TheGuySwann on Nostr: Doesn't seem to be true, in fact while many of the major companies claimed you ...
Doesn't seem to be true, in fact while many of the major companies claimed you couldn't run it without a super computer, after only a couple of weeks after META's weights leaked, the open source community had Llama running on basic consumer hardware. Even got it running (very slowly) on a raspberry pi.
In addition it seems to becoming apparent that just cramming in waaayyy more params isn't a good way to "learn," instead curated, micro step learnings from a huge open source community that get recombined and grow naturally is far faster and can produce equivalent or even better results with just a few Billion params on cheap hardware than the big guys can do with 100s of B of params and a super computer.
I think we are mistakingly thinking of this new tech like previous giant, centralized platforms, but it's beginning to look very much like the positive feedback of network effects may not apply here. Trying to centralized and lock down your weights may be a death sentence in the not-too-distant future.
I'm specifically starting a podcast because I don't think we have the right perspective on this, and I believe our attempts to control & centralize AI will actually produce the *perfectly opposite* outcome as we are hoping for, as basically all rushed, arrogant (no disrespect, but the idea that we are going to "control" this is our inner statist speaking) decisions and knee jerk reactions tend to be. I don't think the above course of action is being proposed out of knowledge, but because we are afraid. Decisions like that historically tend to produce poor results.
TL;DR - I think AI is too powerful and important to let government "control" it or "pause" what is happening. I think that will spell disaster... plus they have virtually zero chance of succeeding anyway. IMHO
In addition it seems to becoming apparent that just cramming in waaayyy more params isn't a good way to "learn," instead curated, micro step learnings from a huge open source community that get recombined and grow naturally is far faster and can produce equivalent or even better results with just a few Billion params on cheap hardware than the big guys can do with 100s of B of params and a super computer.
I think we are mistakingly thinking of this new tech like previous giant, centralized platforms, but it's beginning to look very much like the positive feedback of network effects may not apply here. Trying to centralized and lock down your weights may be a death sentence in the not-too-distant future.
I'm specifically starting a podcast because I don't think we have the right perspective on this, and I believe our attempts to control & centralize AI will actually produce the *perfectly opposite* outcome as we are hoping for, as basically all rushed, arrogant (no disrespect, but the idea that we are going to "control" this is our inner statist speaking) decisions and knee jerk reactions tend to be. I don't think the above course of action is being proposed out of knowledge, but because we are afraid. Decisions like that historically tend to produce poor results.
TL;DR - I think AI is too powerful and important to let government "control" it or "pause" what is happening. I think that will spell disaster... plus they have virtually zero chance of succeeding anyway. IMHO