[Anthropic Brings Zero-Knowledge Proofs from the Crypto World to AI Inference to Prevent Models from Being Secretly "Switched"] Beating AI News reports that Anthropic has open-sourced FlashProve, bringing the zkSNARK (zero-knowledge proof) technology commonly used in the cryptocurrency field to AI model inference. This technology could originally prove the validity of blockchain transactions without disclosing transaction details. Anthropic aims to use it to solve another problem: how can users who pay a high price for an AI model's API know if the service provider has secretly switched it to a cheaper model? FlashProve allows service providers to first generate cryptographic fingerprints for the model weights and then provide mathematical proofs for the inference results. Users can check whether a specified output comes from the pre-committed set of weights without needing to obtain the model weights or rerun the model. Claude also participated in the development and optimization of this system. Anthropic has already completed experiments on Meta's 400 billion parameter model, Llama 4 Maverick. Processing the context of 1024 tokens generates a proof of approximately 19 MB, requiring about 17 seconds for verification using a standard CPU. However, currently each proof can only verify one output token, and cannot yet verify the entire response. The GPU computation time required to generate the proof is also approximately 70–95 times that of ordinary inference. Whether it can be used to identify API service providers that secretly change models in the future depends on when the proof cost can be reduced. [Original Link]