The math that secures blockchains is moving into AI. FlashProve uses a hash-based zkSNARK to prove which model wrote an answer. Still a research prototype: ~70-95x GPU overhead, 1,024-token limit. Early, but the right direction.
anthropic.com/research/prova…
x.com/AnthropicAI/st…
anthropic.com/research/prova…
x.com/AnthropicAI/st…
www.anthropic.com Progress on provable inference for large language models Sharing progress on a powerful type of cryptographic proof known as a zkSNARK, which is uniquely fit for provable inference. New Anthropic research: progress on provable inference for large language models.View quoted post on X
Here, we introduce FlashProve. Building on established cryptographic techniques, FlashProve allows developers to verify whether a given response came from a given AI model. It’s substantially more efficient than prior open-source methods for doing so.
These proofs have a range of applications. They can, for instance, make it harder for attackers to hijack a chat, or enable safety evaluators to verify that the model they’re testing is the same one being served to customers.
For our full write-up, see here: anthropic.com/research/prova…
Further progress in provable inference will likely be an important part of improving trust in AI deployment. To that end, we're open-sourcing FlashProve’s code, here: github.com/anthropics/Fla…
We’re grateful to our collaborators: Alessandro Chiesa, @GiacomoFenzi, @gwberg95, @miranda__christ, Rahul Dalal (rdalal.github.io) and Ben Mayer (@transmissions11).
