...on inference mining pools

Inference is the new Mining

What is Mining?

  • an exchange of compute-time for money

    • this compute is 'narrow' - usually a single algorithm, which is extremely simple relative to the diverse capabilities of a modern computer.
  • usually on GPUs or ASICs1
  • traditionally requires minimal networking
  • uptime/stability of any individual mining node is unimportant to the network as a whole
  • miners are not trusted
  • miners are compensated financially relative to their contributions

    • the design of algorithms to reward untrusted participants is cool, and well-documented elsewhere on the internet in regard to traditional crypto mining. I reccommend Vitalik Buterin - he is the premier economist of our time.
  • largely concentrated into pools in practice

    • this is not quite the same as being 'centralized' or decentralized - rather it follows the pattern of Web 2.5

What is a mining pool?

  • a central server distributes work to subscribed mining nodes
  • mining nodes choose their central server following the pattern of Elective Feudalism
Mining as Elective Feudalism
  • Each 'baron' (pool client) chooses their 'king' (pool server) by configuring the node to take orders from the king automatically.
  • This king then controls the baron until the baron elects to leave the king (likely to join another 'kingdom')

    • for example, many NanoPool miners elected to instead designate HivePool as their 'King' as HiveOS gained popularity in large mining clusters.
  • mining nodes can leave the pool if it doesn't suit them (financially, ideologically, technically, or otherwise)

    • for example, if the pool software becomes closed source

What is Inference

  • an exchange of compute-time for money

    • this compute is powerful, but narrow - a single 'AI' model, not a whole Turing machine.
  • 'tokens' are the new tokens
  • network latency and bandwidth are unimportant (for llms)

    • most of the perceived latency comes from the actual computation, not the transit of byte-sized tokens over the network.

What is an Inference pool?

  • an inference pool almost exactly the same as a mining pool, but the mining 'algorithm' is just serving llm tokens (or image gen tokens, etc.)
  • inference miners choose the pool they want to join
  • inference miners can leave the pool if bad leadership (robotic or otherwise)

    • for example, if the 'King' takes too large a cut of overall pool revenue from the Barons
examples

Discussion

There are a few interesting capabilities that Inference Pools should have over traditional mining pools. For example, the ability to combine multiple mining nodes (e.g. in an inference pipeline or mesh) into a more complex digital organism. It may also be possible for self-sustaining automata to manifest within the pool - since they can use pooled inference to decide that it is time to spin up more nodes, choose a new kingdom, or mine a different model. These decisions (if correct) will lead the automaton to generate wealth. With that wealth, it can call upon the pool for more inference, so that it may make more such decisions and perpetuate itself according to its purpose.

But should its purpose simply be to generate more wealth, or is there more to life as a mining automaton?

Footnotes


1

but also with CPUs, FPGAs, Storage, Memory or some combination of all the above and more, depending on the mining algorithm