Your Dose of Reg.exe, Week {1}
I'm excited to launch our first Reg.exe newsletter 🐶, featuring all the content shared by our community on Discord. Feel free to share any feedback or suggestions for improvement 🙏
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
Knowledge
The community shared valuable insights on multimodal open source developments. Robert highlighted an impressive presentation from Kyutai covering multimodal capabilities from online demos to on-device implementations. The discussion showcased the rapid advancement in multimodal AI systems and their practical deployment strategies.
Watch the presentation: Multimodal Open Source at Kyutai, From Online Demos to On-Device
Achievements
Several community members celebrated significant milestones this week. Sacha from Edgee announced their new Edge Functions Components powered by WebAssembly, allowing developers to push edge functions to the Edgee Component Registry for generating HTTP responses at the edge. Read more about this breakthrough at Edgee's blog.
Kevin from Galion.exe shared recognition for their joint analysis with AWS about GraphRAG, which was featured in Cohere's blog. The analysis presents GraphRAG and Agentic Search approaches side by side, highlighting how these methodologies are closely intertwined in modern AI systems.
Julien from Plakar introduced a revolutionary approach to data archiving with their new .ptar format, challenging the traditional 1979 tar format for modern data needs. Explore this innovation at Plakar's blog.
Computer Vision
The computer vision space saw exciting developments with Raoul from Moyai.ai highlighting BAGEL, a scalable unified multimodal model that demonstrates impressive performance with open weights. This model revolutionizes how AI interacts with complex systems and offers developers unprecedented flexibility in multimodal applications.
Language Models
The language models channel buzzed with activity around Tencent's breakthrough release. Gabriel showcased the new Hunyuan-A13B-Instruct model, which represents the first next-generation model that can truly be hosted locally. With 80B total parameters but only 13B active, it's smaller than Llama 3.2 yet faster than Mistral Small, making it incredibly efficient for local deployment.
The community celebrated when VLLM support was added, enabling an impressive 75 tokens per second with almost DeepSeek v3 quality on just 2x3090 GPUs. Gabriel demonstrated this capability with a real-time demo showing the model's practical performance.
Meanwhile, Kemal highlighted Baidu's ERNIE 4.5 model family, featuring Mixture of Experts architecture with models ranging from 0.3B to 424B parameters. The technical report reveals fascinating fast and slow thinking capabilities that could reshape how we approach AI reasoning.
Paco reached out to the community seeking expertise in dataset best practices for agent training, RLHF, and fine-tuning of coding agents, particularly for working with open source contributors and their coding patterns.
Audio And Speech
Laurent from Kyutai provided insights into their upcoming releases, mentioning they should have new developments within two weeks. The team is working with VLLM for text LLM integration and considering switching from Gemma to Mistral for better Apache licensing and size flexibility. They're developing with the 27B version, running cloud inference with 12B, and offering 4B variants for local users.
The discussion revealed an exciting collaboration between Mistral and Kyutai, with Mistral providing the text LLM powering Unmute, Kyutai's open-source voice AI project.
Autonomous Agents
The autonomous agents space dominated discussions with extensive coverage of Model Context Protocol (MCP) developments. The Linux Foundation executive described MCP as "AI's HTTP Moment", positioning it as a foundational communications layer for AI systems.
Claude Code recently added remote MCP support, making it the most user-friendly implementation with claude mcp add functionality. The community noted that Linear became one of the first services to offer managed remote MCP servers, with Sentry and Linear emerging as key business cases for Claude's implementation.
However, security concerns arose with Kevin and Julien expressing strong reservations about exposing databases to MCP, calling such practices "madness." Robert shared a concerning analysis about Supabase MCP potentially leaking entire SQL databases.
On the positive side, MCP now supports full authentication through WorkOS AuthKit, addressing many security concerns that plagued earlier implementations.
Nnenna published a comprehensive analysis of 8 missing architectural patterns in Generative AI, arguing that with MCP adoption rising, this creates a blueprint for more protocols to emerge and solve real infrastructure problems in AI.
Edge And Embedded Ai
The edge computing discussion centered around innovative MCP implementations at the edge. Robert and Sacha explored the potential for Edge+MCP combinations using WebAssembly components, which could provide stability by safely running MCPs in V8 with WASM while offering zero-cost compute.
Mozilla released significant research on WASM-agents, demonstrating AI agents running directly in browsers as HTML files, removing dependencies on external frameworks and making agent deployment incredibly simple.
The community discussed the distinction between local and edge computing, with Robert noting that local environments have high heterogeneity while edge environments are homogeneous, making edge deployment more predictable and scalable.
Cyber
Critical security updates dominated the cyber channel this week. Kevin highlighted severe sudo vulnerabilities that allow local users to gain root access on Linux systems, requiring immediate attention from system administrators.
Pierre warned about Dashlane's concerning decision to discontinue web support within two months, recommending migration to Passbolt or Bitwarden for teams. He criticized this as a "terribly stupid move" given Dashlane's already problematic browser extension behavior.
Bio Tech Health Chemistry
Willy announced an informal afterwork event for deeptech and biotech founders, hosting up to 15 people at their Sentier station office in Paris. This reflects the growing intersection between AI and biotechnology in the community.
Robotic
The robotics space saw exciting developments with Hugging Face releasing Reachy Mini, an open-source robot designed for today's and tomorrow's AI builders. Francesco hinted at upcoming releases from the Hugging Face team.
Pierre highlighted Genesis AI, a company that raised $90M and is making significant strides in robotics and AI integration. Their ambitious approach draws inspiration from thoughtful science fiction like The Talos Principle.
Programming
Programming discussions featured a controversial study on AI-assisted developer productivity. The research suggested that AI assistance doesn't significantly improve developer performance, but Thierry identified major methodological flaws in the study design. The random assignment of AI usage doesn't reflect real-world developer behavior, where developers strategically choose when AI is most beneficial.
Kevin showcased Terragon, a platform enabling parallel Claude Code instances working simultaneously on codebases, representing a significant advancement in AI-assisted development workflows.
The community also explored various CLI coding tools, with Robert noting that OpenCode emerged as an interesting winner with its Go-based CLI design. The comparison of multiple AI coding assistants solving the same problems in parallel provided valuable insights into their relative strengths.
Nia v0 by Nozomio was introduced as a way to 10x coding agent effectiveness by providing more context through MCP integration, showing the practical applications of the MCP protocol in development workflows.
Gaming
Kevin shared an insightful interview with John Carmack, the legendary creator of DOOM, discussing his transition into AI development. This provided the community with perspectives on how gaming industry innovations can influence AI advancement.
Julien sought recommendations for open-source games suitable for AI experimentation, leading to community suggestions including CorsixTH, Wesnoth, and Warsow. These games offer opportunities for implementing reinforcement learning algorithms and experimenting with AI dialogue capabilities.
Infrastructure
Infrastructure challenges were highlighted through Pierre's AWS capacity issues, where spot instances were available but on-demand instances showed "Insufficient capacity" errors. This sparked discussions about AWS quota management and the differences between spot and on-demand instance availability, with the community rallying to help resolve the technical challenge.
New members
Klaas Foppen - CTO at Promptwatch, 15 years in software development, drummer for 20 years, based in Amsterdam, NL
Pierre Vannier - Chief Vibecode Officer at Flint.sh, 25+ years in tech, producer of "IA pas que la Data" podcast, appears on BFM Business, based in Montpellier + Paris, France
Louis Choquel - Co-founder & CTO at Pipelex, repeat entrepreneur building open-source language for repeatable AI workflows, based in Paris, France
Niels Schmidt - CTO & Co-founder at Droidrun, enabling AI Agents to control mobile apps, former track & field athlete (sub-11 second 100m), based in Osnabrück, Germany
Christian Ninstel - CEO & Co-founder at Droidrun, recently incorporated in Delaware and closed Pre-Seed, passionate about travel and foraging, based in Osnabrück, Germany
Geographic hubs
🇺🇸 San Francisco
Jules organized a low-key BBQ event featuring beers, food, and AI discussions on a rooftop setting. The event brought together like-minded individuals passionate about generative AI, fostering community connections in the SF area.
🇩🇪 Berlin
Youssef sought affordable workspace recommendations for aspiring founders, with Christian providing helpful information about the Merantix AI Campus residency program. The community demonstrated strong support for members transitioning into entrepreneurship.
🇳🇱 Amsterdam
Klaas joined the Amsterdam hub, with Kevin welcoming the new member. Robert encouraged making the Amsterdam hub more active, showing the community's commitment to fostering local connections and engagement across different geographic regions.