zkML
Also: Zero-Knowledge Machine Learning
Producing a cryptographic proof that a specific model produced a specific output from a specific input, verifiable without re-running the model or seeing its weights.
An inference result on its own is an assertion. zkML turns it into something checkable: the prover generates a proof alongside the output, and anyone can verify that the claimed model really did produce it — without access to the weights, and without repeating the computation.
Why it is the joining piece
A model whose output can be proved is a model that can be paid for by another machine. Verification is what allows an automated counterparty to accept a result it did not compute, which is the missing condition for machine-to-machine markets in inference.
What makes it hard
Proving a computation costs far more than performing it, and neural network inference is a very large computation. Practical work concentrates on reducing what has to go inside the proven circuit — for instance keeping hash functions out of it, since hashing inside a circuit is expensive and is the recurring cost in recursive proof systems.