Checking an AI answer is never free, it just charges you in a different currency each way. a) Heavy cryptographic proof (zkML) charges compute. b) Trusted-hardware approaches (TEEs) charge a dependency on the chip maker. c) Consensus-scoring charges accuracy, because it cannot actually prove the model ran. Ours charges a small overhead per response, a company claim we have not had independently benchmarked, running on testnet today. Every one of these is a real cost, and the only wrong move is pretending your option is free.

Excited to share that @icme_labs has been selected for the @0G_labs Apollo Accelerator, backed by @theBBFund and veterans from Stanford's blockchain community @StanfordSBA. 🔥Only 10 teams selected out of hundreds🔥 #ZKML# We're building PreFlight: cryptographic guardrails for AI agents. Plain English policies compiled to formal logic used for enforcement; with a succinct ZK proof of every decision. LLM judges don't work. Humans-in-the-loop, no way! Agents that handle anything important need cryptographic guardrails. That's PreFlight. #0gApolloAccelerator#

Prediction: You won’t lose your job to AI, until your company starts using PreFlight (Automated Reasoning + ZKML). Here is why. Imagine the company fires the HR person and replaces them with an agent. Over time that agent could, and would likely start hallucinating. How would management know that its decision were based on facts (company HR policy)? “AI is a black box!” With AR, they could turn that policy into formal logic and make sure the agents outputs correspond to facts, precisely. No more guessing if it hallucinated. But who wants to go in and check a thousand AI outputs against formal logic proofs - to audit all of this? “No one.” PreFlight wraps these formal proofs into one succinct proof. With ZK it can be verified in under 1s. “All of our policy rules have always been followed.. by the new HR agent.. shi..” Personally, I am unsure if we can undo this invention. The moment you had agents that are guaranteed to follow facts and succinctly verifiable to have done so, is the moment you have reliable AI coworkers..

Great seeing @zk_albi from the @icme_labs team leading two talks. One as a guest speaker on new cutting edge efficient recursion for JOLT. The other his own work on ABBA (lattices). These talks are on YouTube: search “zkSummit Rome”. In our Jolt Atlas (zkML JOLT) repo - upstream work from the a16 helps us greatly 🙏 this year we have ZK (privacy) coming from folding scheme (cross stream), efficient recursion, and a Jolt prover that can work on constrained devices. Down stream: a lot of the core focus has been on EC primitives. A lot of the work needs porting to lattice. One example: add privacy (ZK) in the lattice setting. Another is succinct proofs. Getting one GPU proving one cpu core in realtime.. is likely happening these year.. Let’s go! 🔥🔥🔥

Listened to the most recent @zeroknowledgefm "Is ZK dead? Or has it just begun?" Deep research takes time. The funders were early to the vintage. Now that vintage is becoming real world products. Will make @tarunchitra bullish on ZK again 😆 and ZKML. 'Automated reasoning' turns natural language into formal logic and can be used as AI guardrails that can't be bypassed or ignored. Will CC when our paper drops mid-month🤞.