Pikagi
Eito Miyamura | ♨️ GPU-EVM Profile Banner
Eito Miyamura | ♨️ GPU-EVM Profile
Eito Miyamura | ♨️ GPU-EVM

@Eito_Miyamura

1,136
Followers
415
Following
74
Media
390
Statuses

Building GPU-EVM: 100x EVM execution ♨️ Co-Founder @Gatling_X | Hiring DevOps,GPU Eng: CS Alum @UniofOxford , AI, RL

Joined December 2021
Don't wanna be here? Send us removal request.
Pinned Tweet
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
30 days
♨️ Announcing: GPU-EVM by @gatling_x ♨️ The most performant parallel-EVM in the world by 100x. Massive unlock: foundational infra for training AI/RL models that interact with the EVM, accelerated L2s, MEV, backtesting, and more 🧵+ waitlist below
Tweet media one
26
74
297
Last Seen Profiles
Memorandum

@SKHmemorandum

Senan

@wait4lifestan

Kadir yüksel

@Kadiryk65270903

Samuel Ivaska

@samekeyboards

puffi

@pinkpoffinz

Du

@Du31758789

Trevor “TD” DeLacey

@TrevorDeLacey

温泉まんじゅ

@wizwiztarou

عبدالعزيز النغيمشي

@abohamoud_

Michael

@mic_mazz

Eric S. York

@impactyork

Bayram ERGİN

@bayramergin50

Tennessee Baseball Coaches Association

@TBCAorg

The Bible For Normal People's Opinions

@theB4NP

あやこ

@cwo60xCv0U92387

Louise Arseneault

@L_Arseneault

Kayla Cardillo

@kaylacardillo

Real OT Solutions

@RealOTSolutions

Delilah Leitz

@DeliLeitz

Bria Peacock, MD

@bpeacockMD

Run4Coffee

@run4coffee

$$$$$…

@kennfentyy

Autocamper

@AutocamperBand

Russia in Sudan

@Rus_Emb_Sudan

Miss Josephina❤️🎊

@Missjosephina78

dubuk

@cowboy_borscht

Medusa

@medusa200614

Clayton Sandell

@Clayton_Sandell

Erika Heidi

@erikaheidi

Seblu Stifanosss

@stifanosss

Matt Elento

@mattelento

sh☆z

@certishaz

Claudio Marcio

@PesFcPlay

Academia for Palestine TCD

@AfPTrinity

Lutz Ziob

@lziob

IMCS Pax Romana-UG Official

@ugpax_official

@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
10 months
ZK has potential to save the world against deepfakes. So, we built @zkmicrophone from scratch in 2 days, which can prove: -Audio was recorded on a real mic -Audio edits authenticity while preserving privacy (by innovating on @NoirLang ) How we can fight deepfakes as society👇
Tweet media one
@ETHGlobal
ETHGlobal
@ETHGlobal
10 months
🎙 ZK Microphone The world's first attested microphone, using ZK-SNARKs to protect authenticity and privacy: Trusted audio in the age of deepfakes. @Eito_Miyamura , @dimitrov_radko , @ac_crypto , @lucasege , @manujmishra_
16
46
365
9
24
137
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
2 months
♨️ Over the weekend, I hacked a PoC for SuperParallel: Infinitely EVM-parallelizable (in theory), highest throughput EVM L2 in the world, a sort of "appchain" specialised for stablecoins like @circle , ERC-20, payments, and any parallelizable EVM code. Via mutex concurrency 🧵
Tweet media one
3
21
81
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
19 days
One of the application I'm excited about with GPU-EVM from @Gatling_X is the ability to do Monte Carlo Tree Search (MCTS) extremely quickly (Using the 100x exe. capability) This allows you to "look many moves ahead" and give you far better capability to look into the future
Tweet media one
2
3
36
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
14 days
🧵RL REQUIRES GPU-PARALLELISED SIMULATION: GPU-parallel sim env like GPU-EVM by @Gatling_X is critical for RL e.g. in NVIDIA Isaac Gym, a GPU-parallelised sim env, see the crazy speedup you get just by simulating robots on GPU. What does this hint about next phase of AI?
Tweet media one
1
4
28
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
9 days
Understanding CPUs vs GPUs: Broadly, CPU has avg. ~16 very "fast" cores (🐰rabbit), while GPU has avg. ~10,000 "slow" cores (🐢turtles) i.e. CPU - fast serial processing GPU - fast parallel processing This has implications in throughput vs latency 👇
Tweet media one
2
5
30
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
25 days
Was an extremely exciting and chaotic week coming out with GPU-EVM, thanks to everyone who gave incredible support ❤️💞 you know who you are
2
0
30
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
23 days
💡Unit test is a constant policy function test for bugs X~I(x=c) Fuzzing is a pseudo-uniform policy (+ mut. strategies) X~U One exciting application of GPU-EVM by @Gatling_X is training RL-optimised policy function that prunes search spaces intelligently to where the bugs are
Tweet media one
1
2
29
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
10 months
The solution we've built to fight the negative effects of deep fake audio!
@dimitrov_radko
Radoslav Dimitrov
@dimitrov_radko
10 months
Come check how you properly do attested microphone! @ETHGlobal @ac_crypto @Eito_Miyamura @manujmishra_ @LucasEge
Tweet media one
6
8
40
4
2
29
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
28 days
I think people don't fully appreciate Huang's law as the modern equivalent of Moore's law While this graph showed at the latest NVIDIA GTC is a bit unfair (comparing FP4 vs FP8), the general trend that we see at @Gatling_X is that speedups come from innovations on meta-levels of
Tweet media one
1
2
28
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
30 days
GPU-EVM is an EVM scaling solution so performant that cutting-edge RL-based AI agents can be trained on top of it. Some upcoming unlocked applications 🪄: 1) 🤖 Training AI Agents with the EVM 2) ⚡L2 Acceleration 3) 📈DeFi Simulation / Fuzzing
1
2
28
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
30 days
🎓 Detailed technical report soon on benchmarking methodology, technical capabilities, etc, as we launch products out of the lab, step by steps more details coming soon on how GPU-EVM accelerates AI, MEV, L2s, DeFi and so much more
Tweet media one
1
2
24
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
30 days
❇️ @gatling_x are committed to pushing the SoTA on AI, GPUs, Crypto, Distributed Computing that moves the world. ⚙️ We are hardcore: Off-the-shelf if available. If not, we build it. If you want to join the small, elite, tech-heavy squad, apply at:
Tweet card media
GatlingX Job Board | Notion

📚 Context

gatlingx.notion.site
4
3
24
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
2 years
Had a chance to meet and talk briefly to Vitalik himself at the @Edu_DAO hacker house! (No photos on his request) Turns out he can speak pretty good Japanese 😂 Thanks @BitDAO_Official @WindrangerLabs for providing the opportunity!
Tweet media one
Tweet media two
Tweet media three
1
3
22
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
1 year
Late post - I became a solo team finalist at @eth_vietnam 1.5 weeks ago, representing @sozuhaus , @HomeDAO_live @blockchainox , @Edu_DAO ! Built @tinyEVM , an EVM simulator designed to run on accelerated hardware. Now developing with ML experts to push to real application! 👇
Tweet media one
Tweet media two
Tweet media three
3
3
22
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
30 days
🤔 What is it? GPU-EVM is a breakthrough by our team of Oxford CS/AI Alumni: by GPU-parallelizing the EVM we massively improved the unit economics of EVM computations/second/$. 🤯Our current evm-bench benchmarks beat traditional CPU-based EVMs by 100x (log scale graph below)
Tweet media one
1
1
22
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
28 days
What are some benchmarks you'd like to see included in the GPU-EVM technical report? Already have a few stuff down like latency, unit economics, etc but would like to hear what else is important to people!
3
1
19
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
30 days
Thank you to @exolabs_ team @DCbuild3r @gakonst @__geometry__ @johannestknd for reviewing the materials for this post!
2
0
20
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
30 days
3) 📈DeFi Simulation / Fuzzing: 100x data crunching capabilities to find unaccounted edge cases in DeFi strategies / Protocol designs Think 🔫water gun vs 🚒water hoses for extinguishing trouble. Think Foundry+ in the early days 🌄Fuzzer coming soon
Tweet media one
1
1
19
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
15 days
Interested in pushing the foundations for the new age of exponential GPU computing, away from (dead) Moore's law? Good at GPUs / RL / symbolic solvers? We're hiring at @Gatling_x : Together, we will solve the hardest problems too hard for others to solve
Tweet media one
1
4
18
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
8 days
One of the applications I think will be low-key huge with GPU-EVM by @Gatling_X is EVM bytecode superoptimizers Superoptimizers afaik are computation-heavy, and henceforth you can use the GPU-EVM 100x execution to automatically optimize gas Who's an expert on this topic?
Tweet media one
3
0
18
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
30 days
📝 Wait-list for GPU-EVM by ♨️ @gatling_x : (We're scaling up GPU Capacity) 👀 Full announcement blog here w/ details & nuances: (🇨🇳 Chinese (中文) also)
Tweet card media
Introducing GPU-EVM: The Most performant EVM in the world

Setting the new state of the art on parallel-EVM execution

medium.com
2
1
17
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
30 days
🦾 like @nvidia 's IsaacGym, @google 's brax, and JAX-LOB, GPU-parallelized simulations give 100-250,000x performance improvements in RL training, data processing, and simulation. [src:] 🕹️GPU-EVM simulates the EVM, for general AI agents to interact with
2
1
17
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
30 days
1) 🤖Training AI Agents with the EVM: Remember AlphaGo beating Go world champions by discovering novel strategies? That was done via Reinforcement Learning (RL), by agents training within the Go simulation 💰RL allows training of AI aligned to your reward function, on its own
Tweet media one
1
1
17
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
7 days
🌶️ Hot take: For enterprise, I'm long Google > Open Source AI for intelligence In-context Learning (ICL) seems to work far better than finetuning in many benchmarking in many cases, and large 1m-2m context windows of Gemini-1.5 likely to beat
Tweet media one
1
1
16
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
30 days
SuperParallel PoC I made in a weekend:
Tweet card media
SuperParallel | ETHGlobal

SuperParallel - the highest throughput EVM L2 in the world designed for payments! Remove the bottleneck of longest-interdependent TX by design to EVM-parallelize effectively, while leveraging ETH...

ethglobal.com
1
1
16
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
30 days
2) ⚡L2 Acceleration: Accelerate existing CPU programs with GPU-EVM as a sidekick! 🫰🏻Imagine @monad_xyz as an L2 but utilising the GPU-EVM 100x higher throughput, with clever concurrency state partition optimisation for payment applications (More in 👇)
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
2 months
♨️ Over the weekend, I hacked a PoC for SuperParallel: Infinitely EVM-parallelizable (in theory), highest throughput EVM L2 in the world, a sort of "appchain" specialised for stablecoins like @circle , ERC-20, payments, and any parallelizable EVM code. Via mutex concurrency 🧵
Tweet media one
3
21
81
1
1
16
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
5 days
♨️ @Gatling_X in a single picture
Tweet media one
0
1
15
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
30 days
And lulu from @ABCDELabs !!
2
0
14
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
1 month
"The Bitter Lesson" is a famous note by Richard Sutton, a legendary ML pioneer observing the repeating cycle of mistakes with programs (see meme below) It also hints, to the fact that **search** scaled with compute is the next frontier of AI. But why? And why care?🧵
Tweet media one
2
3
13
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
1 year
Now all details revealed - large lineup of the industry leaders like @StaniKulechov @EliBenSasson @ArthurB ! We've filled the conference with educational ZK workshops, poker competition with @Pickem_Live or pitch competition with @CapitalTagus ! Let your friends in the UK know!
@blockchainox
Oxford Blockchain Society (University of Oxford)
@blockchainox
1 year
With our highlighted speakers announced, here is all the info out now! 📆🔥 SAVE THE DATE! Oxford Blockchain Conference, 18th June! 👉 Grab a ticket here! (It's FREE for students!)
Tweet media one
12
45
67
0
3
13
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
6 days
This is one of the most fascinating outcomes - the convergence of different modalities in their representations, from totally independent models This implies a universal set of representation that all models will converge to Questions: 1) How does synthetic data warp this
@phillip_isola
Phillip Isola
@phillip_isola
7 days
In particular, in an specific idealized world, we show that contrastive learners converge to a representation whose kernel is the pointwise mutual information function over the underlying events. On a simple color domain, this empirically holds. 5/8
Tweet media one
2
3
63
1
0
11
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
1 year
Recently hacked at #ETHTokyo to build an LLM chatbot that can interact with smart contracts directly, and allows easy swaps with @1inch fusion. Thanks to @1inch for awarding us #1 with the main prize! @blockchainox @HomeDAO_live @Edu_DAO
Tweet media one
Tweet media two
1
2
12
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
22 days
What real EVM contracts are interesting as benchmarks for GPU-EVM for stress-testing recursive calls / loops? Would be great if I can just copy-paste the code in one file, so we can run it locally, but can be multi-folder too We should probably have a newer ver. of evm-bench...
3
0
12
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
21 days
My co-founder HEB after coming out of the CUDA mines programming GPU kernel optimisations
Tweet media one
1
0
11
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
6 days
These LLM companies advertise extremely long context windows, but I think one of the things that need more knowledge is the fact that output windows are much much smaller than the "context window", and its occasionally quite misleading
Tweet media one
1
0
11
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
17 days
Recent GSM1k by @scale_AI is I think the best work exposing all of the standard LLM benchmarks contimination (MMLU, HumanEval, GSM8k, etc) The trend seems to be larger models generalising out of distribution, and actually learning reasoning capability
Tweet media one
0
0
10
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
27 days
If only execution was the bottleneck to Gigagas/s 😅 GPU-EVM hits tens ~ thousands of Gigagas/s consistently, but unfortunately there is a lot of DB work outside of execution that needs doing to achieve with a node
@gakonst
Georgios Konstantopoulos
@gakonst
28 days
So what is the gas per second on various chains today? A quick survey shows most chains at <10megagas per second, with Binance's chains leading the charge.
Tweet media one
5
3
46
0
1
10
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
9 months
Absolutely incredible achievement! This is definitely an AlphaGo, OpenAI Five, AlphaFold, ChatGPT moment for real-life robotics with RL Excited to see all the applications coming out
@davsca1
Davide Scaramuzza
@davsca1
9 months
We are thrilled to share our groundbreaking paper published today in @Nature : "Champion-Level Drone Racing using Deep Reinforcement Learning." We introduce "Swift," the first autonomous vision-based drone that beat human world champions in several fair head-to-head races! PDF
Tweet media one
41
341
1K
0
0
10
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
14 days
A consensus the industry is coming to, is that the order of AI takeover will be in order of cheapest and fastest simulation (i.e. availability of high-quality controllable and curatable data) GPU-EVM is a step towards this direction, and providing the high-quality steered data
1
1
9
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
10 months
@dimitrov_radko from our team in particular made huge technical innovations generalising and speeding up proof generation from 5 days to 5 sec on @NoirLang by operating in hash space. The consequences of this is that this can now extend to any medium (e.g. Image, video, etc)
Tweet media one
1
0
9
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
14 days
@DrJimFan @nvidia Take this recent breakthrough - robots balancing on balls! (Superhuman) How do they achieve it? Ans: RL training on GPU-Parallelized sim (Isaac Gym) then taking learnings in sim to real life (sim2real) because training in sim is VERY cheap & fast
@JasonMa2020
Jason Ma
@JasonMa2020
19 days
Introducing DrEureka🎓, our latest effort pushing the frontier of robot learning using LLMs! DrEureka uses LLMs to automatically design reward functions and tune physics parameters to enable sim-to-real robot learning. DrEureka can propose effective sim-to-real configurations
24
112
586
1
0
7
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
10 months
Discovered @siliconcompiler today which is really cool! Allows compilation of source code into silicon, just using a python API!
Tweet media one
0
1
8
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
2 months
🧠 @nvidia NVLink. the magic of centralisation NVLink is the linchpin that makes all LLMs work, that facilitate really fast GPU-GPU connection, as if 100s of GPUs were all one GPU. But did you know behind this, was a $6.9B acquisition of a networking company in 2020? 🧵
@sairahul1
Sai Rahul
@sairahul1
2 months
🚨 BREAKING 🚨 NVIDIA just announced Blackwell. The most powerful GPU in the market. Main features: - AI Superchip, 208B transistors - 2nd Gen Transformer Engine: FP4/FP6 Tensor Core - 5th Generation NVLink: Scales to 576 GPUs - RAS Engine: 100% In-System Self-Test - Secure
33
61
304
2
1
8
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
1 month
@mo_baioumy @sentient_agi @Gatling_X Thanks! Will not underdeliver.... Am very excited about @exolabs_ also!
1
0
8
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
9 days
Good analogy: buses vs cars Buses can transport more people, but takes longer to get from A -> B than cars due to needing to stop at every station Absolute execution of 1 EVM execution in GPU-EVM by @Gatling_X is slower than CPU, but in parallel far faster than CPUs
Tweet media one
0
0
8
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
10 months
Also thank you to everyone for your engagement and interest with the project! Only been a day, but we might quickly become the most engaged @ETHGlobal project ever 👀 Demo: Moving forward, @zkmicrophone will be developed under @HomeDAO_live !
Tweet card media
ETHGlobal Paris 2023 Closing Ceremonies

Join us for the closing ceremonies at ETHGlobal Paris 2023!Paris is the city of light and birthplace of many revolutions. Therefore we are thrilled to be bac...

www.youtube.com
0
0
8
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
14 days
This explains why self-diving never quite worked - the cost of simulation was too high. You can't GPU-parallelise the real world & simulate it. Flipping that, the applications that will be solved by AI will be ones where the cost of simulation is small
1
0
8
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
2 months
@0xdjma This has been the general consensus at the foundation model community for at least 6 months The general consensus is that DPO >>> RLHF and most models now are trained with DPO Here's a 3 minute summary I did on this 7 months ago on DPO intuition
Tweet card media
(Technical) LLM Jailbreak factory (ML Monthly August 2023)

Twitter: @Eito_Miyamura

www.youtube.com
0
0
8
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
10 months
Hardware signed audio is a good start - you can use digital signatures embedded in the secure enclave hardware to verify a audio came from a legit mic However this is brittle and breaks when you want to do something as trivial as: -noise removal -bleeping -compression
Tweet media one
1
0
7
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
30 days
@TheBlock__ Thank you to the Block for featuring us! more details coming soon on how GPU-EVM accelerates AI, MEV, L2s, DeFi and so much more: @Gatling_X
0
0
7
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
10 months
Huge thanks goes to @kobigurk @daniel_d_kang @AnnaRRose for inspiring us! This was a great Hackathon, with a all-star team. @dimitrov_radko @ac_crypto @lucasege @manujmishra_ Github:
Tweet card media
GitHub - Miyamura80/ZKMicrophone: 🎙🔒 ZK Microphone: Trusted audio in the age of deepfakes 🔒🎙...

🎙🔒 ZK Microphone: Trusted audio in the age of deepfakes 🔒🎙 Generative AI is a threat to society. It enables disinformation, manipulation, and political subversion. We've built the world&...

github.com
2
0
7
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
10 months
I think one of the most killer use cases of ZK and crypto is deepfake prevention and digital supply chain management with GenAI ZK privacy-preserving proof of computation (edit) can change how we think about genAI supply chains to keep us grounded in reality
1
1
7
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
10 months
And so our @zkmicrophone is a hardware + software solution, which allows you to record the audio with a hardware signature, and edit the audio as you please while preserving: -Hardware signature (i.e. Audio is real) -Privacy of unedited audio And know everything is legit!
Tweet media one
1
0
7
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
20 days
🌶️ Hot take: There has been no fundamental innovations in AI ever since 2017 on the path to AGI (maybe bar diffusion models) Rather, all technical breakthroughs have come from GPU engineering feats on training larger and larger models (RE: Bitter lesson) 🧵
1
0
6
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
10 months
@kobigurk @daniel_d_kang @AnnaRRose @dimitrov_radko @ac_crypto @LucasEge @manujmishra_ Huge thanks goes to @Quantstamp @ForesightVen @genventurecap for hosting us, and @sozuhaus for being the loveliest people, and never failing to impress Thanks to the @blockchainox crew and all the amazing people I've met during what felt like an extremely short trip!
0
1
6
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
30 days
@Gatling_X featured on the block!
@TheBlock__
The Block
@TheBlock__
30 days
GatlingX comes out of stealth, claiming 'most performant' parallelized EVM
9
11
65
0
0
6
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
10 months
E.g. Let's say a fortune 500 CEO is on a podcast and they accidentally say sensitive info. Then, you can bleep the sensitive part, while preserving the authenticity guarantees that it was recorded on @zkmicrophone , and keeps original audio with sensitive info private
1
0
6
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
1 year
This new NVIDIA DGX GH200 is an absolute game changer - literally 230x more memory than previous generations on the H100/A100 (144TB!!!!!!). Damn let the scaling hypothesis play out, with all your attention
Tweet media one
1
0
5
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
10 months
It's amazing to see huge multimodal models coming out to solve the data deficiency problem! It performs pretty decent on SOTA models - but this is only the beginning! Looking forward to how this field evolves
Tweet media one
0
0
6
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
1 month
@shaunmmaguire If you need help getting access to the right kinds of people at Oxford, let me know. Am based here, and can intro you to the serious startup founders here (which there are very few of, but very high quality nonetheless)
0
0
5
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
7 months
Will be with the team in Boston between 13th-20th, who should I meet? 👋
0
0
5
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
1 month
Liked a framework I heard today: GPT is the Heroku of AI Similar to how Heroku removed need for DevOps to deploy apps for startups (despite how expensive it was) GPT massively lowered barrier of entry for startups to add scalable, robust ML features
1
0
5
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
1 year
Cool points on zero interest rate policies by @gilbert and @djrosent helping capital intensive technology-risk businesses like AI or electric vehicles over the last 13 years, whilst distorting market-risk based businesses like all these marketplaces apps (e.g. Uber)
1
0
5
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
6 months
I think something that I find interesting is that Gemini/Alphacode is going the competitive programming approach to coding, GPT-4 is going the data route via just Vscode / github data brute force ingestion
0
0
5
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
3 months
I found a pretty broken prompt injection while playing out this LLM CTF: that carried me through all the levels besides the final level Hint: It uses an attack on tokenization LLMs aren't secure at all, it'll be hard for any meaningful privacy to be kept
Tweet card media
Gandalf | Lakera – Test your prompting skills to make Gandalf reveal secret information.

Trick Gandalf into revealing information and experience the limitations of large language models firsthand.

gandalf.lakera.ai
0
0
5
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
1 year
Was great speaking on @ZhenRongYap 's Zhen's podcast regarding Blockchains and geopolitics, AI and chips, information supply chains and truth! And watch out for what @ZhenRongYap does moving forward 👀
Tweet card media
Finding Truth in Atoms and Bits

Listen to this episode from Zhens' on Spotify. Eito Miyamura (Twitter: @Eito_Miyamura) is the President of the Oxford University Blockchain Society and is exploring the intersection of AI and...

open.spotify.com
0
2
5
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
5 months
@abstractooor Developing SC vs developing Hardware is actually an even better analogy They have the similar property, that once the product (the contract or hardware product) is deployed, it is very expensive to fix mistakes & often must recall
1
0
4
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
10 months
Really cool the direction Gorilla is taking for API databases!
Tweet media one
0
0
4
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
10 months
Deepfakes and GenAI are threatening the fabrics of our society. Take the @realDonaldTrump deepfakes, @ZelenskyyUa deepfakes announcing Ukrainians to retreat, revenge porn deepfakes. Solution? Hardware attested audio + Attestation on a trusted platform + Proof of audio edits
1
0
4
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
1 month
Is there a calendarly-like tool, with a UX like the following, s.t. it achieves the below outcome?: 1) Person gets cal link 1.5) (Optional) Receiver can connect their cal 2) Cal link shows times the optimal time to book, s.t. all meetings are batched together in one continuous
Tweet media one
2
0
4
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
9 days
KANs are busted. The below colab notebook rewrites KAN network into ordinary MLP with same number of parameters with slightly atypical structure. Universal approximation theorem of MLP is extremely powerful, and KAN is just a hard-to-GPU-scale awkward architecture
Tweet media one
2
0
7
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
20 days
🔮Model sharding + GPU-to-GPU distributed computing is the real magic behind GPT-4 💡Not like blockchain distributed computing, but like: "How do you simulate an LLM across multiple GPUs while it is sharded?" ♨️This is the kind of experience we gain at @Gatling_X
Tweet media one
1
1
3
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
2 months
@Tether_to @Visa @Mastercard @NethermindEth @ETHGlobal If you want to look at the code / slides, here it is: If you have any questions, please ask at: (may not respond on twitter)
Tweet card media
GitHub - Miyamura80/SuperParallel: ♨️ Highest Throughput EVM L2 PoC, ThreadSafe Execution ♨️

♨️ Highest Throughput EVM L2 PoC, ThreadSafe Execution ♨️ - Miyamura80/SuperParallel

github.com
1
3
4
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
11 months
According to crunchbase report, AI investments was 18% of global VC funding, but is down 28% since last year. Both # of deals and total $ invested are declining steadily Interesting how AI seems so hot rn, and yet this is the reality
Tweet media one
0
1
4
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
30 days
@sozuhaus Thanks to @sozuhaus @0xVEER @doctorgm_01123 !! Sozu Haus Vietnam was a life-changing moment for me, and these guys are some of the best people in this industry ❤️❤️❤️
2
0
4
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
5 months
@abstractooor Probably more modularized designs, you can think of Open Zeppelin as the equivalent
0
0
1
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
1 month
This is our core thesis at @gatling_x - if you're interested in pushing the frontier of AI with these learnings, please feel free to DM!
0
0
4
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
10 months
I believe general solution to deepfakes is permissioned deepfakes. As @kobigurk mentioned on his recent post, attestations allows everyone to see the audio is approved by the originator, deepfake or not. GenAI is inevitable - we should control deepfakes through permissioning
1
0
4
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
1 month
(AI Research) CLIP is dead, long live SPARC. You can run segmentation & much richer & fine-grained vision encoding, while having global alignment
Tweet media one
0
1
4
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
9 months
This is awesome! Applying DreamerV3 RL world model ideas to LLMs. Another nice framing to think about this is to think about it is a aligning an LLM to a simulation environment with just raw language
@realJessyLin
Jessy Lin
@realJessyLin
10 months
How can agents understand the world from diverse language? 🌎 Excited to introduce Dynalang, an agent that learns to understand language by 𝙢𝙖𝙠𝙞𝙣𝙜 𝙥𝙧𝙚𝙙𝙞𝙘𝙩𝙞𝙤𝙣𝙨 𝙖𝙗𝙤𝙪𝙩 𝙩𝙝𝙚 𝙛𝙪𝙩𝙪𝙧𝙚 with a multimodal world model!
4
113
531
0
0
4
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
1 year
Again, thank you to @sozuhaus in particular for hosting an excellent environment to work on these cool projects!
1
0
4
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
2 years
@Derekmw23 What's the argument for why metamask can't do this? Surely they could put in switching mechanisms depending on the user
11
0
0
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
28 days
@nican0r Thank you! Yes, we were really inspired by the ToB work as well when building GPU-EVM Things that were especially nice is just how little overhead the EVM has, compared to aarch64 binary exe. environment Also, in the ToB work, they go code -> LLVM -> PTX from remill, which
0
0
3
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
11 months
This is quite a cool PoC of using Zero-Knowledge SNARK proofs for deepfake protection But hashing and proof gen. for more complex edits takes too long - I imagine there is opportunity for tech pruning edit trees efficiently and generating proofs for it
Tweet card media
Fighting AI-generated Audio with Attested Microphones and ZK-SNARKs: the Attested Audio Experiment

By Daniel Kang, Kobi Gurkan, and Anna Rose

medium.com
0
1
3
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
9 months
This is great - maths of unsupervised learning in a extremely insightful way
@DrJimFan
Jim Fan
@DrJimFan
9 months
There're few who can deliver both great AI research and charismatic talks. OpenAI Chief Scientist @ilyasut is one of them. I watched Ilya's lecture at Simons Institute, where he delved into why unsupervised learning works through the lens of compression. Sharing my notes: -
55
432
3K
0
0
3
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
16 days
@cyodyssey Would highly recommend checking out Geometric deep learning by Petar Velikovic and Michael Bronstein and their YouTube notes if you want to dive deep into GNNs They are the leaders in this field, and their lectures have been nothing short of amazing (thigh sometimes too
1
0
3
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
1 month
@VictorTaelin Yup, makes sense FPL like HVM is very parallelizable My questions are more around which haskell applications would actually benefit from it, esp if you take thread divergence into account - e.g. I can see HVM parallelization for stuff like Yesod, Cardano being good (data heavy,
2
0
3
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
10 months
This is only a start of the solution to fighting deepfakes. @zkmicrophone has a lot of limitations (e.g. Playing deepfakes into the mic, which is solvable). But my hope is that this can unlock new waves of innovation ahead!
1
0
3
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
10 months
No more new (ad)ventures
Tweet media one
1
0
3
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
20 days
💡Takeaways: - LLMs are just big big big transformers - "AI Research" should focus more on what is scalable (and obeys the bitter lesson) + optimise around what existing hardware allows you to optimise - GPU Distributed Computing is the 80-20 of innovation over the last 7 years
0
0
3
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
1 month
But more importantly, an additional takeaway is that one should approach solving hard unsolved problems with AI carefully: 1) Ride on the a) Search b) Learning curve that scales indefinitely with compute 2) Leverage search, using LLMs to enhance their capability
1
0
3
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
9 days
Short notebook showing the reduction here:
Tweet card media
KAN is just MLP.ipynb

Colab notebook

colab.research.google.com
0
0
3
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
30 days
@punk5736 Actually the very opposite - A key thing is that we will be able to launch the best fuzzer in the world, powered by GPU-EVM, which will shoot down all the bugs before they make it to the application layer + RL security agents off-chain
0
0
3
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
4 months
Google search will decline in favour of LLMs as informal retrieval systems - in that case, what's the equivalent of SEO in that new paradigm? I think the answer might be more quite profound in the future, where applications will be built LLMs-first than human-first
3
0
3
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
1 month
@VictorTaelin How practical is IN for real applications? Church-Turing thesis gives you the equivalence, but if I remember the proof correctly, it gives you heavy overhead in the transformation - so with that in mind, what is the actual usefulness of massively parallel λ-Calc?
1
0
3
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
28 days
@mo_baioumy @exolabs_ @Gatling_X Thanks Mo! Extremely excited for the world to see what @exolabs_ have been cooking as well!
0
0
3
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
10 months
Proof of computation and edit to the digital supply chain while verifying authenticity. Imo, wealth of opportunity and innovations for this, and how we think about copyright, fair use, derivative works, authenticity Deepfakes aren't an invented problem, they are a problem TODAY
Tweet media one
1
0
3
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
14 days
Extremely surprising, perhaps even more surprising is the fact that 90%+ is coming from non-human TX Is this due to some weird regulatory gap?
@TurnerNovak
Turner Novak 🍌🧢
@TurnerNovak
17 days
Stablecoins are estimated to pass Visa in total payments volume in Q2. Anyone know where most of these volumes are coming from? And what they're being used for?
Tweet media one
61
47
341
0
0
3
@Eito_Miyamura
Eito Miyamura | ♨️ GPU-EVM
@Eito_Miyamura
30 days
@timtimtim_eth Thanks Tim! Always really appreciate your no-nonsense attitude and intellectual feedback!
1
0
3