The number of io.net GPUs is a mystery. What are the problems with decentralized AI protocols?

Author: @rargulati, MartinShkreli;Translation: Vernacular Blockchain

io.net is a decentralized computing network built on Solana, belonging to the Depin and AI sectors, and has received financing from Mult1C0in Capital and Moonhill Capital, and the amount of financing has not been disclosed.

io.net is a decentralized cloud platform for machine learning training on GPUs based on Solana, providing instant, permissionless access to global GPU and CPU networks. The platform has 25,000 nodes and uses revolutionary technology to cluster GPU clouds together, saving up to 90% of computing costs for large-scale AI startups.

Currently built on Solana, it belongs to the currently hot Depin and AI sectors. Today, let’s take a look at the two people on X’s analysis of its GPU and existing problems:

How many GPUs (Graphics Processing Unit, a graphics processor is a chip used for graphics processing) does @ionet have?

X @MartinShkreli did an analysis of four answers:

1) 7648 (when tried at deploy time)

2) 11107 (manually calculated from their resource manager)

3) 69415 (unexplainable number, unchanged?)

4) 564306 (no support, transparency or substantive information here. Not even CoreWeave or AWS have this many)

believes the real answer is actually 320.

Why 320?

Take a look at the explorer page with me. All GPUs are "free", but you still can't rent one. If they are free, why can't you rent them? People want to get paid, right?

The only ones you can actually rent are 320.

If you can't rent them, then they don't really exist. Even if you can rent it, it will increase...

@rargulati said that Martin is absolutely right to question this matter. There are the following problems with decentralized artificial intelligence protocols:

1) There is no cost-effective and time-efficient way to perform useful online training on highly distributed general-purpose hardware architectures. This requires a major breakthrough that I don't know about at the moment. This is why FANG spends more money than all the liquidity of cryptocurrencies on expensive hardware, network connections, data center maintenance, etc.

2) Inference on general-purpose hardware sounds like a good application case, but the development of hardware and software is so rapid that general-purpose decentralized methods do not perform well on most key use cases. You can refer to the latest OpenAI delays and Groq's growth.

3) Infer from correctly routed requests, clusters of gpus co-existing closely with requests, and leverage decentralized cryptocurrencies to drive down the cost of capital to compete with AWS and incentivize enthusiasts to participate. Sounds like a good idea, but with so many suppliers and dispersed liquidity in the GPU spot market, no one is consolidating enough supply to provide to people running real businesses.

4) Software routing algorithms have to be very good, otherwise there are a lot of operational problems with general-purpose hardware from consumer operators. Forget network breakouts and congestion control, if someone decides to play games or use anything that uses webgl, you may experience service outages on one of the operators. Unpredictable supply side causes operational headaches and creates uncertainty for demand side requesters.

These are hard problems that will take a long, long time to solve. All the bidding is just a joke.