Buzzing HN · 中文精选· ltononro·· 9 hours agoSelectedAI score71
Anthropic Releases Verifiable LLM Inference System FlashProve, Achieving 110 tokens/s Proving Speed on 400B MoE
大型语言模型可证明推理方面的进展
AI guide
Anthropic, in collaboration with several universities, has introduced FlashProve, a prototype system for verifiable LLM inference that leverages zk-SNARK cryptographic proofs to verify that outputs originate from a specified model without revealing its weights.
Why it matters
This breakthrough reduces the overhead of zero-knowledge proofs for large models to within 100x of standard inference, offering a software-only verification solution for regulatory audits and model tamper-proofing.
Source:Buzzing HN · 中文精选 · anthropic.com