A curated AI research newsletter covering four major topics: (1) A proof-of-concept self-replicating AI worm from University of Toronto/Cambridge researchers that uses stolen GPU compute to run open-weight LLMs for autonomous vulnerability detection and exploitation, achieving ~37% full attack success rate. (2) Dwarkesh Patel's argument that compute prices will rise dramatically as AI systems become more capable and better monetize hardware. (3) A statement signed by ~1337 AI lab employees from OpenAI, Anthropic, Google DeepMind, and others requesting US government support for international efforts to pace frontier AI development. (4) Research showing current AI systems excel at engineering tasks but fail at creative, paradigm-shifting research — with Claude Opus 4.8 producing papers rated 'Reject' and 'Strong Reject' at NeurIPS 2026 — contrasted with OpenAI's internal Astra model solving ten open math and CS problems.
2K Impressions1 Comment