What's New in AI

Professor Matthew Schwartz used Claude Opus 4.5 to complete a theoretical physics paper through 270 sessions and 51,248 messages — calling it 'vibe physics.'

A viral Reddit thread draws a striking parallel between LLM hallucinations and split-brain confabulation — and Anthropic's own research on Claude confirms the uncomfortable comparison.

Three days after Google's TurboQuant dropped, a community project called RotorQuant uses Clifford rotors to achieve 10-19x faster KV cache compression.

Security researchers found nearly 3,000 unpublished Anthropic documents in a public data cache — including details on Claude Mythos, a new model with unprecedented cybersecurity capabilities.

A federal judge in San Francisco barred the Department of Defense from designating Anthropic as a supply chain risk, calling the move 'arbitrary and capricious.'

A Reddit user tested Qwen3.5-397B on dual NVIDIA DGX Sparks and a Mac Studio M3 Ultra 512GB. The results split the community — neither machine wins outright.

Mistral AI releases Voxtral TTS, a 4B parameter text-to-speech model that outperforms ElevenLabs in blind tests. Open weights, 9 languages, 90ms latency.

Alex Karp says the future belongs to skilled tradespeople and neurodivergent thinkers. Everyone else should be worried. Here's why he might be right.

Chollet's new benchmark drops the same week Jensen Huang declared AGI. GPT-5.4 scored 0.26%. Claude Opus 4.6 scored 0.25%. The gap with humans is 99+ points.

The ARC-AGI creator fires back at NVIDIA's AGI claims with a benchmark that every frontier model fails. His argument is simple: if humans can do it cold, AGI should too.