Something about OpenGradient's verification architecture that didn't fully register until I sat with it: $OPG #OPG @OpenGradient positions itself around verifiable AI inference — the idea that you can prove a model ran correctly without trusting the operator. That framing leans on ZKML, which is real, but it's also computationally prohibitive for almost any model running at practical scale today. What the protocol actually defaults to is TEE attestation — trusted execution environments backed by Intel and AMD hardware. Those aren't the same thing. ZKML gives you a mathematical proof; TEE gives you a hardware manufacturer's attestation that the execution environment wasn't tampered with. One requires no trust assumption beyond the math, the other quietly relocates the trust from the operator to the silicon supply chain. I don't think this invalidates the project — TEE is a meaningful improvement over unchecked inference — but there's a gap between "cryptographically verified AI" and "attested execution on trusted hardware" that the narrative tends to compress. The interesting question is whether ZKML ever becomes compute-feasible at scale, or whether TEE just quietly becomes the permanent default.