Will a recognized entity create a GPU compute sharing platform that let anyone share GPU for models training or inference before 2026 ?
You're probably familiar with 'Folding@home' project that enable anyone to share compute power to solve scientific problems related to the folding of proteins.
Will we see a similar compute sharing project but focused toward training or inference of LLMs ?
1,000People are also trading
@tedks I see ! BitTensor seems to check many boxes of this market, the only point that let me hesitant to resolve is that the tech seems still naissant and have other prerequisite that I consider would not 'let anyone' use this, like having a crypto wallet and having to go throught a variably complex node configuration, having to buy and stake TAO tokens to start 'mining'. Also it's not really clear to me what they actually do with your compute when you start 'mining'.
I wouldn't be surprized if a wrapper layer system would make the use of BitTensor or equivalent possible with a lower entry barrier and no initial token to provide would exist in the coming months. I this case it will 100% fit the description to me.
For example if BitTensor or other release a small weight all in one client script you can run and configure on a mac or linux machine (windows optional) and would not require buying tokens from the start that would count (having to pay to share compute seems odd to me)
I exclude any solution that would require paying for additionnal hardware or distant compute like AWS instance for example.
@Guillaume As far as I understand (which is not very far), when you are "mining" in a subnet, you are running inference on a task defined by the validator, and the successful "miner" is the one whom the validator picks as the best answer. You don't seem to have to stake any TAO to start mining: see https://github.com/opentensor/prompting?tab=readme-ov-file#running. You could package this into a Docker container with a script that would create a wallet, register you into the subnet with proof-of-work (no tokens required), then start mining.
This is basically renting at scale, as far as I can tell, not "sharing" per se: each miner is running their own, self-contained task, so that might be against the spirit of the market.
There are other BitTensor subnets that provide training, see https://github.com/bit-current/DistributedTraining?tab=readme-ov-file#running-a-miner-on-hivemind--a-step-by-step-guide
Hypertensor summary paragraph really seems to fit to the market, I'll try to follow the release in Q3 when it comes and maybe test it :
Hypertensor gives users the ability to contribute computational power towards machine learning models in a decentralized network. In return for contributing computational power, incentives are rewarded. Incentives are based on a peer-to-peer ranking system that is based on throughput, model hosting, availability, ping, stake balance, and other parameters. Hypertensor is built incorporating Hivemind, a library for decentralized deep learning across the Internet. Hivemind is built using libp2p, a well established p2p library that many blockchains use, such as Ethereum and many others. We put forward the concept of incentivizing the contribution of computational power towards machine learning models by integrating blockchain technology that will validate throughput, secure the models, and act as a payments & transactions infrastructure.
I think this p2p ai compute technology is really coming ...