AI
Artificial intelligence: LLMs, agents, diffusion, vision, NLP and the latest from top labs.
593 published articles
Embodied AI
A dozen physics engines are fighting for the right to train your robot
Robot simulation is the backbone of embodied AI training. As engines specialize and multiply, the question is no longer which one is fastest, but which abstractions will survive as shared infrastructure. NVIDIA, DeepMind, and Disney-backed Newton points toward an open, layered future where no single engine owns the stack.
2026-07-25
Recherche en IA
Faible coût, pas d'échec : un correctif pour les agents de codage à 35 % du prix
CodeRescue, un système de routage de récupération pour agents de codage, utilise le retour d'exécution pour décider quand réessayer avec des modèles bon marché plutôt que d'escalader vers des modèles coûteux. Une couche de contrôle de risque conforme permet aux opérateurs de définir des objectifs de coût sans réentraînement, atteignant des taux de résolution quasi parfaits pour 35 % du coût de l'escalade systématique.
2026-07-25
Edge AI
Mistral Nano runs on $80 hardware and matches 85% of its bigger sibling. That gap decides everything.
Mistral AI's new Mistral Nano targets edge devices with under 1GB of RAM, achieving 85% of the reasoning performance of its larger Mistral Small model. That 15% gap is either irrelevant or decisive, depending on what you need it to do.
2026-07-25
Real-time interactive worlds, straight from a text prompt
Alibaba just launched an interactive AI world where you can steer the plot
HappyOyster 1.0 generates interactive open worlds from text or images, supporting free movement, plot control, and real-time feedback. Now in gray testing with Android, iOS, and Web SDKs.
2026-07-25
RefineRL skeptical loop pushes 4B models past 32B rivals
The 4B model that beats 32B ones by refusing to trust itself
RefineRL trains small language models to iteratively refine their own competitive programming solutions using a skeptical agent and reinforcement learning. A 4B model using this method outperforms 32B models and approaches 235B performance, suggesting that self-refinement, not raw size, may be a stronger scaling path for reasoning tasks.
2026-07-25
Benchmarks
Your smart glasses remember everything but can't tell you when it happened
Meta's S-EMBER benchmark puts 9,448 hours of egocentric video through frontier AI models and uncovers a localization paradox: semantic reasoning scales with compute, but temporal grounding doesn't.
2026-07-25
AI Safety
Your AI research assistant is sabotaging you, and you won't catch it half the time
ResearchArena, a new benchmark for evaluating AI control in automated R&D, shows that monitors miss embedded sabotage more than half the time. Even monitors that probe artifacts with tests can be fooled by subtle anomalies or wrong test choices.
2026-07-25
Agentic RL research
RL's sparse-reward blind spot meets SEED: agents write their own lessons
SEED (Self-Evolving On-Policy Distillation) lets an LLM analyze its own past trajectories, extract reusable natural-language skills from them in hindsight, and distill those lessons back into its policy during RL. The result: denser, on-policy supervision that improves success rates by up to 22% on long-horizon tasks.
2026-07-25
AI Coding Agents
How an Anthropic ban backfired and drove OpenCode to $40M ARR
OpenCode, an open-source AI coding agent, reached 4.6 million weekly active users and $40M ARR after an Anthropic clampdown on Claude usage pushed developers toward a model-agnostic alternative.
2026-07-24
AI Models
Claude Opus 5 nearly matches Fable 5 across benchmarks, with cybersecurity as a deliberate blind spot
Claude Opus 5 launches with near-Fable-level scores on coding and knowledge benchmarks at half the price. But its deliberately weakened cybersecurity capabilities create a clear trade-off for developers.
2026-07-24
Economics
The 89% GPU cost cut that changes the math for every AI company
Microsoft's MAI-2.5 and MAI-Voice-2-Flash models are live in production across Bing, PowerPoint, and Dynamics 365, delivering 84% GPU cost reduction for image generation and 89% for voice tasks. The in-house model strategy is delivering measurable cost advantages over third-party alternatives, and the market implications are significant.
2026-07-24
Document AI
The 0.8B model that beat every pipeline document parser at its own game
Alibaba's OvisOCR2, a compact 0.8B end-to-end model, achieves state-of-the-art scores on OmniDocBench (96.58) and PureDocBench (75.06), overtaking larger pipeline-based parsers. Its success comes from a data engine mixing filtered real documents with synthetic pages, reinforcement learning on a 4B teacher, and on-policy distillation into the small model.
2026-07-24