Multi-head attention divides a model’s query, key, and value representations into multiple attention heads. Each head can learn to emphasize different positions or types of relationships, such as local syntax, long-range dependencies, object associations, or alignment between modalities. The head outputs are combined and projected into the next representation. This structure is a central component of transformer models and provides more expressive interactions than a single attention calculation with the same representation. Individual heads are not guaranteed to have a simple human-readable role, and some may be redundant. Head count, dimension, masking, memory use, and optimized kernels affect both model quality and computational efficiency.
Anthropic unveils a revised Responsible Scaling Policy with a Frontier Safety Roadmap, regular Risk Reports, and clearer separation between company commitments and industry recommendations.
ŌURA unveils a proprietary large language model for women’s health, combining clinical research and biometric data to deliver personalized, privacy-first AI guidance through Oura Advisor.
Anthropic reveals industrial-scale campaigns by DeepSeek, Moonshot, and MiniMax to extract Claude’s capabilities via fraudulent accounts, highlighting national security and AI safety risks.
Former DeepMind researcher David Silver has raised $1 billion for London-based Ineffable Intelligence, aiming to build a superintelligence that learns autonomously through experience.
Google DeepMind has introduced AlphaGenome, a new AI model designed to analyze long DNA sequences and predict how genetic variations may influence disease development.
OpenAI now uses age prediction to adjust ChatGPT’s safety settings for teens. Users 18 and older can verify their age to disable extra restrictions.
OpenAI has invested in Merge Labs, a brain-computer interface startup founded by CEO Sam Altman, as part of a $250 million seed round valuing the company at $850 million.
Google released a major upgrade to its Gemini Deep Research agent, giving developers access to advanced autonomous research capabilities through the Interactions API. The company also introduced DeepSearchQA, a benchmark for evaluating multi-step web research agents.
Google introduced the Titans architecture and the MIRAS framework to enable AI models to handle massive contexts and update their internal memory while running, improving performance in long-sequence tasks.
OpenAI co-founder Ilya Sutskever says the era of scaling compute and data as the main driver of AI progress is ending, arguing that meaningful advances now depend on true scientific breakthroughs and solving AI’s poor generalization.