Events

🇫🇷 NeurIPS Unplugged in Paris (December 10) – TTY, Dust and 2050 will host live Q&A with AI researchers including Stanislas Polu, then drinks at Dust, in partnership with Nebius.

🇫🇷 NeurIPS After Hours in Paris (December 17) – Pierre Marion and Gaël Varoquaux (Inria, scikit-learn) will debrief NeurIPS’ main themes, then open Q&A. Hosted by TTY, Dust and 2050, in partnership with Nebius.

🇫🇷 State of AI 2026 launch in Paris (November 3) – Air Street presented this year’s State of AI report at Google, with talks on robotics data and autonomous air defense.

🇺🇸 Backprop Summit in San Francisco (November 17) – A tokens& day on frontier models and AI infra, with Jürgen Schmidhuber, Sara Hooker and Wes McKinney.

🇫🇷 Agent Stack Series #00 x Holo AI Night in Paris (November 18) – Vercel and H Company will give talks on System 1 models, followed by Q&A, food and drinks.

Audio

🗣️ tts-engine gives Reachy Mini a voice – Open-source Python engine by TTY member Thibault Hervier that streams speech chunk by chunk from 12 engines, local (Kokoro, Piper, Qwen3-TTS) or hosted (ElevenLabs, Cartesia). On a Pi 4, first audio came fast, keeping up with real time did not, per its benchmarks. Reachy Mini also performed a slam with the voices.

Autonomous Agents

🧳 Jinko launches travel rails for agents – Jinko, co-founded by TTY member Shuai Zhang, launched publicly: a travel search and booking layer that lets AI agents answer travel questions, avoid getting blocked by airlines like United and American, and earn commissions. Exa and Pally already used it.

Biotech, Health, and Chemistry

🧬 Carbon-X revives the text-plus-modality debate – Carbon-X, a text-plus-DNA model from Hugging Face (distinct from Carbon and Carbon-A), aimed to let an LLM interpret DNA sequences. It reopened an old question: do text-plus-modality models actually use the second modality, or just the text?

  • Community take w/ Jean du Terrail of Living Models: Reasoning across modalities has real merit, but evaluation has always been the weak spot. In visual question answering, “the model just needs to look at the word ‘tennis’ to answer yellow and be right lots of the time, despite being totally unable to reason over the image.” For DNA plus text, “this is even worse: answers to all the questions they test on are already present in the base LLM training data,” which makes leakage very hard to assess. Hugging Face has been prioritizing speed over quality, “which is just a strategy, not necessarily bad in itself.” Since Flamingo and LLaVA in 2022-2023, text-plus-radiology, omics (ChatNT) and histopathology papers mostly planted the same “first text plus modality” flag with the same flaw. “People are desperate to do RL on biology, and using MCQ with text is the only way to do it right now.”

🔬 OpenAI BioBasicsBench tests wet-lab agents – A 100-task benchmark built with enzyme-design company Ligo checked whether agents could read raw lab evidence like gel images, microscopy stacks and plate photos. GPT-6 Astra scored 66.7% but still let familiar assumptions override experimental controls.

🧫 Flagship merges perturbation biotechs into Etcelera – Flagship Pioneering folded Cellarity, Etiome and Empress Therapeutics into Etcelera, which trains models on how cells respond to genetic, chemical and biological perturbations, building on a foundation model trained on 150 million cells.

Image, Video & 3D

🖼️ Pruna turns MiniMax-H3 into P-H3-Image – Pruna rebuilt the MiniMax-H3 video model into an open-weights text-to-image and editing model, betting that video training already teaches objects from many angles, lighting on materials and body motion. The result was 7x smaller (20GB vs 143GB) and 5x faster, rendering 8MP images in about 21GB of VRAM.

  • Every component was reworked: the text encoder was distilled from 32B to 4B parameters, the backbone shrank from 33B to 20B, and distillation cut generation from 50 steps to 8.

  • It rendered 2 to 16 megapixel images, was strongest at photorealism, portraits with real skin texture and film stills, and could combine up to 8 reference images into one scene.

  • Weights shipped in bf16, 8-bit and 4-bit for Diffusers and ComfyUI under the MiniMax H3 Community License, which carries geographic restrictions, while dense text and small hands remained weak spots.

💸 Linum trains P-JiT for $21K – Two-person lab Linum trained a 2B-parameter text-to-image model directly in pixel space, skipping the usual VAE compressor, which cut image tokens 4x more than standard setups. Weights were released under Apache 2.0 as a step toward video models.

✏️ Krea 3 technical report on editing – Krea 3 edited images from up to 8 references while keeping characters and products consistent, finishing in 4 to 8 seconds, which Krea claimed was 3-5x faster than Nano Banana Pro. Reinforcement learning cut text rendering errors by 25%.

🧑 Atmee v2 renders real-time video avatars – Atmee v2 created a talking video character from one image and 10 seconds of audio, with contextual expressions and gestures for live calls. Its creator, whose AR startup Google acquired, claimed it ran up to 50x faster than Seedance 2.5 and Wan 3.0.

Cyber

🔐 Anthropic FlashProve proves which model answered – Anthropic’s open-source FlashProve used zero-knowledge proofs to show a response came from committed model weights without revealing them. One H100 proved Llama 4 Maverick at about 110 tokens per second, at 70-95x normal GPU time, with 1,024-token contexts.

  • Community take w/ Mario Cornejo: “This is really interesting, especially when the whole internet starts saying ‘they nerfed the model’ a few days after every release.”

Infrastructure

🖥️ 2501.ai ranks open models for infra – TTY member Gabriel Olympie’s team at 2501.ai tested 24 open-weight models on fixing broken servers and handling IT requests. DeepSeek v4.1 Flash nearly matched the leader on one server, and Qwen3.8 27B reached 93% of the top score on consumer hardware.

  • Only 2 of 24 models followed every company rule across 60 incident runs, and tailing rotated log files was the most common violation.

  • Fixing machines decided rankings more than understanding requests: Kimi K3 led on fixes despite ranking last among the top 10 at reading requests.

  • Cost per solved incident varied 14x, and Qwen3.8 2.4T cost 4.6x more per fix than Qwen3.8 27B while ranking lower.

Language Models

🧠 GPT-6 Astra edits Qwen weights directly – Researchers had GPT-6 Astra read and rewrite raw network weights with no training. Editing 0.016% of Qwen3.5-0.8B removed a reasoning loop and lifted MATH-500 from 7.9% to 48.1%, but it failed on Gemma 3 270M, which had no such flaw. Liquid AI’s Maxime Labonne posted a breakdown.

  • Community take w/ Gabriel Olympie of 2501.ai: “It overclaims and underproves. The one impressive feat is the zero-shot writing of weights for playing small games. The Qwen edit is mostly spotting one defective neuron, and the result does not generalize to other models or non-defective ones. A more suitable title would be: a frontier coding agent can do the work of a capable interpretability engineer on raw weights, cheaply, and that sometimes beats a weak training baseline.”

🪞 JEPAs prove the Platonic Representation Hypothesis – A paper co-authored by Yann LeCun turned the idea that models trained on different modalities converge to one shared picture of the world into a theorem. The condition: training pairs must be “good views”, two observations of one scene, like two photos or two captions of the same thing.

  • Next-word prediction, the way most LLMs learn, falls short: the start and end of a document share the author’s style, so the model learns style instead of the world and needs 42x more data to catch up.

  • A small 109M-parameter text model trained on good views lined up with how image models see the world better than any LLM tested up to 72B, and making LLMs bigger barely helped past a few billion parameters.

  • The small model was also better at judging whether two sentences mean the same thing, and the results held across every family of image model tested.

🎞️ LeVJEPA 1.5 Large closes V-JEPA gap – Lukas Kuhn’s LeVJEPA video encoder, whose first paper Yann LeCun co-authored, reached version 1.5 Large: it matched V-JEPA on image and video benchmarks with 5.6x less compute and over an order of magnitude less data.

✒️ Thinking Machines ships Inkling 2 Small – A 594B mixture-of-experts model (17B active) that takes images and targets agentic coding, vision and knowledge work. Weights stayed closed on Tinker at launch, with an Apache 2.0 release planned after external safety testing and red-teaming.

🤯 QwenPU runs an LLM inside Qwen – After Qwenputer compiled a RISC-V CPU into Qwen 3.5 4B’s weights, the same developer compiled a neural processing unit into the model and ran Karpathy’s tiny stories260k LLM on it, with matrix multiplications and attention computed bit-exact.

🌊 Ricursive releases Ripple-65B Preview – A 65B dense model pretrained from scratch on 12T tokens for agentic coding. Ricursive claimed it beat much larger US open-weight models on TerminalBench 2.1 (80.52) for a $32M total R&D cost, and it fits on a DGX Spark. Weights are on Hugging Face.

  • The team chose a thin-and-deep dense Transformer (80 layers, narrow width) over mixture-of-experts to keep infrastructure simple.

  • Reinforcement learning reached that score with 1.4M rollouts, roughly 70x fewer than Reflection Beam.

  • A team of 1 to 5 people trained it at 40% hardware utilization with no custom kernels.

🤝 Hightouch pairs Jev with Opus – Hightouch used LLMs to spot customer records wrongly merged into one person. Jev handled easy cases and escalated uncertain ones to Claude Opus 5.5, reaching 98.8% accuracy versus 99.0% for Opus alone at over 4x lower cost, thanks to well-calibrated confidence.

Programming

🛒 Mistral acquires SkipLabs for code generation – Mistral bought SkipLabs, home to TTY member Hugo Venturini, which builds technology to make AI code generation faster, more reliable and safer. The team joined Guillaume Lample’s science group in Mistral’s fifth acquisition of the year.

⚡ Cognition releases SWE-2 Flash – Devin maker Cognition launched a fast, efficient coding model and claimed it beat every American open-weight model on leading evals, including ones 3x larger. Weights were not public: Cognition planned a program to share them securely with partners.

Robotic, World AI

🌐 3D World Simulator hits 50 FPS – An action-conditioned world model that predicted robot scenes in 3D, producing consistent views from any angle, depth, geometry and touch signals at 50+ FPS on an RTX 5090. Policies trained only inside it transferred to real robots.

🎮 World Labs Frameboy streams Atlas worlds – A real-time demo for exploring generated worlds, powered by an optimized version of the Atlas autoregressive world model with pixel-perfect camera control and streamed from cloud GPUs at up to 15 fps, two months after Atlas launched.

🤸 ACP playground puts humanoids in browsers – A browser demo ran a 23-joint humanoid in MuJoCo, steered by keyboard, gamepad or phone, with a contact-planning policy at 50 Hz that vaulted, climbed, jumped and stood on a vaulting wall.

New members

🇫🇷 Denis Leroy – Principal Engineer @ SoundHound · Builds voice-native conversational AI agents for brands and cars. Systems and OS generalist with 10 years in the Bay Area (Stanford, SGI, Sun and several startups), into Linux and FOSS. Enjoys cooking, wine tasting and watching basketball. Special power: I can explain baseball. Paris, France

Contributors This Week

Gabriel Olympie, Félix Raimundo, Robert Hommes, Amine Saboni, Denis Leroy, Pierre Chapuis, Emmanuel Benazera, Jean du Terrail, Mario Cornejo, Mehdi Medjaoui, Nancy Wang, Quentin Dubois, Shuai Zhang, Thibault Hervier