Comment by marcus_holmes

I think this phase of centralising power is part of the never-ending cycle of centralisation and distribution - mainframes -> PCs -> websites -> apps, and so on round we go. We will get a "data centres -> Personal LLMs" phase of the cycle which distributes it again.

So my hope is that LLMs become local in a few years.

We've been sitting around 16Gb of RAM on a laptop for 10-15 years now, not because RAM is too expensive or difficult to make, but because there's been no need for more than that for the average user. We could get "normal" laptop RAM up to 16Tb in a few years if there was commercial demand for it.

We have processor architectures that are suitable for running LLMS better/faster/efficiently. We could include those in a standard laptop if there was commercial demand for it.

Tokens are getting cheaper, dramatically, and will continue to do so. But we have an upper limit on LLM training complexity (we only have so much Internet data to train them on). Eventually the race between LLM complexity and processing speed will run out, and probably with processing speed as the winner.

So my hope is that our laptops change, that they include a personally-adapted very capable LLM, run locally, and that we start to see a huge variety of LLMs available. I guess the closest analogy would be the OS's from "Her"; less typing, more talking, and something that is personalised, appearing to actually know the user, and run locally (which is important).

I don't see anything stopping Linux from doing this too (but I'm not working in this area so I can't say for sure).

Obviously we'll face the usual data thieves and surveillance capitalism along the way, but that's part of the process.