Wyatt Benno
@wyatt_benno
Followers
2K
Following
5K
Media
45
Statuses
1K
Cooperative network for verifiable compute @novanet_zkp (DeSCI) I work where AI meets cryptography: @kinic_app . “What do you do for others? #runawayZK
Chicago / Tokyo
Joined May 2015
1/ We made verifiable AI (zkML) 3-7x faster than everyone else! Here's how. We asked a simple question. What if we modified a16z Crypto's sumcheck based & lookup-centric zkVM "JOLT" for ML operations?👇🧵
11
28
119
The Definitive Guide to ZKML (2025). https://t.co/IJkTY2vMou This is meant to be a reference for us all to check in 12/mo to see progress! Leaving edits open until Dec 12th -> DM for tweaks, corrections, or additions.
blog.icme.io
AI systems are making decisions that move money, affect health outcomes, and control autonomous agents. But how do you verify that an AI actually computed what it claimed to compute? That it used the...
1
0
11
I am doing an in-depth post on the zkML in 2025. Let me know if your team is doing something within the space and I will plot it in.
0
0
0
At its core, this is what zkML aims to address: AI that doesn’t lie to you. @icme_app is on it.
David Sacks says the biggest risk of AI was described not by James Cameron in The Terminator but by George Orwell in 1984. “I almost feel like the term ‘woke AI’ is insufficient to explain what’s going on because it somehow trivializes it.” “What we’re really talking about is
0
1
3
Not all zkML proofs are about proving inference of giant LLMs! Classification proofs like this — verifying agent spending decisions against policy constraints — are lighter, faster, and actually deployable right now. zkML usecase have actually been underlooked.
Keep up the pace! Spending policy proofs for agents transacting in USDC
0
0
1
The more verifiable inference we enable with zero knowledge proofs — the more usecase we open up 🤝
@NovaNet_zkp using Circle Gateway to enable agents to spend USDC seamlessly across supported chains.
0
1
2
FHE is getting over-thrown around as an alternative to ZK. If you're not doing verifiable compute on encrypted values, the server can just return gibberish! Every crypto usecase of FHE needs verifiable compute under the hood (ZK, TEE, or consensus).
FHE is getting over-thrown around, to suggest its an alternative to MPC. ~Every crypto usecase of FHE is in practice actually MPC under the hood. If you need the data provable to a third party, its being encrypted to an MPC committee.
0
0
2
Silo: AI memory owned by Big Tech Permission slips.: AI memory owned by Web2 startups Sudo: AI memory owned by You. Cryptographically. The design space explodes when you don't need permission to build & you don't need to worry about middlemen tampering. Everyone wants to be
0
0
0
x402 agents are about to power $10T in autonomous transactions by 2028....🔥 But there is a problem for cross border multi-lingual agents.. Let's say your agent operates in Japan, consulting an English knowledge base, autonomously paying for services with USDC. You can't audit
Multilingual AI memory sounds great until you realize: if the sources are in English but you speak another language.. - You can't audit because you don't speak English. - How do you verify the provider didn't make shit up? - Did someone swap in malicious sources? With embedding
0
2
2
Multilingual AI memory sounds great until you realize: if the sources are in English but you speak another language.. - You can't audit because you don't speak English. - How do you verify the provider didn't make shit up? - Did someone swap in malicious sources? With embedding
0
1
2
Having fun adding some complete gems to this AI memory. https://t.co/KlossS63Bw
It happened again... At a halloween event over the weekend I met a dad. He asked me: 'What should I do in Tokyo?' After a decade in Japan, I get this constantly. Before: I would spend 20 mins of copy-pasting links, restaurant names, fav spots... Now: I just send them my
0
1
3
It happened again... At a halloween event over the weekend I met a dad. He asked me: 'What should I do in Tokyo?' After a decade in Japan, I get this constantly. Before: I would spend 20 mins of copy-pasting links, restaurant names, fav spots... Now: I just send them my
kinic.io
Stop re-explaining yourself to every AI tool. Kinic creates a portable memory layer that works across ChatGPT, Claude, custom agents, and more. Cryptographically secured, user-owned — your context...
1
4
15
Finishing up a visit in NYC this week.. the vibe of the builders and events here has been amazing. I am going to close out the last mtg with a 'rousing pre-battle calls to arms' on trustless agents 😅. HEARKEN! HEARKEN UNTO ME! x402. erc-8004. zkML. TEE. JOLT. Accelerate! 🏇⚔️
1
0
6