AlphaGo creator bets $1.1B against LLMs

WIRED

David Silver's new startup Ineffable Intelligence has raised $1.1 billion to build superintelligence through reinforcement learning instead of LLMs, arguing that current AI approaches are fundamentally limited by their reliance on human-generated training data rather than independent learning and discovery.

TechCrunch

Thinking Machines Lab, valued at $12 billion, is engaged in a bidirectional talent war with Meta, recruiting PyTorch co-founder Soumith Chintala and numerous researchers while securing multibillion-dollar cloud infrastructure from Google with access to Nvidia's latest GB300 chips.

Scientific American

An amateur using ChatGPT Pro solved a 60-year-old Erdős conjecture that stumped expert mathematicians by prompting the AI, which discovered a novel proof method no human had considered. The breakthrough required expert refinement but reveals potentially broader applications, marking a significant advance in AI's mathematical reasoning capabilities.

Video intelligence in 2026 achieves hour-long understanding and real-time on-device tracking through universal compact encoders, adaptive temporal compression, and deployment-optimized quantization. The shift from specialist models to multi-task universal encoders (sub-100M parameters) and query-aware token budgeting enables practical deployment from mobile devices to cloud, though streaming inference and sub-watt AR use cases remain open challenges.

Testing Catalog

Anthropic is developing Bugcrawl, an AI agent for Claude Code that automatically hunts for bugs across entire codebases, expanding its suite of automated code analysis tools to compete with OpenAI, xAI, and Google in repository-scale code reasoning.