Your Dose of Reg.exe, Week {4}
Despite the summer break, it has been an intense week with extensive technical discussions, jaw-dropping releases from Alibaba and others. Also, major achievements and insights from the community.
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
Ask for Help
YC Application Support - Pierre Vannier sought help for a friend preparing Y Combinator applications. Kevin Kuipers provided a comprehensive list of entrepreneurs from YC in the community including founders or alumni from Cubic, Mastra, Defer, Twenty, Awen, Nya, Algolia, and Meilisearch.
Job Board
Founding AI Engineer at Moyai.ai - Robert Hommes posted the opening for a technical founding role requiring experience with transformers/TRL, VerL, and the Unsloth framework. Position focuses on agentic advanced analytics and fine-tuning analytical reasoning models. Company is backed by Speedinvest and Galion.exe. If you know a good fit, 📨 contact him
Events
Upcoming Talk on MCP - As discussed last week, Arnaud Jean from AWS shared the abstract for his talk at the 🎭 World AI Cannes edition. The topic will cover strategic approaches to MCP and multi-agent orchestration for Agentic AI development, scheduled within the Next AI Tech track. Nikolay Rodionov from Alpic expressed interest in potential collaboration
Achievements
Open Source release of reDB: Tommi Hippeläinen announced a new release of reDB, a distributed data mesh platform enabling seamless data access, replication, and migration across relational, NoSQL, graph, vector, and other databases. Key features include real-time replication, zero-downtime migration, a unified schema model, policy-driven data obfuscation, MCP integration for AI, and full self-hosting for complex pipelines across on-prem, hybrid, and multi-cloud environments. 📰 GitHub Repository
Knowledge
XBOW Alloy Agents - Robert Hommes shared an innovative approach to building agents from foundational models: 📰 XBOW – Agents Built From Alloys
AI-Generated Neural Network Architectures - Julien Seveno discussed how AI is generating novel NN architectures that converge faster, calling it an 🔬 ’AlphaGo Moment’ for Model Architecture Discovery. A statistical methodology debate followed, with Benjamin Trom (Mistral AI) and Raoul Ritter raising concerns about potential p-hacking in reported performance improvements
Genesis Physics Engine - Kevin Kuipers shared excitement about new AI learning capabilities: 🎞️ This AI Learns Faster Than Anything We've Seen!
Computer Vision
FLUX.1-Krea-dev - Gabriel Olympie mentioned FLUX.1-Krea-dev, a model released by Black Forest Labs available on Hugging Face. Pierre Chapuis noted the model uses a guidance-distilled version similar to Flux-dev, though the raw pretraining version remains private
Datastacks
Supabase Edge Functions discussion - Kevin Kuipers sought advice on extensive usage. Pierre Chapuis (who’s the current maintainer of the supabase-bin package in the Arch User Repository) shared experience noting limitations with multi-threading and 2-second CPU time limits. Alternative platforms suggested included Cloudflare, Netlify, and AWS Lambda for better serverless experiences
Language Models
Qwen3 Model Series Excitement – The community buzzed with anticipation over Alibaba’s latest releases. Gabriel Olympie said, “Your 30B model is so good that you can compare it to Kimi and DeepSeek and it still looks good 😂.” Kemal Toprak Uçar (Numberly) joked that he was already planning to replace his 14B model.
Unsloth quickly provided GGUF versions: 📰 Qwen3-Coder-30B-A3B-Instruct-GGUF
Qwen3-30B-A3B-Thinking Model - A thinking model was also released with impressive benchmark results: 📰 Hugging Face Model
Qwen3-Coder-Flash Announcement - Kemal shared the official announcement with native 256K context (supports up to 1M tokens with YaRN) and optimized for platforms like Qwen Code, Cline, Roo Code, Kilo Code, etc: 📰 Qwen Twitter Announcement
Alibaba's "Killing Spree" 🔫 - Gabriel Olympie commented that "between that and WAN 2.2, Alibaba is on a crazy killing spree, 5 more days and they get the nuke", highlighting the rapid succession of impressive model releases
GLM-4.5 MoE Models - Kemal announced "something big just dropped again" with two MoE models featuring 355B/32B and 106B/12B parameters, claiming to be the best open-source model across 12 benchmarks: 📰 GLM-4.5 Blog Post
Hardware Investment Discussions - The rapid model releases sparked hardware upgrade conversations. "I've been hesitating to get a RTX Pro 6000 Max Q for a few weeks, I think it's time to invest" said Gabriel.
Pruna.ai Optimization Success - Rayan Nait Mazi reported that all WAN 2.2 models are optimized with Pruna.ai, achieving impressive performance of 1 second video generation every 5 seconds on Replicate with WAN 14B 🔥
Long Context Technical Discussion - A conversation emerged about context length capabilities. Benjamin Trom praised Gemini 2.5 Pro as "the goat for effective long context." Anirudh Kulkarni from Google Labs shared a bit dated paper (🔬 Ruler, also praised by Benjamin) that outline of how different models degrade after specific context length.
Programming
Kilo Code testimonials - Website improvements were discussed, including adding logos and customer testimonials for this new open-source AI-native IDE. Kevin Kuipers has replaced Claude Code with it. Marek Kalnik offered Justin Halsall from Kilo Code a quote from Theodo
Geographic hubs
🇳🇱 Amsterdam
Power of Europe - Kevin Kuipers served on the jury at the 🎭 Power of Europe Hackathon organized by Tech Makers, and Timothy Lindblom from Natively.dev (sponsored by Orq.ai, Mistral AI, Koyeb, Kilo Code, and Weaviate)