Progress on provable inference for large language models(anthropic.com)
232 points by ltononro | 118 comments
Anthropic research post introducing FlashProve, an open-source zkSNARK system for cryptographically proving that LLM responses came from a committed model, with Lean-verified guarantees.