Huge congrats to Succinct’s Rahul Dalal, one of the cryptographers behind FlashProve, @AnthropicAI’s provable inference system built on zkSNARKs. Why: Today, AI runs on a “just trust me” model. Users can’t verify that a response came from the model they chose, or that an attacker didn’t tamper with it along the way. As AI takes on higher-stakes work, that create tension for users, labs, and regulators alike. What: Researchers from Anthropic, EPFL, MIT, and @SuccinctLabs built FlashProve, a system that lets a model provider commit to a specific model and then prove each response was produced by it, without revealing the weights. Under the hood, it uses zkSNARKs, the same cryptographic technology Succinct uses to secure $13.5B in deposits on Ethereum L2s. So What: FlashProve proved Llama 4 Maverick, a 400B-parameter LLM, at 110 tokens/sec on a single H100 GPU, with ~19 MB proofs. That puts proving at roughly 70 to 95x the cost of running the model. The previous standard for the same model was ~0.02 tokens/sec. What’s Next: FlashProve is still a research prototype, but it’s a step change toward the question we asked with Flock earlier this year: how close can proving get to computing? The closer it gets, the less reason there is to release any output without a proof attached. Congrats again to Rahul and the FlashProve team, and great to see @Starknet, @Mantle_Official (proven by Succinct), and several zkML projects cited.