However, the article also points out the high cost of implementing ZKPs in machine learning, suggesting an alternative approach called opML, where proofs are only created when results are challenged. This could significantly reduce the computational effort to verify every inference. The author concludes by expressing hope that the integration of blockchain and AI will lead to a future where the reliability and fair use of algorithms are substantiated through robust verification processes, fostering a deeper sense of trust and transparency for users.
Key takeaways:
- The AI space is notorious for being closed-source, which means there is no transparency over how the models work, how many parameters they truly have, or if they are even run when we send a query to them.
- Blockchain technology can be used to create a cryptographic proof of a model’s inference, providing transparency and trust in the AI models.
- Zero-knowledge proofs (ZKPs) are a cryptographic mechanism that can prove a specific machine-learning model has been executed without revealing the computation, serving as an attestation mechanism.
- Despite the potential of ZKPs in machine learning, the cost of doing zkML is currently incredibly high, but optimistic rollups, where proofs are only created when results are challenged, could be a cost-effective alternative.