Your Dose of Reg.exe, Week {8}
AI is moving fast: 👑 MCP is everywhere, 🔬 DeepConf & CodeMonkeys research, PyTorch ZenFlow, IDE shifts with Codex & Qoder, plus major moves in vision, infra, and robotics.
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
Events
🎭 MCP Connect - Event for MCP builders with talks from Datadog, Alpic, Mux, Twelve Labs and MCPJam. (🙏 Nikolay Rodionov @ Alpic)
🇺🇸 San Francisco, September 10th
🤖 Think Outside the Bot Hackathon - Virtual hackathon co-hosted by Linkup and Qdrant pushing boundaries of vector search. Submit projects until September 16th, winners revealed at Vector Space Day in Berlin. (🙏 Boris Toledano @ Linkup)
🇩🇪 Finale in Berlin, September 26th
👉 Participate here!
🎛️ Engineering Night MCP Special - Technical deep-dives on Model Context Protocol from Alpic, Linkup, and Dust teams. (🙏 Jules Belveze @ Dust.tt)
🇫🇷 Paris, September 10
Knowledge
Rich Sutton presents OaK Architecture - 🎞️ The father of RL introduces a vision for SuperIntelligence from Experience. “The father of RL is back 😉” (🙏 Emmanuel Benazera @ Jolibrain)
Blackwell GPU optimization guide - 📰 Small guide for deploying and optimizing LLMs on RTX Pro 6000 Max-Q Blackwell architecture. (🙏 Gabriel Olympie @ Reg.exe)
Language Models
🔥 Evaluatorq framework released - 🛠️ Open-sourced production-tested AI evaluation framework for testing AI features without building infrastructure. (🙏 Sohrab Hosseini @ Orq.ai)
🔥 Nous Research releases Hermes 4 - 📰 New reasoning model based on Llama 3.1 with improved post-training that breaks Llama alignment and excels on RefusalBench, includes a 405B version. “I didn’t get much time to test it, but Hermes models typically have very strong post-training. They usually start by breaking Llama’s alignment and then try to perform well on RefusalBench. There’s also a 405B version.“(🙏 Pierre Chapuis @ Finegrain)
Autonomous Agents
✈️ Alpic and Kiwi partner on a flight search and booking MCP server - A big step in MCP’s path toward becoming the standard. Kiwi.com can now be queried via LLM in your favorite AI client (how to). Alpic has also released an 🛠️ Automatic instruction generator for MCP servers supporting Claude, Cursor, Cline, VSCode, and other popular tools (🙏 Nikolay Rodionov @ Alpic)
🔥 DeepConf architecture breakthrough - 🔬 Research paper achieving up to 99.9% accuracy on AIME 2025 while reducing generated tokens by up to 84.7% compared to full parallel thinking. (🙏 Kevin Kuipers @ Reg.exe)
🔥 GEPA prompt optimization - 🛠️ Framework for optimizing text components using reflective text evolution, enabling 35x more compute efficiency than GRPO with 10% better results. (🙏 Gabriel Olympie @ Reg.exe)
🎛️ Rube Universal MCP - 🛠️ Composio's unified interface allowing AI to talk to 500+ apps without extra setup, from Slack to Notion to NASA. “Composio was one of the first companies to pivot to the MCP aggregator. I tried Rube, but unfortunately it wasn’t working for me. Too bad, because the idea is really cool.“ (🙏 Kevin Kuipers @ Reg.exe, Nikolay Rodionov @ Alpic)
🧠 AgentFly online learning - 🔬 Paper on LLM agents that learn without fine-tuning by saving all actions in memory instead of updating weights. (🙏 Kemal Toprak Uçar @ Numberly)
🧠 CodeMonkeys scales test-time compute - 🔬 Stanford research resolving 57.7% of SWE-bench issues. “In comparison with DeepConf, this one is mainly about massively scaling the number of agents. I don’t really see it as groundbreaking. It feels more like brute force and tweaking the knobs“ (🙏 Kevin Kuipers @ Reg.exe)
Computer Vision
📸 Gemini 2.5 Flash Image pricing disruption - Google's Nano-Banana model priced at $0.039 per image (about 1290 tokens), significantly undercutting OpenAI while topping Artificial Analysis Image Editing Arena.
“This is really state of the art image editing available to anyone” (🙏 Georges Gomes @ ‹div›RIOTS)
“It’s not that it’s very cheap, it’s that OpenAI is very expensive. Four cents per image is a typical price for generation or editing, for operations that take under 10 seconds like this. (…) For instance, that’s what FAL charges for Flux Kontext Pro, while Qwen Image Edit is priced at 3 cents for 1024×1024.” (🙏 Pierre Chapuis @ Finegrain)
Cyber
🔐 SHA3 Sponge Function explained - 🎞️ Computerphile video on SHA3 as the next big cryptographic hash function waiting in the wings if SHA2 bugs are found. (🙏 Kevin Kuipers @ Reg.exe)
Infrastructure
🗜️ Weaviate's 8-bit Rotational Quantization - 📰 New vector quantization algorithm utilizing random rotations to compress vectors by 4x while improving speed-quality tradeoff. (🙏 Kevin Kuipers @ Reg.exe)
💎 Alibaba develops new AI chip to challenge NVIDIA - 📰 OMGPU reported (this article is in French, [English version: China's Alibaba develops new AI chip to help fill Nvidia void]) on Chinese companies' massive investments in developing local AI solutions, with Alibaba creating a new chip for technological independence. (🙏 Antoine Millet @ Scaleway)
Programming
👩💻 Qoder, new agentic IDE - New platform from Alibaba, 🛠️ Qoder is offering free “2000 credits” on their advanced model. Pretty promising, but they don’t provide yet a paid plan to use it extensively. (🙏 Kevin Kuipers @ Reg.exe)
👩💻 Cursor-Linear integration - 🛠️ New integration allowing Cursor background agents to turn Linear issues into pull requests automatically. (🙏 Arnaud Porterie @ Vibe)
🎁 Grok Code Fast freebie - Kilo Code alongside most AI platforms give access to an uncapped 7-day trial on the last xAI programming model. Grok Code Fast claims to be a “speedy and economical reasoning model”. (🙏 Kevin Kuipers @ Reg.exe)
🧰 DeepSpeed ZenFlow for LLM training - 📰 PyTorch blog discusses new stall-free offloading engine for LLM training. (🙏 Benjamin Trom @ Mistral, Pierre Chapuis @ Finegrain, Robert Hommes @ Moyai.ai)
🙋♂️ Feedback needed for Pierre: “We’ve been using DeepSpeed for a long time—our entire training stack is built on it. I was just curious about the new feature called ZenFlow.”
👩💻 Codex Improvements with GPT-5 and IDE integration - Reports suggest some users are moving away from Claude Opus in favor of Codex. Although Reg.exe members remain unconvinced about its overall superiority, they do praise its clean TUI. (🙏 Kevin Kuipers @ Reg.exe, Benjamin Trom @ Mistral)
Robotic
🧠 Boston Dynamics Atlas with Large Behavior Models - 🎞️ Collaboration with Toyota Research Institute developing end-to-end language-conditioned policies enabling Atlas to accomplish long-horizon manipulation tasks. (🙏 Kevin Kuipers @ Reg.exe)
New Members
Anand Pajaniradjane - Founder @ Scope · The AI Search Growth Engine. Former boxing national champion in France, plays two instruments, spent 4 years deep in LLM world through research + industry. Based in 🇺🇸 San Francisco, US (PST)
Jean du Terrail - Founding AI scientist @ dotomics · Building foundation models on plant genomics to develop the plants of tomorrow. Background in AI research (mostly in healthcare). Based in 🇫🇷 Paris, France (CET)