Progress on provable inference for large language models

Anthropic, EPFL, MIT and multiple researchers have made progress on a type of cryptographic proof called "zkSNARK" for large language models. Currently, when a user talks to an AI model, they have no way of knowing for certain that the responses came from the specific model claimed by developers. This is because most AI model weights are kept private.

Provable inference aims to enable a user with limited resources to verify that a response came from the requested model without running the model itself. The researchers released FlashProve, an open-source prototype for provable inference that can generate proofs at over 100 tokens per second for a 400B-parameter model using one H100 GPU. The release includes code, Lean formalization, and a technical description of the protocol.