[spion]

I'm not entirely sure about the centralization part. We already have models that can run on consumer hardware and are freely available to use for anyone (e.g. code-llama 34B is actually a viable gpt 3-5 replacement, if not slightly better)

Training these is still out of reach, but fine tuning is getting close (LoRA) and running them is almost easy at this point.

The products we've built so far are power-centralized, but augmentative. Where we go from there is up to people, not the nature of the technology. My hope is decentralized and augmentative, but the worst case scenario is indeed centralized and substitutive.

[danpalmer]

By centralisation I don't necessarily mean of the models. It might be the training, training your own model requires expensive hardware (even just a high end graphics card is out of reach for most of the world).

But also, a model running on your phone generating AI content is likely to be cheaper and not as good as human curated content in whatever form that is.

I think compute can be decentralised, while the power is still centralised, or at least those at the low end lose out on quality.

[hdjfkfbfbr]

I think you need a caveat... Right now.

There was a time where computers took up the space of whole rooms and had much much MUCH less processing power than the phone I am currently writing this comment on...

The same thing will happen with ai.

[octacat]

oh, believe me, regulations would be lobbied. I.e. to run an AI model as a part of some service, you would need to get certification (a lot of money). Even now you could see a couple of talks on tech conferences from "non-profit" organisations. Also, web services are highly centralised, so there are all the chances to get highly centralised AI services too.