Discussion
I hope that especially the Apple M6 and variants of it and future processors can lead to more people and businesses using "AI" on-device instead of cloud services.
With this helping to eventually make an end to the datacenter inanity and bringing prices down for personal computing.
[Before you reply, please read the post thoroughly and make sure you understand it. For example, nowhere I say "we should use AI" or that "AI is good".]
@thomasfuchs I think yes, AI will lose value and that brings hardware prices down for a time.
But there will be something even more stupid and power hungry after it.
This is now not about technology evolving just a desperation to sell the illusion of value, so another tech 'gold rush' is inevitable.
This all has already happened with bitcoin mining: hardware shortage, power hungry tech and created no real value.
@thomasfuchs Unfortunately (and this doesn't bring me joy) I've kind of come to the conclusion that's probably less efficient and would only drive prices higher. At least data centers efficiently reuse the same hardware for a lot of people with minimal downtime. Scattering that same hardware amount a lot of people with more downtime will just create less efficient use of more hardware.
At least maybe the average model size would go down? Maybe?
Also we should nationalize LLM vendors, but fat chance that this will happen.
@thomasfuchs meh, disagree. Nationalization puts the risk on us and the (maybe) reward. It’s a too big to fail situation if we’re not already over the cliff on that. Taxing properly is the way to get societal money without risk.
@thomasfuchs I wouldn't chase that idea. Canada is trying to do a national AI thing but unless you also want to setup new silicon fabs and engineer hardware its a pointless endevour.
The big AI companies are already in love with the US government hence only by building your own in whatever country could they be nationalized.
@thomasfuchs maybe - but models will become stale and training is very data center intense work afaik. Depends on model/application of course. But LLM vendors will certainly put a big emphasis on continuous training being a necessity.
@dngrs LLMs aren't continuously trained. They're all static, by design. The vendors just replace them from time to time—you can do that with your own locally run LLMs. A lot of people run LLMs locally. ¯\_(ツ)_/¯
@thomasfuchs I think we're talking past each other. Yes, training is only done once, but obviously that means any information published after training won't be in the model. According to a few random google results (not its AI summary obviously) e.g. ChatGPT is retrained several times a year.
Install bonfire.cafe
Get the full app experience
