Bitcoin's proof-of-work mechanism secures permissionless blockchains but at a steep cost: miners burn enormous amounts of electricity solving hash puzzles that serve no purpose other than extending the chain. A new arXiv preprint proposes an alternative that repurposes that computation. Instead of hashes, miners would perform machine learning inference, with the resulting model outputs or predictions serving as the proof of work.
Under this scheme, the computational effort required to secure the network would simultaneously produce something of value—trained inference results or "knowledge"—rather than a meaningless nonce. The abstract frames this as a shift from mining hashes to mining knowledge, though it does not provide details on how the inference tasks are validated, how difficulty is adjusted, or how the system resists game-theoretic attacks.
Because the source is a single announcement abstract, there is no independent evaluation or comparison with existing alternatives. The idea is provocative, but the practical feasibility remains an open question.