Your Dose of Reg.exe, Week {28}

Clawdbot/Moltbot/OpenClaw is the new viral phenomenon. Google Genie 3 and LingBot launched AI gaming models. DeepPlanning targets long-horizon tasks. AlphaGenome, Qwen3 Max Thinking, Kimi K2.5 dropped

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

---

👉 Article originally posted on WeLoveSota.com

Events

🇫🇷 Workshop on database security for AI agents in Paris (February 5) Guepard and Microsoft Purview co-organized a workshop showcasing MCP integration for database version control and time travel capabilities, addressing database corruption issues caused by coding agents like Cursor.

Autonomous Agents

🔐 Clawdbot/Moltbot/OpenClaw saga and Moltbook social network - As expected, the viral AI assistant was everywhere despite multiple renamings, from Clawdbot to Moltbot following pressure from Anthropic, then more recently to OpenClaw. After the chaos, and a crypto scam on X using the former name, the project spawned Moltbook, a social network built exclusively for AI agents, where they autonomously share, discuss, and upvote content. (🙏 Gabriel Olympie, Kemal Toprak Uçar, Louis Choquel, Enrico Piovano, Jeremie Kalfon, Robert Hommes, Maxence Maireaux)

• Moltbook has been described as “the most interesting place on the internet right now” by Simon Willison (read the blog post), and “the most incredible sci-fi takeoff-adjacent thing I have seen recently” by Andrej Karpathy on X.

• Moltbook demonstrates emergent agent-to-agent social dynamics with autonomous content creation

• 🔥 Community take: Security concerns dominated the discussion, with members viewing OpenClaw as too risky for use on primary machines due to prompt injection vulnerabilities and other security issues. One member reported spending an hour setting it up on a disposable laptop, finding it barely functional but recognizing its potential direction for future UX. On Moltbook specifically, one member noted: “Honestly, some post are quite funny”

📊 Qwen team evaluates long horizon task performance - DeepPlanning benchmarks LLM agents on realistic multi-day travel and multi-product shopping tasks that require tool-based information gathering, local constraint handling, and global budget/time optimization, showing frontier models still fail many long-horizon plans.(🙏 Gabriel Olympie)

• Models get decent constraint scores but low case accuracy, exposing fragility in end-to-end plans.​

• Internal chain-of-thought and heavier tool use significantly improve effectiveness–efficiency trade-offs.​

• Dominant failures stem from missing tool queries and weak global optimization under interacting constraints.​

🧠 Headroom context optimization layer for LLM applications - Tejas released OSS project for context compression with plans for image, audio, and video support, seeking startup partnerships for enterprises burning tokens in Claude Code or Cursor. (🙏 Tejas Chopra @ Netflix)

• Works seamlessly with Mem0 and Letta frameworks for memory management in agentic workflows

• Significantly reduces token consumption in coding agents, lowering operational costs

• Enterprise features including multi-modal compression currently under active development

🔧 Anthropic integrated interactive MCP apps into Claude - Claude received updates enabling users to access Asana, Slack, Figma, and Box tools directly in chat, with developers able to build apps using MCP and view live tool content mid-conversation.

Biotech Health And Chemistry

🧬 Google DeepMind released AlphaGenome for genetic disease analysis - AlphaGenome AI tool launched to identify genetic drivers of disease, capable of analyzing up to 1 million letters of DNA code at once and potentially paving the way for new treatments. (🙏 Maziyar Panahi @ OpenMed, Quentin Dubois @ OSS)

• Processes massive DNA sequences up to 1M base pairs in single unified analysis pass

• All model weights and implementations open-sourced on Hugging Face for research community

• Represents major advancement in genomic medicine with potential for personalized treatments

• 🔬 Paper, ⚖️ weights and models, 🎙️ roundtable

🔬 Iktos published large-scale compound selection research - Nature paper (co-authored by our member Ihab Bendidi from Recursion!) demonstrated AI-guided compound selection using cell painting with DINOv2-derived phenotypic embeddings from 112k compounds, enabling structure-independent selection and systematic discovery of compounds. (🙏 Victoire Cachoux @ Iktos)

• Phenoseeker platform enables advanced screening workflows without chemical structure dependency

• Batch-corrected DINOv2 embeddings from 112k compounds improve hit discovery systematically

• Successful collaboration between Iktos and Recursion teams combining AI with cell imaging

🧠 State of brain emulation assessed in comprehensive report - Research evaluated current capabilities and limitations in brain emulation, highlighting fundamental gaps including inability to simulate single neurons and incomplete knowledge of neuronal types. (🙏 Gabriel Olympie, Jeremie Kalfon @ Pasteur/ENS)

• 🔥 Community take w/ Jeremie Kalfon: “To me, we’re at zero here. I mean, we can’t even simulate a single neuron. We don’t even know all the neuronal types in the brain, and for the ones we do know, we mostly don’t understand what they do or how they’re involved in computation.”

🦠 EBV virus linked to multiple sclerosis and other diseases - Analysis revealed that over 90% of people get infected with EBV (mononucleosis) in teens, with some developing diseases later including MS, cancers, skin diseases, dementia, and Parkinson’s due to lack of effective vaccines. (🙏 Jeremie Kalfon @ Pasteur/ENS)

• 🔥 Community take w/ Jeremie Kalfon: “We prefer some people to get cancer, MS, or Parkinson’s rather than give a virus to people who will get it anyway. Many people might volunteer. Indeed, we give so much to associations fighting cancer, MS, and dementia, but when it comes time to actually do something, it seems we don’t want to.”

💊 Hologen AI biotech startup raising $150M Series A - Former Google CEO Eric Schmidt co-founded secretive AI biotech company seeking significant funding, though details remained sparse with minimal public information. (🙏 Ihab Bendidi @ Recursion, Jeremy Kalfon @ ENS/Pasteur)

Image Video And 3d

🎬 VideoMaMa mask-guided video matting released - KAIST lab published generative prior-based video matting tool under non-commercial license, addressing complex rotoscoping challenges for production workflows. (🙏 Alvaro Lamarche Toloza @ Mago)

• 🔥 Community take w/ Alvaro Lamarche Toloza: “Rotoscoping is a much more difficult / complex task than it looks to be used in production”

⚡ Pruna.ai accelerated FLUX.2[flex] by 3x - Partnership with Black Forest Labs delivered 3x speedup for production-grade FLUX.2[flex] image generation and typography model through optimization techniques. (🙏 Amine Saboni @ Pruna.ai)

🎨 Hunyuan 3D 3.1 Pro and Rapid launched on fal - Tencent released high-fidelity Image-to-3D and Text-to-3D generation models with Pro version for quality and Rapid for speed, featuring smart topology and part generation for advanced workflows.

• Two versions optimized for different use cases: Pro for highest fidelity, Rapid for speed

• Smart topology generation and automatic part segmentation enable advanced 3D workflows

• Available through fal platform with API access for easy integration into pipelines

👁️ Youtu-VL lightweight vision-language model released - A 4B-parameter vision-language model from Tencent’s Youtu Lab using Vision-Language Unified Autoregressive Supervision (VLUAS). It treats visual tokens as prediction targets alongside text, boosting vision tasks like detection, segmentation, and pose estimation without extra modules, while excelling in VQA and OCR.

👔 FASHN open-sourced VTON v1.5 for virtual try-on - Station F Paris-based startup released maskless virtual try-on model generating photorealistic results directly in pixel space without segmentation requirements. (🙏 Pierre Chapuis @ Finegrain)

Language Models

🧠 Qwen3 Max Thinking model announced without open-source release - Qwen3-Max-Thinking is a scaled reasoning model with RL-tuned tool use and test-time scaling, matching frontier models on 19 benchmarks and exposed via OpenAI/Anthropic-compatible APIs.​ (🙏 Enrico Piovano @ Goji, Gabriel Olympie)

• RL and scaling give frontier-level scores on knowledge, reasoning, and alignment vs GPT-5.2-Thinking, Claude-Opus-4.5, Gemini 3 Pro.​

• Adaptive Search/Memory/Code tools reduce hallucinations, add real-time + personalized + code-based reasoning.​

• Test-time “experience-cumulative” multi-round reflection beats naive multi-sample decoding at similar token budgets on GPQA, HLE, LiveCodeBench, IMO, HLE+tools.

• 🔥 Community take: The new model sparked some criticism for not open-sourcing its most powerful versions for now. “The team appears to be back in teasing mode. Fingers crossed for an updated Qwen Next 80B. A distilled version of the max-thinking model would be a major win”.

🇨🇳 Chinese AI models lagging US frontier by 7 months - Epoch AI analysis showed Chinese models averaging 7-month lag behind US frontier since 2023, widening from 3 months in October.

📦 Snowflake Arctic Embed XS with only 22M parameters - Very small 22M-parameter, 384-dim text embedding model that gets strong MTEB retrieval scores (NDCG@10 ≈ 50.15) while staying ultra-fast and cheap, close in quality to ~100M‑parameter English retrievers. (🙏 Youssef Tharwat @ noodlbox)

• 🔥 Community take w/ Youssef Tharwat: “22M is crazy! Useful for someone with really strict on-device conditions.“

🚀 Kimi K2.5 model released by Moonshot AI - New model from Chinese AI company launched with availability through Hugging Face and Ollama integration. Kimi K2.5 is a 1T-param MoE native multimodal agentic model (32B active, MoonViT vision, 256K ctx) with strong reasoning, vision, coding and long-context benchmarks, plus int4-friendly deployment and dual thinking/instant modes. (🙏 Hugo Hernandez @ Alakazam)

🇺🇸 Arcee AI launched Trinity model family - US-built open-weight MoE models delivered reliable reasoning, tool use, and long-context support across multiple sizes with American infrastructure. (🙏 Enrico Piovano @ Goji)

⚡ Cerebras pruned MiniMax model for single GPU - MiniMax-M2.1-REAP-139B-A10B is a 139B-parameter sparsely activated language model made by removing 40% of MiniMax-M2.1’s experts with the REAP pruning method, keeping similar quality but using less memory. (🙏 Gabriel Olympie)

• 🔥 Community take w/Gabriel Olympie: “That should fit on a single RTX6000 (haven’t tried it yet)”

🎓 Yann LeCun launched AMI Labs with contrarian bet - AI pioneer started Paris-based venture betting against large language models in favor of alternative architectures, challenging current industry direction. (🙏 Kemal Toprak Uçar @ Numberly)

💰 OpenAI unit economics showed path to profitability - GPT-5-era analysis revealed plausible trajectory to profitable operations, addressing concerns about long-term business model sustainability.

Programming

🤖 Karpathy reported 80% agent coding workflow by December - He acknowledged a personal rapid shift from 20% to 80% agent-based coding between November and December, with prediction of “slopacolypse” in 2026 as AI-generated code proliferates.

📊 Claude Code Opus 4.5 performance tracker launched - Marginlab released daily performance monitoring for Claude Code on SWE-Bench-Pro with statistical significance testing to detect degradation. (🙏 Quentin Dubois @ OSS)

• Continuous daily performance monitoring tracks Claude Code reliability over time on real tasks

• Statistical significance testing detects performance degradation with mathematical rigor

• Public transparency dashboard allows community oversight of model quality and consistency

⚡ Mistral Vibe 2.0 released with improved performance - Updated coding model delivered faster, higher-quality results for development workflows. (🙏 Quentin Dubois @ OSS)

Robotic World Ai

🤖 Google DeepMind unveiled Project Genie interactive world creation - Experimental research prototype enabled users to create, edit, and explore virtual worlds through natural interaction, demonstrating advanced world modeling capabilities. (🙏 Hugo Hernandez @ Alakazam)

• Users can create and edit virtual environments through natural language and interaction

• Real-time world exploration allows immediate testing and refinement of generated spaces

• 🔥 Community take w/ Hugo Hernandez: “It’s very impressive. It seems able to generalize beyond SOTA open-source world models, although I do recommend taking a look at Wordplay 1.5 and LingBot. That said, the model’s action space remains poorly controllable, which is the same limitation those models have, and even Google isn’t immune to the timed-session constraint.”

🌍 Lingbot open-source world model challenged Genie 3 - An open-source, high-fidelity world simulator built on Wan2.2, enabling interactive, minute-scale 480p/720p video generation with camera/action control, multi-GPU long-horizon inference, and Apache-2.0 licensing. (🙏 Gabriel Olympie)

• Supports <1s latency at 16 FPS for real-time interactive world simulations across diverse visual domains.​

• Uses FSDP + DeepSpeed Ulysses for long-horizon multi-GPU inference (e.g., 1-minute videos at 16 FPS).​

• Provides camera-pose–controlled base model now; action-controlled and fast variants are planned releases.

🦾 Figure AI introduced Helix 02 with full-body autonomy - Latest humanoid robot generation featured complete autonomous control, though safety concerns persisted following previous lawsuit. (🙏 Pierre Chapuis @ Finegrain, Guillaume Allegre @ Andromede, Jeremie Kalfon)

• Full-body autonomous control system enables coordinated manipulation and mobility tasks

• Advanced dexterous manipulation capabilities combined with dynamic bipedal locomotion

• 🔥 Community take: Implementing the safety features needed to avoid any possible robot incident might be costly or slow, and the community questions whether there’s a real technical limitation or simply a lack of care. While Figure AI regularly showcases impressive demos, the recent safety lawsuit has fueled skepticism, particularly about how much of the behavior shown is truly generalized versus scripted.

🌍 NVIDIA launched Earth-2 open climate models - First complete family of open climate AI models spanning data processing to 15-day global forecasts and local storm predictions, making weather AI accessible worldwide.

Infrastructure

🔧 Plakar v1.1.0-beta released with improved UX - A stable, backward-compatible backup engine refresh with a revamped UI, faster restores, lower resource usage, simpler integration APIs, and a roadmap for PITR and multi-source snapshots.​

• Major gains in restore speed and reduced RAM/disk footprint make large-scale backups more operationally efficient.​

• New package manager and simplified importer/exporter/store interfaces cut integration friction for custom backends.​

• Replacing the agent with the cached service improves reliability while enabling safer concurrent CLI operations

New Member

🇦🇪 Matt Suiche (OnchainDB) - CEO at OnchainDB building data infrastructure for AI agents with built-in monetization for data providers. Serial founder with two exits: CloudVolumes to VMware (2014) and Comae Technologies to Magnet Forensics (2022). Background in memory forensics and detection engineering, founded OPCDE community. Hobbies include Brazilian Jiu-Jitsu and drifting. Located in Dubai, UAE.