How We Explored Weight Space Programming With One AI Rewriting Another
A ThorstenMeyerAI.com report describes MIT weight-space experiments and three follow-up tests, including a narrow benchmark gain and unresolved safety risks.
Checking The Claim That ChatGPT Handles One-Fifth Of Google Searches
OpenAI’s one-fifth comparison is unverified outside the company, while independent models and website traffic panels show lower, differently measured figures.
Progress On Provable Inference For Large Language Models
Anthropic’s open-source FlashProve prototype generates cryptographic proofs for LLM responses, but proving inference remains far slower than running a model.