Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Anthropic open-sources FlashProve prototype for verifiable AI inference

AI summary

Anthropic open-sourced FlashProve, a research prototype that uses zero-knowledge proofs to let users check whether a response was produced by a model whose weights were previously committed. The proof does not reveal the weights, intermediate values or which experts a mixture-of-experts model activated. Anthropic says tests using Llama 4 Maverick produced proofs of about 19 MB for 1,024-token requests, while verification took about 17 seconds on a 24-thread CPU without a GPU or access to the weights. The company reported proof generation at about 110 tokens per second on one Nvidia H100 and at 70 to 95 times the GPU time of ordinary inference under its comparison conditions. The prototype supports only mixture-of-experts models and contexts of up to 1,024 tokens; side-channel attacks and formal equivalence between the Lean specification and executable implementation remain unresolved, according to the supplied reports.

In-depth insight

FlashProve offers a cryptographic way to audit model identity and output provenance without exposing model weights. The reports identify possible uses for safety evaluators, model providers and regulators, while also describing the project as intermediate research rather than a deployed verification system.

New Anthropic research: progress on provable inference for large language models.

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: https://www.anthropic.com/research/provable-inference

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: https://github.com/anthropics/FlashProve

We’re grateful to our collaborators: Alessandro Chiesa, @GiacomoFenzi, @gwberg95, @miranda__christ, Rahul Dalal (http://rdalal.github.io) and Ben Mayer (@transmissions11).

#Anthropic#FlashProve#Llama 4 Maverick#Meta#zkSNARK#Lean#NVIDIA H100#Apache 2.0

Original · 1 min readRead the original
Group chat

No comments yet. Start the conversation.