Original author: Cathie, Hyper Oracle
Early summary
Latest articles by Vitalik ButerinExplores the intersection between blockchain and artificial intelligence (AI), focusing primarily on how AI can be applied to the crypto world, and explores four intersections: AI as actor, AI as interface, AI as rule, and AI as a target.
The article discusses the prospects and challenges at these intersections, highlighting issues of adversarial machine learning attacks and cryptographic overhead. The article mentions the possibility of using cryptographic forms such as zero-knowledge proofs to hide the inner workings of a model, while pointing out the challenges of cryptographic overhead and black box adversarial machine learning attacks.
Finally, the article discusses techniques for creating scalable decentralized private AI and considers applications to AI safety and AI as a gaming target. The article concludes by emphasizing the need for careful practice in these fields, but expressing expectations for the prospects of the intersection of blockchain and AI.
0. Crypto + AI Application Prospects and Challenges
In Vitaliks latest article, he discusses the intersection of artificial intelligence and cryptography and identifies two main challenges: cryptography overhead and black-box adversarial machine learning attacks.
Vitalik believes that artificial intelligence and cryptocurrency have great potential. Artificial intelligence can play a key role in helping cryptocurrencies become better, such as as a game interface or game rules.
1. Challenge: Cryptography Overhead
a) Has the cryptographic overhead problem been solved?
While Vitalik sees great potential in AI x Crypto, he noted that one of the main objections is the cryptographic overhead. Currently, the most mainstream on-chain AI/ML method is zkML, which compiles ML models into zk circuits so that cryptographic proofs can be verified on the chain.
“AI computing is inherently expensive,” and with cryptography, it becomes even slower.
Vitalik believes that the problem of cryptographic overhead has been partially solved:
Artificial intelligence calculations and their cryptographic overhead are suitable for high acceleration, and there is no unstructured type of calculation like zkEVM.
Over time, more efficient ZooKeeper cryptography schemes will be invented and the overhead will be greatly reduced.
b) Currently, the overhead is 1000x.
However, this approach is far from practical, especially for the use case Vitalik described. Here are some relevant examples:
The zkML framework EZKL takes approximately 80 minutes to generate a proof for a 1 M-nanoGPT model.
According to Modulus Labs, zkML is >>1000x more expensive than pure computation, with the latest reported figure being 1000x.
According to EZKLs test, the average proof time of RISC Zeros random forest classification is 173 seconds.
In practice, waiting several minutes to get a human-readable explanation of a transaction generated by an AI is unacceptable.
2. Solved by opML
a) opML: Optimistic machine learning
At the end of the article, Vitalik mentioned: I look forward to seeing more attempts at constructive use cases of artificial intelligence in all of these areas, so that we can see which of them are truly feasible for large-scale applications. We believe that zkML is in It is not feasible at this stage and the above applications cannot be realized.
As the inventors of opML and creators of the first open source implementation of opML, we believe that opML can solve the cryptographic overhead problem through game theory, making AI x Crypto possible now.
b) Security through incentives
opML solves the cryptographic overhead problem of on-chain ML while ensuring security. For simplicity, we can use Arbitrums AnyTrust hypothesis to evaluate the security of the opML system.
AnyTrust assumes that there is at least one honest node per claim, ensuring that the submitter or at least one validator is honest. Under AnyTrust, security and validity are maintained:
Security: An honest validator can enforce correct behavior by challenging incorrect results from malicious nodes, thus punishing them through an arbitration process.
Validity: The proposed outcome is either accepted within the maximum period or rejected.
Comparing AnyTrust and Majority Trust, opMLs AnyTrust model is more secure. AnyTrust maintains high security and is superior to Majority Trust under various conditions.
c) User privacy > Model privacy
Vitalik also addressed the issue of model privacy in the article. In fact, for most models (especially the small ones that zkML currently supports in practice), it is possible to reconstruct the model with enough inference.
For privacy in general and user privacy in particular, opML appears to lack inherent privacy features due to the need to keep the challenge public. By combining zkML and opML, we can achieve just the right level of privacy, ensuring safe and irreversible obfuscation.
d) Implement AI x Crypto use cases
opML can already run Stable Diffusion and LLaMA 2 directly on Ethereum. The four categories Vitalik mentioned (AI as player/interface/rules/objectives) can already be implemented with opML without any additional overhead.
We are actively exploring the following use cases and directions:
AIGC NFT (ERC-7007), 7007 Studio wins the Story Protocol Hackathon
On-chain artificial intelligence games (such as Dungeons and Dragons games)
Prediction markets using ML
Content authenticity (Deepfake verifier)
Compliant programmable privacy
Prompt market
Reputation/Credit Score
3. Summary
With opML, we can eliminate the challenges posed by cryptographic overhead, retain decentralization and verifiability, and make AI x Crypto viable today.
