Frontier AI only feels scary because most of it still runs like a black box with a nice logo on top. What @inference_labs is doing with DSperse + JSTprove is basically saying: “ok, prove it.” I went back to the DSperse paperwhere they describe it as strategic cryptographic verification of high-value subcomputations instead of full-model SNARK cosplay. Slices of the model become independently provable units, so you can lock down safety gates, anomaly detectors, or private heads without dragging the whole network into ZK hell. On the other side, JSTprove plus the open-source repo in their GitHub org gives you an end-to-end pipeline on Polyhedra’s Expander backend: feed an ONNX model in, CLI spits out proofs and auditable artifacts, no deep crypto background required. ❯ DSperse decides what is worth proving ❯ JSTprove handles how it’s proved and exposed Most “frontier AI” stacks still ask you to trust benchmarks, not behavior. This combo flips it: you choose the failure modes you refuse to gamble on, and you wire proofs exactly there. If you’re building serious systems in robotics, finance, or agent land, the next honest step is simple: map your threat model, then ask which parts you’d actually route through DSperse + JSTprove. If the answer is “none,” you’re not really chasing trust yet #zkML