Abstract
TensorCash is a protocol for making AI inference efficiently verifiable. It lets an inference provider attach a proof-of-inference transcript to any AI model run. Users, aggregators and validators can check that a model was executed under declared conditions without trusting the provider. The same proof-of-inference primitive can then be used as proof of work in the sense introduced by Nakamoto [2008]: TensorCash introduces a Bitcoin-derived blockchain and native token whose block production is secured by useful inference. AI providers can secure a cryptocurrency while serving commercially valuable tasks, converting inference into decentralised economic value.
Contents
- Introduction: The Parable of Potato-Land
- Back on Earth
- TensorCash Protocol
- TensorCash Building Blocks
- Related Work
- Proof-of-Inference and Proof-of-Work
- Conclusions
- Appendix A. Inference Fuzziness and Reproducibility
- Appendix B. Entropy Collapse and Collision Resistance
- Appendix C. Attack Vectors Analysis
- Appendix D. Privacy
- Appendix E. Inference Example
The full paper — figures, equations and appendices — is in the PDF above. Authored pseudonymously by Imosuke Takakuni.