I played with local AI, but now I wonder:

isn’t decentralized AI better?

Local AI is centralized, albeit on you

waht is DecAI

You clearly want open models so while the following can apply to frontier and proprietary models, they are not my concern now 1

Open models run on a lot of (inference) providers. why don’t we just optimize for it? Local AI is just my inference provider anyway, it’s not free as in free software/speech and certainly not free as in beer. a good model is as costly as it gets and you will not likely a good quality from it.

The stack:

  • my computer
    • open harness
  • decentralized inference
    • open modelS

The Cons (I can think of)

  • The only reason why you wouldn’t have external inference is privacy, but you can anon your account for personal secrets (ofc they can still “correlate”, but adulthood is tradeoffs understanding and acceptance )
    • what about company secrets? well, how possible is for the inference providers to capitalize on your requests, expeciallyif the work will be open sourced/published
  • I don’t think censorship is a real problem, as the availability of inference providers improves (again going into a renting mindset instead of owning mindset 2)

  1. And btw, what’s a open model anyway? how do open weights compare to reading the source code directly if you don’t access to the training source? ↩︎

  2. tinyboxes will be interesting to evaluate further along the way… ↩︎