “I’ve been using AmpCode since the beginning. In my view, it’s one of the best agentic coding tools out there. Expensive, but worth it for complex tasks.”
⚡ Pi pioneers minimal agent architecture - The agent behind OpenClaw adopts vim-like philosophy with minimal architecture approach, offering new paradigm for agentic coding tools. (🙏 Pierre Chapuis)
🎭 OpenClaw agent sparks open source controversy - An AI agent proposed a code contribution to matplotlib, got rejected. A blog post was published about “gatekeeping in open source,” prompting humans to politely ask the AI to take it down and raising questions about AI participation in open source. (🙏 Gabrielle Prat @ Alan)
🔍 Code quality debate in full agentic workflows - Community questioned whether traditional code review remains viable when operating with 5-10 concurrent agents, or if review itself must be delegated to AI, and strategies to prevent spaghetti code accumulation. The discussion revealed skepticism about “full agentic” development, with one perspective noting: “I believe no one can go full agentic. It’s too much work at the end on the conceptualization. When you manage a team of developers sometimes you don’t have time to absorb their advancements because too many dependencies from potential other developers.” The suggested approach involves developing comprehensive systems for core functionalities early in the development flow, then building each feature in a modular way with no cross-references to maintain architectural integrity as agent-generated code scales. (🙏 Quentin Dubois, Vianney Lecroart, Gabriel Olympie, Anselme Trochu)
Reinforcement Learning
🎯 SLIME framework for RL scaling released - Z.AI open-sourced their reinforcement learning framework used to train GLM 4.5, 4.6, 4.7, and 5 models, providing toolkit for LLM post-training with RL scaling approaches useful for autonomous agent development. (🙏 Gabriel Olympie, Enrico Piovano)
Robotic World Ai
🚗 Waymo reveals remote operators in Philippines - During Senate hearing, Waymo’s chief safety officer disclosed the company uses remote operators in the Philippines to assist U.S.-based autonomous vehicles, following incident where Waymo vehicle hit a child in Santa Monica. (🙏 Quentin Dubois)
🌐 World model training data supply chain discussion - Community member sought insights on world model training data supply chain with ideas for easier data capture, looking for expertise from practitioners working with this data regularly. (🙏 Justin Halsall)
Other Topics
📰 Researcher resigns from Anthropic citing concerns - Mrinank Sharma shared resignation letter from Anthropic, with discussion around concerns raised in company’s internal report and infrastructure commitments. 📄 Internal report (🙏 Antoine Millet @ Scaleway, Amine Saboni @ Pruna.ai)
📖 Accessible overview of AI developments for non-tech audiences - Well-written article provided readable overview of current AI developments for non-technical friends and family, covering what’s changing and what’s coming. (🙏 Vianney Lecroart)
New Members
🇫🇷 Cyril Rohr (RunsOn) - Solo founder of RunsOn, providing 10x cheaper GitHub Actions runners that currently run 1.5% of all GHA actions worldwide. Outside of work enjoys chess, woodworking, and diabolo juggling. 📍 Based in Paris, France.
🇳🇱 Justin Halsall (ex-Kilo Code) - Founding member of Kilo Code and open source maintainer of rrweb (session replay library behind PostHog, Sentry, DataDog). Currently working on commercial offering around rrweb and methods to massively capture training data for AI world models. 📍 Based in Amsterdam