
We've been acquired by @tfh_technology to apply research @worldnetwork ๐
SF
Joined July 2022
- Tweets175
- Following2
- Followers14.1K
- Likes739
Modulus Labs retweeted
Been looking at AI x Crypto project all day - and have to admit, whoever wrote this on @ModulusLabs's landing page deserves a raise and some more.
Couple bars above their peers.
Modulus Labs retweeted
๐ช
We are taking "Proof Day" to SF on April 17 with @ModulusLabs & @nascentxyz & @EdgeAGI!
Dive into an evening of insights with speakers @shumochu, @dysfunctorial, @Zac_Aztec, @pumatheuma, @realDanielShorr, @nayr_oac, Junda and others. Plus, enjoy a complimentary dinner. ๐ฝ๏ธ
๐ช
In the meantime, check out our blog for what it took to prove a real LLM, why specialization is the future, & the endgame for AI x crypto!
Medium blog: medium.com/@ModulusLabs/chapโฆ
Eth block: etherscan.io/tx/0x7a629cf5bbโฆ
Follow @ModulusLabs for the latest in ZKML, AI x Crypto, & more ๐ค
๐๐ญ๐ก๐๐ซ๐๐ฎ๐ฆ ๐ก๐๐ฌ ๐ฌ๐ฉ๐จ๐ค๐๐ง ๐ข๐ญ๐ฌ ๐๐ข๐ซ๐ฌ๐ญ ๐ฐ๐จ๐ซ๐!
That's right โ we have completed the world's 1st ZKP of a 1B+ param LLM inference, before verifying the proof on Ethereum mainnet (block 19427725) ๐คฏ
What did Ethereum say? Learn more in ๐งต
medium.com/@ModulusLabs/chapโฆ
So it's possible โ nice! But the proving overhead is far from acceptable...for now ๐
As we demonstrate in our latest paper, specialization offer 1,000-10,000x speed-ups for ZKML!
In fact, we already have some exciting new updates cooking at the bleeding edge ๐๏ธ
We're excited to unveil ๐
๐๐๐๐๐๐๐๐โthe world's most powerful ZKML prover
Our first implementation achieved a mere 180x proof generation overhead vs. the same (non-verifiable) computation on an M2 Mac.
Highlights below, read all about it here: medium.com/@ModulusLabs/chapโฆ ๐
This was no easy task ๐ฅต
We had to split GPT2-XL's 48 decoder blocks into 144 sub-blocks, before proving on our massive 1TB memory, 128 core CPU cloud clusters for nearly 90 hours!
That means that โ using generic prover systems โ zkLLM overhead is over 1,000,000x!!!
GPT2-XL is a 1.5B parameter behemoth of an AI model!
And to see whether ZKML has finally advanced enough to make proving LLMs feasible at all, we decided to put our proving chops to the test and find out!
Oh! It said: "magic", btw <--> openai.com/research/gpt-2-1-โฆ
๐๐ญ๐ก๐๐ซ๐๐ฎ๐ฆ ๐ก๐๐ฌ ๐ฌ๐ฉ๐จ๐ค๐๐ง ๐ข๐ญ๐ฌ ๐๐ข๐ซ๐ฌ๐ญ ๐ฐ๐จ๐ซ๐!
That's right โ we have completed the world's 1st ZKP of a 1B+ param LLM inference, before verifying the proof on Ethereum mainnet (block 19427725) ๐คฏ
What did Ethereum say? Learn more in ๐งต
medium.com/@ModulusLabs/chapโฆ
Modulus Labs retweeted
For a while now, Iโve said that AI is the EVERYTHING narrative.
These days, AI headlines and tokens are dominating our timelines.
With so much happening, it feels like we're living in an accelerating time chamber. How do we make sense of it all?
Introducing Generalized.
A collaboration between @aiarena_ and @ModulusLabs hosted by yours truly alongside @sihing_guppy and @realDanielShorr.
This new weekly X Spaces series features engaging discussions with innovators at the forefront of the AI x Web3 ecosystem. Whether you're new to the scene or a seasoned pro, youโll want to tune in to hear the latest insights, debates, and conversations.
While many shows are focused on NFTs and gaming, the intersection of AI x Web3 remains largely unexplored. Generalized aims to fill this gap.
Our mission is to simplify AI and help you become more informed while navigating this dynamic ecosystem. We aim to inspire you to explore and embrace the coming changes, and we promise to do this in a way that's informative but not dense, educational but not pretentious.
I want to also give a special shout out and thank you to @SamSteffanina for inspiring us with his incredibly successful Gamified series, and for giving me a peek into his creative process. The work that he and his team put into every show cannot be understated, and it will be a benchmark that we set for Generalized.
Join us for the first show next Tuesday, Mar 12th at 12PM ET, 9 am PT.
It's time to level up.
Modulus Labs retweeted

cami@camiinthisthang28 Feb 2024
Colombia
ColombiaConnected via United States App StoreAccount-level information, not a live location or per-post device.
YOURE REALLY GONNA BET AGAINST THIS TEAM? @ModulusLabs @ende_sh
Replying to @weikeng
Not to spoil the next paper but... it works even better for deep neural nets ๐
Excited for the future of specialized ZK? Eager to get your hands dirty?
We're launching the Modulus Early Access Program (MEAP1) later this year for builders just like you, eager to explore ๐
๐๐๐๐๐๐๐๐
Sign up for your chance here: itrrvgq8qar.typeform.com/to/โฆ!! ๐๏ธ๐
A heartfelt thanks to the brilliant minds and advisors who've guided us. Special shoutouts to @zhenfei_zhang, Riad Wahby, @_weidai, @SuccinctJT, and @VitalikButerin for their invaluable insights and support ๐๐ง
Want to learn more? Dive into all the details with our paper, โScaling Intelligence: Verifiable Decision Forest Inference with ๐
๐๐๐๐๐๐๐๐โ
Pretty graphs and so much more await you ๐๐ at github.com/Modulus-Labs/Papeโฆ
Our third key insight is generalizing the linear-time prover algorithm from Libra [XZZ+19] to the dataparallel case, allowing ๐
๐๐๐๐๐๐๐๐ to achieve a strictly linear proving time for even dataparallel, unstructured layers requiring wiring predicates โ-๐ฆ
Our second key insight is a novel layered claim combination strategy, built over the interpolation protocol in Time-Optimal Interactive Proofs for Circuit Evaluation [Tha13] and augmented with a claim grouping heuristic, shaving proof generation time by ~40-60% โโฌ๏ธ
So how did we achieve such results from that insight?
A small armada of techniques! Chief among which includes refining and expressing the entirety of the verifiable decision forest circuit from the Zero-Knowledge Decision Tree paper [ZSZF20] in a structured form-factor ๐ฒ๐ฒ๐ฒ
A close look showed that one prover โ the GKR one used in zkCNN [LXZ2021] โ stood out among all others
An even closer look revealed why: as weโve described extensively, GKR circuits are layered, and admit a blazingly fast prover over linear and structured operations ๐๏ธ
13 months ago, we asked: "What's the real cost of verifiable machine learning inference?"
Six proof systems later, we got our answer: itโs astoundingly expensive at ~10^5 to 10^6 overhead over naive inference!
But the ๐ (or in this case, ๐) is in the details...
We are so thrilled to finally be releasing: โThe Cost of Intelligence: Proving Machine Learning Inference with Zero-Knowledgeโ๐ธ๐ค๐ธ
A quick๐งตon our survey paper, supported via EF grant and a cohort of heroic cryptographers and builders
(Medium blog: medium.com/@ModulusLabs/chapโฆ)
We're excited to unveil ๐
๐๐๐๐๐๐๐๐โthe world's most powerful ZKML prover
Our first implementation achieved a mere 180x proof generation overhead vs. the same (non-verifiable) computation on an M2 Mac.
Highlights below, read all about it here: medium.com/@ModulusLabs/chapโฆ ๐













