Wyatt Benno
@wyatt_benno
Making AI output succinctly verifiable, using formal methods and cryptography. Serial Technical Founder | @icme_labs What do you do for others?
296 Following 1.9K Followers
Three things you should learn today.
1. Lattice are here to stay (security + performance is amazing)
2. Lattice papers are often long 😅
3. ZK (privacy w/ Lattice); Akita will have it soon 🤞 and we already know its possible with Lattice folding 👀.
Show more
Amazing results.. 10x is insane!!!
Fast CUDA ZKML inference inbound (40/tokens a second or more) Which would make it comparable to SoTa plus a few other amazing benefits only homomorphism can provide.
Show more
4/ On a MacBook, CPU-only Lattice Jolt proves well over 2 million 64-bit RISC-V cycles/sec (curve-based Jolt was ~1 million).
Our new Apple Metal GPU backend takes that to over 10 million — a ~10x jump in a single release, with plenty more to squeeze.
Show more
Terence Tao just dropped a banger 🤯
Super cool!
Proud to announce ProveKit v1! Devs can build authentication with ZK privacy, running on all user devices.
2 years ago I started this project to show that every users’ phone is capable of ZK proving a passport document, under a minute, no trusted setup, 128bit post quantum secure. Thanks to collaborations with many world class teams, and several breakthroughs, this is now reality! Thank you everyone involved 🙏
ProveKit uses Noir as it's programming language, so try your circuits today!
Show more
True or false?
Any specific task a large generalists AI agent can do, a specialist can do just as well if not better.
More like "human-in-the-loop" explained.
"A person who can acquire no property, can have no other interest but to eat as much, and to labour as little as possible." - Adam Smith
Techno-ownership === web3.
incredible take
What the end of 2026 looks like for zkML.
- GPUs are getting more powerful.
- ZKP algorithms are getting better (10,000x vs 1,000,00x).
- Model inference itself is becoming more efficient.
The convergence of these 3 facts enables you to build much more w/ Jolt Atlas (v2 paper inbound). If the cost gets low enough & the speed fast enough, universally people will use ZKP over TEE (disruption!!) for many adversarial & on-device use cases. Enterprise today already choose it for verifiable AI agent guardrails w/ zero trust; I expect these trends to continue. Any good reason why they would not??
Show more
In March I laid out performance stages for zkVMs. By the end of 2026, we should hit “Stage 2”: 10,000x prover overhead relative to native execution. And that's when things get interesting for applications beyond blockchains.
Show more
“we sandboxed the agent”
meanwhile the agent:
I think it’s amazing that in the business & IT world, we hear “AI”, “AI!” & “AI!!” multiple times per day.. but yet 80% of US adults almost never use it. If you stand in any random room with 10 people.. 8 will be unfamiliar with what it’s used for at all!
CRaZy.
Show more
Lean4 is a performance tuned kernal c++.
Probably lots of soundness bugs still to be found.
Lean4Lean is written in Lean (proof of correctness for parts) but is 20% - 50% slower.
In adversarial settings this matters.
Show more
Shoutouts to Ramana Kumar for refuting the Collatz Conjecture in Lean, *as checked by Comparator!*
Imagine when you can prove AI guardrails to 3rd parties.. in under 1s per run (locally). Moreover, you can fold thousands of these and anyone can verify them with zero trust, once in under 100ms. Fantasy or this year!?
For contrast, today we are told to store all AI guardrails logs ($ millions); just in case. And proving they ran and worked takes weeks of review (I see you CFOs) and exposing both logs and your policies to auditors which contain highly private info.
Show more
@badcryptobitch @StoffelMPC I see, thanks.
Latest generation of zk is far more performant than your expectations. Jolt for instance proves 1 million RISC-V cycles using only 2-3GB of RAM (soon will go down to <1GB).
In general, zk is converging to ~3 orders of magnitude overhead over native computation
Show more
What if novel AI math breakthroughs become unverifiable, due to the lack of people willing or able to check them? Probably, going to happen.
saw an ad.. working on the fear that "the new guy can use AI and gets paid 40% more..", but the protagonist who has been working for years, in spite of knowing more is on the chopping block.. AI is just that great!
WHY in the WORLD would the protagonist not just use the same AI to skill up.. ? Why would they pay a service extra to use a free LLM tool? The Ad seems to be shooting itself in the foot. Let me know if you see similar.. 'fear slop'.
Show more
Great piece. The thesis is right, formal verification is finally practical thanks to AI!
One gap: Lean requires proof experts, relies on interactive proving, and produces non-succinct proofs. SMT-based verification (Dafny, ICME PreFlight, etc) automates the proof step entirely. More importantly, you can translate natural language intent directly into formal specs via automated reasoning. No tactics, no proof engineering. Some systems hit 99% and climbing with minimal human battle testing.
We call this vericoding. Same goal, different tooling. And you can wrap the entire pipeline in ZK so every verification result is succinctly verifiable.
Wrote about it here:
Show more
Many people have claimed that with AI-assisted bug finding, secure code (and hence trustless anything) will be impossible.
I have a much more optimistic take, and AI-assisted formal verification is a major part of the reason why:
Show more
Now let’s make this process succinctly verifiable and you get 🥁 vericoding.
NL -> specs -> review -> formal proofs with code. If you use smt and tools like Dafney you can wrap solvers in ZK. If you use Jolt Atlas (zkML) you can wrap conversion models in ZK; fold them all together.
It took you 20h to do this with your agents.. it should take me 1s to verify it 😜
Show more
We can now fully rewrite most software in @leanprover and prove it correct:
- Compiler module rewrite (AI) from Rust to Lean
- Full FFI integration
- All unit and integration tests pass
- Formal spec and proofs!!
- Under 20h wall time (unnoticed pauses)
Show more
Now let’s make this process succinctly verifiable and you get 🥁 vericoding.
NL -> specs -> review -> formal proofs with code. If you use smt and tools like Dafney you can wrap solvers in ZK. If you use Jolt Atlas (zkML) you can wrap conversion models in ZK; fold them all together.
It took you 20h to do this with your agents.. it should take me 1s to verify it 😜
Show more
We can now fully rewrite most software in @leanprover and prove it correct:
- Compiler module rewrite (AI) from Rust to Lean
- Full FFI integration
- All unit and integration tests pass
- Formal spec and proofs!!
- Under 20h wall time (unnoticed pauses)
Show more
Multiple “web3 AI” people I met this month.
Me: how do you secure that?
People: TEE.
Me: how do you get those proofs to fit on chain? Wrap them in ZK?
People: I am not sure… need to ask the team. Uh, i don’t think so… we use TEE…
Show more
Prediction: in 2027 almost all of agentic commerce will be secured by ZKP; And not for the reasons you expect.
ZKP adoption is accelerating in RL 📈
