Curated coverage of artificial intelligence breakthroughs, research papers, industry moves, and policy developments.
Anthropic's latest flagship model brings expanded reasoning capabilities and a million-token context window, enabling complex multi-document analysis in a single pass.
A new architecture combining chain-of-thought verification with symbolic solvers scores 92% on IMO-level problems, surpassing prior methods by 15 points.
The first compliance deadline under the EU AI Act has arrived, requiring providers of high-risk AI systems to register in the EU database and meet transparency obligations.
The Blackwell Ultra B300 series promises 2.5x inference throughput over the previous generation, with major cloud providers announcing day-one availability.
Researchers detail the engineering challenges and architectural decisions behind training a 2-trillion parameter sparse MoE model on 16,000 GPUs, with implications for efficient scaling.
A community-trained open-weights model achieves GPT-4V-level accuracy on visual question answering and document understanding, with full weights and training code released.
Enterprise users can now build and deploy custom AI agents that operate across Word, Excel, Teams, and Outlook — marking a shift from chat assistants to autonomous task completion.
The revised framework introduces mandatory red-teaming guidelines for frontier models and new metrics for measuring societal impact of deployed AI systems.
A survey of 140+ papers comparing RLHF, DPO, RLAIF, and constitutional methods finds that direct preference optimization consistently matches RLHF at a fraction of the compute cost.