Les derniers reportages, explications et analyses dans AI Models.
28 articles
As major AI labs prepare for public offerings, the industry faces a reckoning over soaring token costs. Shifting from flat-rate to usage-based pricing and the withdrawal of investor subsidies signal a new era of corporate budget constraints and evolving risk disclosures.
OpenAI has deployed a major architectural overhaul to ChatGPT’s memory system, introducing the Dreaming V3 mechanism. The update directly addresses persistent issues with stale and inaccurate recollections, cuts compute requirements to one-fifth for free-tier usage, and expands memory capacity for paid subscribers while preparing for a broader public rollout.
A bipartisan US House coalition has released a draft of the Great American Artificial Intelligence Act of 2026, proposing biannual safety audits for major AI labs, a three-year freeze on state-level AI development regulations, and expanded whistleblower protections.
Google has released Gemma 4 12B, a new open-source AI model featuring a novel Unified architecture that processes text, images, audio, and video natively without external encoders. Optimized for consumer hardware, it runs efficiently on 16GB RAM laptops and supports a 256K context window, advanced reasoning, and local deployment via the AI Edge Gallery.
ModelBest, Tsinghua University, and the OpenBMB community have released MiniCPM-V 4.6, a highly efficient 1.3B-parameter multimodal model. Featuring Instruct and Thinking variants, it outperforms larger competitors in key benchmarks, operates on just 6GB of RAM, and leverages the new LLaVA-UHD v4 architecture to accelerate on-device AI deployment across smartphones, PCs, and IoT ecosystems.
Stability AI has released Stable Audio 3, a next-generation latent diffusion model for high-quality stereo audio generation. Featuring a novel 4096x compression autoencoder and optimized architecture, it delivers variable-length audio generation at unprecedented speeds while open-sourcing smaller and medium model weights.
OpenAI’s advanced reasoning model has autonomously cracked the Erdős unit distance problem in combinatorial geometry, a mathematical challenge that has puzzled researchers since 1946. The breakthrough highlights AI's evolving role from computational assistant to independent scientific discoverer.
YouTube is transitioning from creator-led AI disclosures to automated detection, applying prominent, permanent labels to videos featuring significant photorealistic AI. The policy update aims to improve transparency as generative video models rapidly advance.
DeepMind introduces AlphaProof Nexus, a multi-agent AI framework that combines large language models with formal verification to solve long-standing mathematical conjectures, including 56-year-old Erdős problems and a 15-year Hilbert function challenge.
Baichuan AI has introduced Baichuan-M4, a specialized medical large language model that claims to reduce factual hallucinations to just 3.3% using a novel reinforcement learning algorithm. Alongside the model, the company launched Baixiaoyi, an AI family doctor integrated into WeChat, aiming to shift AI healthcare from passive consultation to proactive, long-term health management.
OpenAI announces its reasoning model has solved a famous unsolved geometry conjecture from 1946, with backing from mathematicians who previously criticized the company's premature claims about AI mathematical capabilities.
Le dépôt de prépublications influent arXiv imposera des bannissements d'un an aux chercheurs qui soumettent des articles contenant des preuves évidentes de contenu généré par l'IA non vérifié, intensifiant ainsi les efforts visant à lutter contre la recherche assistée par LLM de mauvaise qualité.