AI is getting better at generating answers, but how do we know those answers actually came from the model we intended to use? I think this is a much bigger challenge than it might seem. That's why Anthropic's FlashProve caught my attention. It introduces a way to verify that a response was generated by a specific AI model without exposing the model's weights. What makes this even more interesting is the cryptography behind it. FlashProve uses hash-based STARK techniques to build its proofs, bringing post-quantum security properties into the verification process. There are a few details I find particularly important: Cryptographic techniques already used in systems like Starknet. Hash-based proofs that don't rely on elliptic-curve assumptions. Formal verification in Lean, helping establish confidence in the protocol's correctness. Earlier foundational work involving Cairo and Stwo. Another interesting connection is Alessandro Chiesa, StarkWare co-founder, being among the researchers behind this work. For me, this shows how technology developed for verifiable computation can reach beyond blockchain. As AI becomes part of more important decisions, trusting what a model says may no longer be enough. We also need ways to verify that the model actually produced the output. That's where I think the combination of AI and cryptographic proofs becomes really valuable. @Starknet underlying cryptographic technology is now relevant to a much broader conversation than blockchain alone.