Proof of Track 2025-03-09: Truly Onchain AI
// proof of track is back! and i’m thinking about enabling it to be daily and more lightweight!
Truly Onchain AI (not verifiable ones)
Onchain Verifiable == Onchain
It is always hard to put all computation onchain, because execution environments like EVM are with limited compute resources.
To enable all computation to be “onchain”, people adopted offchain computation to be verified onchain, and then rebranded “onchain verifiable” to “onchain”. For examples, layer 2s with zkEVM or optimistic fault proof vm.
But for these verifiable systems, there’s still gap between the onchain verification logic and offchain component + bugs introduced with such complex architecture.
So natively onchain is still superior than verifiable onchain. Just like btc > wbtc in bridging terms.
Same for Onchain AI
Running AI models (the ones that you think of) is hard enough, so running them onchain inside EVM is much much much harder.
At ORA, we took the same approach to enable AI model (inference) onchain, with Optimistic Machine Learning (https://arxiv.org/abs/2401.17555).
Also, other verifiable onchain AI approaches includes zkML and crypto-economic based AI (like restaking in eigenlayer).
You can check out Haseeb’s great explainer about all these: https://x.com/hosseeb/status/1773146428594090473.
Though they are “fixed” with verifiable AI, running AI directly inside smart contract / completely onchain is still a cool challenge.
Some Truly Onchain AIs
The MNIST digit classifier runs entirely within a single smart contract, just EVM inference with 94% accuracy and 41 million gas per inference.
Small model but it works, and the deployment fee is afford. This is why I wrote this post today.
Kind of like a language AI model. It is a contract you can run to get semi-coherent phrases, all entirely running on eth mainnet.
These two are the ones on my mind. Should be some more.
Wondering if parallelized evm will help something like this.
