Highly concerned how fast this is growing now. Over this week alone, seeing the network grow by almost 1000 unique connected nodes, every day, to almost 7500 live, connected nodes now.
Things don’t grow this fast, and also, shouldn’t grow this fast.
We seem to have flipped from individuals painstakingly running the home desktop client to industrial “miners” running a node on servers because it is hard and costly to run an always connected node. But people are doing it. And the network is currently holding up (we never stress-tested to these levels).
So
@HyperspaceAI is the largest distributed network of open source models from
@AIatMeta, also
@GoogleAI, and also
@MistralAI.
Thank you to the tireless work of the open source community who meticulously craft these models. Qr have connected 671 such LLMs in one distributed peer-to-peer network.
Hyperspace has now crossed over 4000 connected nodes for the first time.
This is a baffling metric, and makes Hyperspace one of the top global P2P networks in existence today. Some Q&As:
What does this even mean ?
* Each one of these nodes is available to serve inference requests for other random peers on the network.
* Each node has it's own unique identity on the network, is running an open source model either in the browser, in the desktop client or on the terminal/on a server.
* Again, this is one, interconnected network.
* Each one of these nodes is also being sent a Proof-of-FLOPS matrix multiplication challenge on an hourly basis based on their VRAM pledged to verify the authenticity of their compute provided.
How is this network supported ?
To support these 4000 connected nodes, the Hyperspace matchmaking server currently gets over 100 million requests, every day. There are engineering optimizations which we need to do across the client and the network stack in order to reduce this and make the system more efficient, in addition to pursuing a decentralization roadmap.
The average latency is also in the 500ms range - which makes the peer-to-peer network slow for a human eye for which anything over 100ms is perceptible delay. Thus such networks are better suited for use by AI agents on long-running parallel tasks where intelligence scales as a function of parallel compute, and not live feedback to humans in real-time, where it is better to run AI on-device directly (as the aiOS is built for). We view AI as a complex multi-system engineering challenge, not just on a one-off well-trained model.
How did we get here ?
Earlier this year we started seeing a spike in number of nodes, which quickly overwhelmed the system. These nodes originally were low quality, but due to the sheer number of them and across Windows, Mac and Linux, we had to do a fundamental architecture re-write across the client nodes and the network, in addition to building up an incentive structure to get more reliable nodes on the network. It is due to a summer of working hard and writing code, that we now have a stable network of good nodes which we did not have before.
People running nodes now are essentially miners, so they are incentivized to run more powerful machines with more useful models as the system specifically incentivizes that.
How does it compare to other P2P networks ?
Think about Bitcoin, it has just over 19k connected nodes, where each node is competing in a cryptographic challenge in order to earn the right to mine a block. That network is fully permissionless, unlike the Hyperspace network which currently uses a centralized architecture for orchestrating consumer peer-to-peer nodes much like Uber does.
Think about Ethereum, which collectively has just over 6k unique nodes, where each node in addition to some compute also has to provide an economic stake in order to participate in the network. That network is permissionless as well. Likewise for Solana, with 1363 unique nodes serving the network today.
It is the Lindy effect, the permissionless aspect, the much higher requirements and the long-running nature of these networks which give their underlying currency a value in the real world. In the case of Hyperspace, the underlying system currency called "flops" is an in-game worthless currency meant to use applications running on it: it is not permissionless and it is not a blockchain, and can theoretically be manipulated by the company.
NOTE: we are on a path to make the network fully permissionless in the years ahead, and you can see that roadmap in the flops announcement. Currently we are bootstrapping the habit of participating in such P2P AI networks which consumers did not have before us and using social, utility and cryptoeconomic incentives at play.
Will there be an API ?
Yes, this is coming soon and will have additional capabilities such as running embedding jobs, serving vectors and more.
Will it integrate with other decentralized AI protocols ?
Yes, we believe in offering choices to users, in this case the mining nodes on whichever protocols they want to support on their devices. We take care of the AI-related plumbing as part of aiOS such as providing access to 671 models, 3.2 TB vector database with 400 million embeddings, real-time web search and other attributes of the system to make it useful.
Will there be a full blockchain node one day as an add-on ?
Yes, we are firm believers in consumer devices being suitable for block production and validation process, and entirety of our research and engineering has focused on that for the past few years (we have written multiple papers in that regard).
Think Full Self Driving in your Tesla, for those who want it.