AI Math Still Needs the Human Mathematician
AI now handles hard proofs in minutes and makes an experienced mathematician roughly 50% more productive, yet its proofs remain tangled and poorly abstracted. Collaborative 'centaur' work still needs humans to own structure and understanding. — lesswrong.com
Python's Experimental Compiler Hits a Dead End
Python's experimental just-in-time compiler is being removed from the standard interpreter after its maintainers withdrew the proposal to make it a supported feature. The steering council had paused the work until the proposal answered questions about maintenance, compatibility, and performance. — lwn.net
Unlimited OCR 2 Parses Books in One Pass
Unlimited OCR 2, a document-reading model, parses book-length documents in a single pass by compressing attention to a fixed set of text-visual references. It achieves state-of-the-art results on ultra-long document benchmarks. — huggingface.co
Hermes Agent Builds Skills From Experience
Hermes Agent, a self-improving AI assistant, creates skills from experience and refines them during use, while building a persistent model of its user across sessions. It runs on anything from a cheap virtual server to serverless infrastructure and keeps working even when the user is away. — github.com
AI Agents Fuel Bursts, Then Burnout
AI agents have pushed an engineering leader back to their busiest coding pace, with the current year already surpassing every full year since that early peak. The work arrives in exhausting bursts followed by burnout, as cheap implementation keeps turning into doing more. — oioannou.com