Frontier models, research results, agents, and developer tooling.
Meta revised its commercial terms of service to drop the word "train," but legal experts say the new language still gives the company broad room to use scraped content for AI products.
DeepMind's new release signals a major update to its Antigravity platform, though the announcement offers no technical specifics.
NVIDIA and the Barcelona Supercomputing Center are expanding startup access to sovereign AI supercomputing and training through a new accelerator programme.
Ai2's FairStress measures whether language models resist changing decisions equally across demographic groups, not just whether final answers match.
New regulatory moves in the US and China are testing the open-weight ecosystem even as DeepSeek releases a new model and its internal inference stack.
IBM Research's ALTK-Evolve shows that distilling past attempts into prompt guidelines can outperform frontier models on a customer-service benchmark without weight updates.
Inception's new diffusion LLM claims a massive speed and long context, but benchmark results are self-reported.
New case studies show Google DeepMind's AI co-scientist generating testable hypotheses for liver disease, from explaining patient-specific drug responses to repurposing existing medicines.
DeepMind's Co-Scientist is being used by researcher Clare Bryant to identify genetic triggers in emerging infectious diseases.
A new MIT experiment suggests AI agents can learn through self-directed play, though skeptics say the behavior may be role-playing.
A practical guide shows how LLMs can assist in the full security lifecycle of source code, from threat modeling to patching.
A hospital partnership with OpenAI shows how AI can support rare disease diagnosis and reduce administrative load.
OpenAI highlights how Braintrust leverages Codex and GPT-5.5 to accelerate prototyping and code iteration.
OpenAI disrupted two separate AI-powered influence operations likely tied to China, aimed at shaping US tech policy debates.
Mistral's new Timestral family, led by a 209M-parameter model, sets state-of-the-art results for zero-shot time-series forecasting and runs on CPU.
NVIDIA's move makes its GPU orchestration stack freely available to developers, potentially reshaping how Kubernetes clusters manage AI workloads.
Hugging Face says TorchNeuron makes running its models on AWS Trainium easier than before.
New benchmarks show GPT-6 models far outperform predecessors on building software and migrating COBOL code, at lower cost.