2026-08-27
minor edit 2026-08-28
What is the bottleneck for ZK proofs for ethereum smart contracts for frontier AI models?
Disclaimer
- Quick Note
- Not an exhaustive list
- I support a complete ban on frontier AI research. If the ban ends up being restrictive enough to also prevent the app below from being built, I am probably fine with that.
Background context
- I am interested in this question because I want there to exist trustless (or trust-minimised) censorship-resistant oracles. My use case for this is to pay people to leak information about an organisation. But there exist many other use cases.
- Oracles require a data source and some computation over that data source. The computation part can often be done by AI. If you had an ethereum smart contract verify ZK proof of LLM inference, that entire step becomes trustless and you don't even need to call it an "oracle" anymore.
- Example: Have a PoS DAO on ethereum publish merkle root of the latest monthly wikipedia archive. This is the data source. Have the PoS DAO on ethereum also publish SHA256 hash of the latest AI model. A user can now take any page from this archive, run any inference over this, and trustlessly prove this to a smart contract. You can use this to bet on or even incentivise any event that some wikipedia page talks about.
ZK proofs for ethereum smart contracts for frontier AI models
Pieces required
- ZK stuff
- Write cryptography papers defining a ZK proof system for a given instruction set architecture
- Convert cryptography papers into code
- Get audits and real-time test-in-prod for the code
- ML and optimisation stuff
- Convert ML model in this instruction set architecture, so you can do ZK proofs over it.
- Optimise the ZK proof system so it is fast
- Ethereum stuff
- Write ethereum smart smart contract that verifies ZK proofs in this proof system
- Get audits and real-time test-in-prod for the smart contracts
Many of the pieces seem to be coming together individually, but as of 2026-08, no single company has solved all the pieces if I understand correctly.
Some pieces that exist as of 2026-08
- EZKL has working audited code for ZK proofs of frontier AI models. But they don't have this deployed on ethereum.
- EZKL has working audited code for ZK proofs of stuff that is deployed on ethereum. But they don't have frontier AI models.
- RISC Zero has working ethereum smart contracts for ZK proofs over RISC architecture. They haven't deployed a demo for frontier AI models yet.
- SP1 has audited ethereum smart contracts for ZK proofs over RISC architecture. They haven't deployed a demo for frontier AI models yet.
- DeepProve has worked on optimising speed for ZK proofs of LLMs. They haven't deployed this for ethereum.
- Attestable also claims to have worked on optimising speed for ZK proofs of LLMs. Closed source though.
- [Link to Attestable front page]
- Same for jolt labs
- See also: ERC 7992
I am still confused if there exists any bottleneck.
- AFAIK people are nowadays moving to GCR / sumcheck based ZK proof systems to generate proofs faster for ML models. (idk what that this, haven't studied yet)
- For my use case, proof generation speed is not really a bottleneck though. The oracle will run optimistically anyway, you only need to generate proof in case of a dispute. And you'll be willing to spend atleast $1000 worth of compute and one day worth of time, to generate a proof for like 1000 tokens or something. For reference, Attestable claims 85 tokens/second for 30B model on H100.
Subscribe
Enter email or phone number to subscribe. You will receive atmost one update per month.
Alternatively, you can subscribe via RSS, or a third-party service like blogtrottr
Comment
Enter comment