Your Dose of Reg.exe, Week {9}
This week, our community discussed: GenAI app rankings, vibe coding retention, MCP’s growth, Vaire’s reversible chip, plus new tools in ASR, safety, edge AI, and coding.
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
🇸🇪 Nordic.js Conference - October 2-3 in Stockholm. Two-day conference bringing developers together to learn, share, and get inspired with world-class speakers. (🙏 François Lecroart @ Sakod.fr)
General
Top 100 Gen AI Consumer Apps Released - 📰 Andreessen Horowitz dropped their 5th edition analyzing which AI apps people actively use and what's making money. Notable findings include surprising allstars with companies relatively unknown despite following the field, and a hockey stick growth pattern for Supabase driven by vibe coding trends. (🙏 Robert Hommes @ Moyai.ai)
Web has 11 new entrants (vs. 17 in March 2025), driven by actual traffic growth.
Mobile, 14 new apps as App Stores removed “ChatGPT copycats,” making space for original AI apps.
Google adds multiple distinct products as they’re now tracked separately, like Gemini (#2 web, after ChatGPT, with approximately 12% of ChatGPT’s web traffic). AI Studio (developer sandbox) or NotebookLM (#13 web).
Gap tightening as Grok and Meta make moves.
Vibe-coding shows staying power: US cohorts on a leading platform post >100% revenue retention for months after signup, meaning overall spend grows even with churn.
Knowledge
🧑💻 CRM and Taxonomy’s Limitations - Following Kevin Kuipers' essay 🗞️ "Vibe-Coding as a VC: We Need To Eat Our Own Dog Food”, community discussed challenges with rigid taxonomy systems in CRM and ERP-like tools. Examples from gaming illustrate the difficulty of classifying emerging genres (much like with new startups) where consistency quickly becomes impossible. LLMs can play a role in navigating this fluidity. (🙏 Antoine Millet @ Scaleway, Kevin Kuipers @ Reg.exe)
With conventional relational databases: “It's really hard (and I believe impossible) to land on a solid taxonomy. But if you ask "I am looking for a game when you can drive and shoot, packed with a lot of action", a RAG (…) would give you most of the time quite a pretty decent result.”
“Airtable’s Interface feature is clever, but your UX has to conform to their data architecture (…) Often, it’s far more convoluted than what you’d do with pure SQL”
AI Safety and Ethics
⚖️ AI Replacing Developers Debate - Extensive discussion on ethical concerns around AI tools potentially replacing developers (🙏 Pierre Chapuis @ Finegrain, Thibaut Bayer @ SensCritique, Kevin Kuipers @ Reg.exe)
Key points included offshore code farms and service companies facing elimination, and the "only replace juniors" pipeline problem.
LLMs far from replace all engineers, but this is the goal.
Enhancing or replacing developers is not unethical: it’s part of the innovation.
Companies like SAP and Duolingo already replacing non-tech teams with AI.
"I want to replace people. If a job can be automated it's not worth doing for humans at scale."
Audio and Speech
🗣️ Real-Time Meeting Summarization - Discussion on using small-sized LLMs with native audio support to detect boundary conditions and topic shifts in dialog. Goal is generating incremental meeting minutes with best accuracy and shortest latency while keeping compute reasonable. (🙏 David Leroy @ Blynt, Laurent Mazare @ Kyutai)
Kyutai’s streaming ASR outputs text with small delay → you could feed sentences/segments into an LLM for “incremental summaries.” tuning it into a pure text problem, easier to experiment with than working on audio directly.
Kyutai’s ASR includes semantic VAD (pauses) and will handle speaker changes in the future.
Topic shifts are harder within ASR (limited context), but a small fine-tuned LLM could handle them better thanks to longer context.
🔊 WaveBlender Physics-Based Sound - 🎞️ Two Minute Papers coverage of NVIDIA tech using physical properties of scenes (soft, bouncy, liquid objects) to create sound calculated from physical laws rather than AI. Must-watch for game creators. Full paper here (🙏 Kevin Kuipers @ Reg.exe)
Autonomous Agents
🎛️ MCP content and discussions this week (🙏 Robert Hommes @ Moyai.ai, Nikolay Rodionov @ Alpic, Kevin Kuipers @ Reg.exe)
Playwright Most Used MCP Server - Cursor revealed at WorkOS MCP night that Playwright is the most popular MCP server in Cursor usage.
MCP Universe Benchmark - 📰 New benchmark using Model Context Protocol servers to test agents on 231 challenging, practical tasks across 6 domains including location navigation and repository management.
MCP UI Rich Experiences - 🛠️ New tool enabling MCP servers to deliver interactive web components that can be rendered by MCP clients, moving beyond text/JSON outputs. 100% open-source 🔥.
Alpic’s 🗞️ announcement to “build the first MCP-native cloud platform”.
👩💻 LangChain Open-SWE - 🛠️ Open-source asynchronous coding agent garnering community interest, and seeking real-world experience. (Contact Taekmin Kim @ Bg.app)
Language Models
📑 POINTS-Reader Research - 🔬 HuggingFace paper on distillation-free adaptation of vision-language models for document conversion, challenging assumptions about synthetic data conveying less information. (🙏 Kevin Kuipers @ Reg.exe)
Infrastructure
🤯 Vaire Computing Breakthrough - 📰 EE Times reported on Vaire's energy recovery demonstration with reversible computing test chip. (🙏 Fabien Niel @ Quantiq.io)
Rodolfo Rosini, CEO: “There are enormous misconceptions about reversible computing. When you say ‘reversible computing,’ people say, oh, we’re still so far away from Landauer’s limit. That’s true, but CMOS can’t get to Landauer’s limit, and we’re now less than an order of magnitude away from CMOS hitting [its own] terminal limit.”
Hannah Earley, CTO: “When you re-architect CMOS to operate reversibly, you’re still using the same signal energy, so the energy that ends up on the gates of these transistors is still the same half CV squared. So even though the transistors are not burning energy as you charge them and discharge them, that energy still needs to go somewhere. You need to couple this with a resonator that is able to recover that energy after each clock cycle and then re-deliver it in the next clock cycle.”
Rodolfo Rosini, CEO: “Everything – eventually, we want to do everything,” Rosini said. “We’re convinced that in 20 years, every new chip is going to be reversible, either quantum or classical computing. There is no other way.”
Cyber
⚔️ Cipher Attacks in LLM Context - 🔬 Research introducing CiFR benchmark for testing safeguards against cipher-enabled harmful fine-tunes. Simple activation probe can detect cipher-enabled harm >99% while keeping false positives low. (🙏 Kevin Kuipers @ Reg.exe)
Datastacks
📘 Jupyter Agent Dataset Release - 📰 New dataset built from 7TB of Kaggle data + 20k notebooks, creating real code execution traces using Qwen3-Coder and E2B. Training dramatically improves ability to execute code and analyze data. Approach aligns with ByteDance's OmniSQL approach. (🙏 Robert Hommes @ Moyai.ai)
Edge and Embedded AI
⏩ Optimizing AI Inference with Edge Computing - 📰 Edgee's research on how edge computing speeds up AI inference by offloading tokenization and RAG to the edge, improving latency, reducing costs, and enhancing user experience. (🙏 Sacha Morard @ Edgee)
Programming
🛠️ Zed Releases Agent Client Protocol Update - Zed's ACP standardizes communication between code editors and coding agents. Claude Code now available in public beta directly in Zed, though integration feels somewhat rushed. (🙏 Kevin Kuipers @ Reg.exe, Benjamin Trom @ Mistral)
🔥 Cocode Open-Sourced - Internal tool for updating docs and generating changelogs powered by Pipelex library. Turns 4-hour doc updates into 30-second commands. 📹 Demo video available. (🙏 Louis Choquel @ Pipelex)
Robotic
👁️ DINOv3 Vision Backbone - 🔬 Meta's 7B general-purpose vision backbone whose frozen features outperform prior methods on classification, detection, and segmentation without fine-tuning. (🙏 Louis Manhes @ Genario)
“Has anyone tried to align its features with a text encoder (using a CLIP/LiT-style contrastive loss) to perform zero-shot image classification or open-vocabulary segmentation?” (👉 Contact Louis Manhes @ Genario)
Ask for Help
🧑💻 Meta Pixel/CAPI Expert Needed - Looking for freelance or advisory experts in Meta Pixel/CAPI for hourly consultation on conversion tracking architecture from both Frontend/Backend perspectives. (👉 Contact Lior Oren @ Replika)
🎭 Big Data Paris Conference Question - Community member seeking feedback on whether the conference is worth attending. (👉 Contact Sohrab Hosseini @ Orq.ai)
Job Board
🚀 Backend Engineer at Odysseus Space - Aerospace and defence startup of 30 people developing laser communication systems seeks backend engineer for distributed systems. Apply here (🙏 Marco Agnan @ Odysseus)
Community
🍕 TechLunch.exe - First TechLunch.exe of the season, hosted by Margaux, covering topics from data centers to model evaluation and hiring tech talent in the AI era. Summary of discussions to come. For Q4 Paris visitors, DM for next lunch invites. (🙏 Margaux Wehr @ SOTA w/ Eloise Castarelli @ Jimini.ai, Jules Belvezes @ Dust, Guillaume Daix @ Datadog, Vlad de Turckheim @ Konvu)
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)
Julien Kilo - Co-Founder & CTO @ Unknown (yet) · Building an AI company around RAG. Software Engineer with 10 years in the startup ecosystem, passionate about web ecosystem and AI. 🇫🇷 Based in Lyon, France (CET)
Taekmin Kim - Vibe Coder @ bg.app · Building autonomous coding agents running in background. Previously Machine Learning Engineer at Meta and TikTok focused on GPU optimization. 🇺🇸 Based in San Francisco, USA (PST)