Anthropic's FlashProve proves which model wrote a reply, and its 72x cost (theainewsreport.com)

Anthropic has introduced FlashProve, an open-source prototype that employs zero-knowledge proofs to verify which specific AI model generated a given output without revealing the model's internal weights. This innovation addresses the critical issue of trust in AI services, allowing users to confirm that the response they received came from the declared model, thereby preventing the substitution of cheaper or modified models. The prototype showcases a significant processing cost, operating at approximately 72 times the time of standard model serving, processing around 110 tokens per second on an H100 GPU, while sticking to a context limit of 1,024 tokens. The implications for the AI/ML community are substantial, particularly for API users concerned about model integrity and authenticity. FlashProve utilizes a mixture-of-experts architecture to enhance efficiency, proving outputs by only activating a fraction of the total model parameters for each token. However, the prototype is not without limitations; it currently supports only greedy decoding and lacks coverage for floating-point operations, sampled outputs, and specific prompt types. With its accessible code and formalized security claims, FlashProve opens the door for deeper scrutiny and verification in AI deployments, suggesting potential next steps for regulators and model providers to establish trust in AI-generated content.