AI
Artificial intelligence: LLMs, agents, diffusion, vision, NLP and the latest from top labs.
593 published articles
Embodied AI
Mistral's tiny 8B model just made lidar optional for office robots
Mistral AI's Robostral Navigate is an 8B model that uses only one RGB camera to achieve 76.6% on R2R-CE benchmarks, outperforming multi-sensor approaches by 4.5 points. Built with simulated data and a token-efficient prefix-caching method, it generalizes across robot types and adapts to unseen obstacles.
2026-07-17
Intelligence Artificielle
Google tue NotebookLM pour le remplacer par un agent de code, et c'est un signal
Google renomme NotebookLM en Gemini Notebook, ajoutant un ordinateur cloud sécurisé pour l'exécution native de code. L'outil se synchronise désormais avec l'application Gemini et sera bientôt disponible dans AI Mode de Search, transformant un assistant de notes en hub personnel d'IA.
2026-07-17
Infrastructure Intelligence
Your AI agent just blew 140,000 tokens on a weather check. Now you can see why.
Alibaba Cloud's Agentic OS (ANOLISA) now includes AgentSight, an observability widget that reveals exactly how agents burn tokens, per session, per dialog, per skill call. With agents consuming tens of thousands of tokens on trivial tasks, the tool promises to turn vague monthly bills into an actionable cost ledger.
2026-07-17
Multimodal AI
Mistral's 12B model just embarrassed a 90B one. The scaling orthodoxy has a problem.
Pixtral-12B matches or beats models seven times its size on multimodal benchmarks, without sacrificing language performance. Mistral also releases a new open benchmark for practical vision-language evaluation, challenging the idea that bigger is always better.
2026-07-17
AI Research
The one training trick that stops AI agents from freezing in production
Current LLM agents crumble under real-world randomness. NoisyAgent exposes them to controlled noise during training, improving both robustness and general benchmark performance. The paper suggests the field has been overfitting to pristine conditions.
2026-07-17
LLMs & Models
Mistral's cascade recipe shrinks LLMs without killing reasoning
Mistral's cascade distillation shrinks large models into small ones while preserving reasoning and vision. The 3B variant packs capabilities that used to require ten times the parameters, and it's all Apache 2.0.
2026-07-17
AI Research
The AI agent bottleneck isn't exploration. It's knowing what good looks like.
Two new Hugging Face papers tackle the same core problem from opposite directions: how to make AI agents reliably evaluate their own actions. AJ-Bench builds a benchmark for environment-aware judge agents, while HeavySkill argues the best judge lives inside the model's parameters.
2026-07-17
AI Image Generation
The one thing Midjourney still can't do: render a Chinese poster correctly
Zhipu AI's GLM-Image fixes a glaring blind spot in AI image generation: accurate Chinese text rendering, for free, with no limits or watermarks. A practical tool for Chinese-language creators.
2026-07-17
benchmark breakdown
The 2.8 trillion parameter model that beats the frontier on the benchmarks that matter
Kimi K3, the 2.8T-parameter open model from Moonshot AI, trails frontier proprietary models on most broad benchmarks, but leads on SWE Marathon, Terminal-Bench 2.1, BrowseComp, and others. The detailed table reveals where its architectural bets on KDA and Stable LatentMoE pay off.
2026-07-17
Open-weight models
Kimi K3 is open, 2.8 trillion parameters strong, and still hunting for a win against the frontier
Kimi K3, the largest open-weight model at 2.8T parameters, lands with novel attention architectures and a 1M-token context. On benchmarks it competes in the upper tier but trails Claude Fable 5 and GPT 5.6 Sol, challenging the narrative that open models are closing the gap to proprietary frontiers.
2026-07-17
AI Labs & Research
An AI that can see its own mistakes and undo them
A training-free AI sampler that generates images and text together, retracting decisions when cross-modal signals contradict each other. The paper's CO₂Jump outperforms existing methods and suggests a structural fix for a blind spot in multimodal AI.
2026-07-17
AI Strategy
Alibaba's $53 billion bet: from lab curiosities to hospitals and pig farms
Alibaba details a first-half 2026 AI update: a new token business unit, Qwen3.7-Max, video model HappyHorse, and a $53 billion infrastructure build-out. The strategy moves Alibaba from research to real-world enterprise and consumer AI deployment.
2026-07-17