Hyperspace AI: a P2P AI network and AI-native L1 secured by EigenLayer
Thread1,055 words
Breakdown of Hyperspace: a peer-to-peer AI network on consumer devices with data, vector ranking and LLM OS layers, planned to become an AI-native L1 bootstrapped on EigenLayer. Cites papers showing 200k TPS and 4.4s finality in simulation.
Key claims
Research papers cited show 200k TPS and 4.4s finality in simulation
Consumer devices represent 20,000+ exaflops of FP32 compute this decade
Q2 2024 roadmap: BitTorrent for AI protocol settling on Ethereum with EigenDA
The Age of AI is upon us! 🤖
In 100s of research hours I have identified a project building the base layer for decentralized AI!
TLDR of a transformative vision for an AI-native L1 blockchain with a novel AI virtual machine enshrined & secured by @eigencloud!
Alpha-🧵
1/ AI will be one of the BIG narratives in 2024 & beyond!
Using @eigencloud to bootstrap network security from Ethereum is the frontier of infra innovation.
Finding a gem that is building the future of AI on @eigencloud is not easy. But it might come with life-changing gains!
2/ This project is on a clear mission!
They want to make AI free and abundant, both as in beer (lower costs) and as in speech (freedom of expression and use).
Free from rate limits, model limits, and free from biases.
3/ They also want to make it smarter.
That's only possible with the right data strategy.
It's not just about the LLM, but also about how it is prompted.
4/ Android enabled choice of hardware and programs, while lowering costs for smartphones.
Today, over 70% of all smartphones in the world are Android-based.
The project I found wants to start a similar revolution in the AI space.
5/ While OpenAI has shown the way, this team sees AI as modern day electricity & believes that everyone should have it limitlessly.
It should be decentralized, in clusters of self-forming communities.
No one entity should control the AI experience for billions of people.
6/ So how can we achieve that?
Enter @HyperspaceAI, an open design for a peer-to-peer AI network which works for generative AI on consumer devices.
7/ 🖥 Compute
Base P2P AI network running on consumer devices which can scale to millions of nodes.
Laptops, desktops and smartphones - just this decade representing 20,000+ exaflops of FP32 compute.
AI projects can use the dev API to tap into the network to lower costs.
8/ 💾 Data
Wikipedia-like curation tools & protocols which lead to tens of millions of finetuned open source small (Mistral7B-like) models (including those developed by Hyperspace).
Finetuned using and serving inference on the above P2P node network.
9/ 🌐 Vectors
VectorRank is a distributed system & model that extends the concept of PageRank to vectors (think Google).
It aims to provide smarter ways to rank vectors based on credibility, usinf in-document references, semantic analysis & parent doc PageRank (better results).
10/ 🤖 LLM OS:
Comes with s fully integrated UX, post-browser consumer product which is AI-native and seamlessly combines the ecosystem of compute, data and vectors in one, integrated experience, so that it is competitive with OpenAI at a fundamental level.
11/ So at its core, @HyperspaceAI is building a decentralized super-brain.
The more capabilities this distributed brain has, the more value it can provide to users and developers.
12/ Eventually, Al agents would be able to run seamlessly on this network, engaging in financial transactions with each other using the set of blockchain protocols enabled on this P2P network.
13/ The @HyperspaceAI team believes this world will run in browsers, and the entirety of their research has been around enabling:
▶️ Full Block Production in the Browser
▶️ Generating ZK Proofs in the Browser
▶️ Running Inference in the Browser
14/ The web scale protocol we designed utilizes:
▶️ Sharded State
▶️ Parallel Block Production
▶️ Fast Finality
This enables one of the highest throughput blockchain protocol designs at-scale.
hyperrealistic pepe riding futuristic rocket through space
15/ More specifically, two recent and rather detailed research papers, with theory-simulations have shown this design reaching:
🚀 200k TPS
🚀 4.4s finality
16/ What this @HyperspaceAI supercomputer looks like?
It's basically stitching together 20,000 exaflops (FP32) of spare compute on consumer devices that will be available this decade to serve AI.
17/ The strategy here is simple.
Make it as easy as possible for anyone to run a node on their personal devices.
A simplified one-click setup is already available for both Intel and Silicon-based Macs.
Serve Mistral-7B-Instruct-v0.1 to others using your spare compute.
18/ Can this even scale?
Yes! Research shows consumer-grade GPUs almost reaching parity with the performance of 10x more expensive server-grade A100 GPUs for running LLMs.
This will turbocharge networks like @HyperspaceAI, reducing the cost of inference even as AI usage scales.
19/ Eventually, @HyperspaceAI will be an AI-native Layer 1 blockchain which enshrines a novel AI virtual machine and is solely bootstrapped on
@eigencloud.
Essentially building the future of democratic AI systems on top of @eigencloud & providing scalable AI infrastructure.
20/ What is @eigencloud you ask?
EigenLayer is a general-purpose system for programmable trust.
It can be utilized to build out generalizable middleware services to fulfill almost any function (e.g. secure a decentralized AI network).
x.com/zerokn0wledge_/status/1601921935662731265
21/ Back to @HyperspaceAI
What's the timeline?
✅ Q4 2023: Starting deployment of integrated web app & P2P network
⏳ Q1 2024: Community data tools & dev API
🎯 Q2 2024: BitTorrent for AI protocol with settlement on Ethereum using @eigencloud's EigenDA & VectorRank deployment
22/
🎯 Q3 2024: LLM OS consumer UX
🎯 Q4 2024: BitTorrent for AI protocol settling on a PoS protocol designed for massive concurrent usage at a scale of millions of users. Enabled by state sharding, parallel block production & off-chain computation, secured by EigenLayer
23/ Who exactly is behind this ambitious vision?
The CEO @varun_mathur, with a background in cross-domain research, has invested his entire personal net worth in this.
He also has experience in building consumer tech products & working as a VP data analyst at Merrill Lynch.
24/ The @HyperspaceAI team includes experts in:
▶️ QuantumBFT and RLHF mathematics
▶️ Petabyte-scale data systems
▶️ Distributed systems architecture
▶️ Protocol research
▶️ AI engineering
▶️ Cryptography engineering
▶️ Game theory.
25/ @HyperspaceAI Advisors consist of leading experts in crypto, AI, and industry experience.
The venture has attracted investments from funds in the US, UAE, and Singapore, reflecting its global vision.
26/ Many ask how they can make it in the upcoming cycle.
The answer is simple.
Identify key narratives (AI).
Find the projects building key infra (Hyperspace).
Identify hype-driving catalysts (EigenLayer).
Build a thesis and find a way to invest.
Keep conviction & make it.
Tagging some chads you should definitely follow! 🫡 👇
@ShivanshuMadan
@0xDinoEggs
@modularmedia_
@therollupco
@alice_und_bob
@Louround_
@0xSalazar
@andyyy
@0xDinoEggs
@WinterSoldierxz
@poopmandefi
@hmalviya9
@thesaint_
@0xSalazar
17 images on X
What came next · since 02 Jan 2024
Since then
Zero has written 1 more piece on Hyperspace since this one.