As predicted Ai needs cryptography for verification. The need in a multi-tennant environment with hosted clouds & model routers, that need increases. People need to verify they're getting the model they're paying for & to prove to regulators what model/weighs theyre using as these models become the backbone for securoty & enterprise workflows in regualted insdustries. ZKP was always in the card as the complimentary solution for AI when the service demands third party trust in the AI blackbox. but note that ANthropic's need to prove inference differs from crypto's use of ZKP. crypto-native projects optimized for cheap on-chain verification, so they wanted small proofs and accepted slow provers. Anthropic optimized for off-chain auditors and regulators, so it accepted 19 MB proofs and got a fast prover. Anthropic chose STARKs that can generate proofs fast - A batch of 24 requests proves in about 223 s, which works out to about 110 tokens/s. Thats roughly 72x the GPU time of serving the same requests on vLLM. That led them to adopt STARK proofs that are both quantum secure & GPU friendly. But & there's a huge but - what Anthropic is really proving to third parties is just that the model weights in the hash commitment, were what was run as an integer in response to a single prompt. There's not enough to prove what exact model was actually run, just that the weighs were what was committed. That implies there are pretty intuitive ways to game this - commit to a cheaper model with similar weighs, rewriting promptes selectively cheat on the rest of the traffic that's not being samples or proven, randomness selection can steer tokens within the error rate of the proof https://t.co/VHR8DG0vbx