Your Dose of Reg.exe, Week {5}
GPT-OSS and HRM have been the main focus this week, thanks to the brave members who remained active during the summer break ☀️. Our NL format has been slightly refined with less automation 🧑💻
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
Computer Vision
Alibaba's Qwen-Image arrived in Diffusers - New 📰 PR #12055 adding Qwen's image model to Hugging Face ecosystem
Language Models
The OpenAI major announcements. The first open-weight releases since GPT-2, GPT-OSS-117B and GPT-OSS-20B (Apache 2.0 license), have been welcomed, but also met with mixed reactions due to poor performance by the 117B model on independent benchmarks. There is also suspicion that it was trained on synthetic data, reflected in its weak real-world knowledge. Although So far, GPT-5 has drawn little reaction given its early release, aside from some sarcasm about the livestream and the graphs presented.
Hierarchical Reasoning Model discussion - Sapient's 27M parameter brain-inspired model, claiming 40.3% on ARC-AGI with just 1000 training examples using HRM uses algorithmic learning instead of chain-of-thought reasoning (🔬 Hierarchical Reasoning Model, 🧑💻 Github), has sparked discussion:
The 27M parameter size, while minimal for language fluency (🔬TinyStories), performs well on specialized tasks.
Although, the approach is computation-intensive and limited to grid worlds.
Also, comparisons to LLMs are considered misleading, as specialized systems already outperform LLMs significantly on ARC-AGI benchmarks.
The participants view this as a breakthrough despite the specialized-vs-generalist comparison being unfair.
Additional resources on the ARC-AGI topic: 🔬ARC-AGI Without Pre-training, 🔬 ARC-AGI Challenge 2024, and 🔬 ARC Prize 2024 Winners & Technical Report
RL Renaissance emerging - AGI House published 📰 first blog post on reinforcement learning's comeback as pre-training plateaus
Flash Linear Attention library highlighted as gold resource - 📰 Efficient implementations of state-of-the-art linear attention models
Audio And Speech
Voxtral deployment interest growing - David Leroy from Blynt planning real-world testing, sharing experiences soon
Programming
Kilo Code gaining traction - Users reporting improved clarity over Claude Code, though experiencing some teething problems with timeouts and regressions
Anthropic's vibe coding philosophy shared - 🎞️ Erik Schulntz presentation on building-first culture at Code w/ Claude event
Casey Muratori's influence discussed - His "Clean Code, Horrible Performance" 🎞️ video changing perspectives on OOP and clean code. Amine Dirhoussi shared 📰 Rust implementation article with Casey's feedback
Gaming
Google DeepMind unveils Genie 3 - Revolutionary world model generating interactive 3D environments in real-time at 24fps/720p. Creates playable worlds from text prompts, maintaining consistency for minutes. 🎞️ Watch the mind-blowing demo blew people’s mind.
New Members
Jeremy LE - Freelance ML Engineer specializing in Computer Vision. 35-year-old cognitive systems enthusiast building ML startups in gaming/AI. Enjoys discovering new tech and meeting new interesting people. Based in Paris, France 👔 Linkedin
Ask For Help
AI virtual staging solutions discovered - Vianney Lecroart is looking for a tool that can decorate photos of an empty home by adding furniture without altering the original background. No perfect solution seems to fit the request. Alternatives mentioned include 🛠️ Finegrain Chat, 🛠️ InteriorAI for home staging, OpenAI (GPT Image 1) or Imagen.
OpenAI API 403 error from Cloudflare Workers ✅ - Issue traced to edge location restrictions, blocking OpenAI’s servers to Cloudflare Workers from South Korea. Solution found using 🛠️ Orq.ai as middleware, simplifying embeddings, search, and RAG operations.
Job Board
🧠 Senior Software Engineer @ Rippletide - Building neuro-symbolic AI with hypergraph architecture for hallucination-free agents. Strong Python/TypeScript required. Contact 👔 Louis-Nicolas Roussel.
🕹️ Founding Software Engineer @ Arcade AI (Paris) - Join 👔 Remi Kaito building AI-powered game creation platform. Looking for product-focused engineer with agent-building experience. Check out their 🎞️ Slenderman demo and 🎮 playable game.
Geographic Hubs
🇺🇸 San Francisco
👔 Benjamin Trom (Mistral) will be in San Francisco from August 10 to 19. He’ll meet 👔 Robert Hommes, who’s already on site. Feel free to reach out to them and join.
🇳🇱 Amsterdam
Product Summit AI launching October 7 - Full-day event focused on scaling AI products in production. New landing page just went live at 🎭 productsummit.ai
☀️ Amazing contributors this week: Gabriel Olympie, Pierre Chapuis (Finegrain), Robert Hommes, Kevin Kuipers (Reg.exe), Benjamin Trom (Mistral) Karim Matrah (Contrast), Amine Dirhoussi (Quivr), Kemal Toprak Uçar (Numberly), Vianney Lecroart (Lemlist), Sohrab Hosseini (Orq.ai), Louis Manhes (Genario), Raoul Ritter, Sacha Morard (Edgee), Jeremy Lê, Guillaume Lesur (Wire), Julien Seveno, Remi Kaito (Arcade AI), David Leroy (Blynt), Louis Choquel (Pipelex), Justin Halsall (Kilo Code), Fabien Niel (Quantiq.io).