5 ms·
The dose makes the poison, what if a new method was 1,000,000x more efficient would there still be more total money spent on GPUs because of it. What if efficie
by cjbgkagh 2y ago
The dose makes the poison, what if a new method was 1,000,000x more efficient would there still be more total money spent on GPUs because of it. What if efficiency brought it down such that inference on CPUs that people already have is good enough. I see a large demand curve but not an infinite one.
For example laptops are getting faster every year but most people I know are not clamoring for the latest laptop, same with phones, good enough is good enough and people would rather save the money. If a laptop came out that was 50x faster for the same price people wouldn’t buy 50x more laptops, they would wait even longer to upgrade.
- aurareturn 2y agonew method was 1,000,000x more efficient would there still be more total money spent on GPUs because of it In my opinion, yes. If scaling laws continue at 1,000,000x, we will turn the whole planet (or Mercury) into a giant GPU.
- cjbgkagh 2y agoDo you have a use case for your opinion? On the assumption that we’d want super intelligence to do interesting things, consider that we currently have many very smart humans and society is not maximizing their use, quite the opposite.
- whatshisface 2y agoVery smart humans don't want to do random tasks, they want to study number theory.
- cjbgkagh 2y agoNot all, and many smart humans are prevented from doing what they want to do for all sorts of reasons. One reason being others feel threatened by them, but those same people will feel threatened by AI, probably more so.
- aurareturn 2y agoDo you have a use case for your opinion? Use the ultra smart AI to convince you that I’m right. :)
- cjbgkagh 2y agoOh I see, the underpants gnomes path to profitability. Makes sense.
- roenxi 2y ago> what if a new method was 1,000,000x more efficient would there still be more total money spent on GPUs because of it Obviously yes; 6 orders of magnitude would result in insane amounts of money swooping in. The market has currently been going crazy over relatively minor improvements compared to that. > If a laptop came out that was 50x faster for the same price people wouldn’t buy 50x more laptops, they would wait even longer to upgrade. That would be equivalent to decades of improvement in the tech. The last time we did that (over the last few decades) we saw massive increases in spending on computers.
- cjbgkagh 2y agoHence my very first point; the dose makes the poison. Yeah, it would be super impressive, perhaps another point SMT solvers algorithms have improved faster than the hardware for a very long time and are incredibly useful. But we are not spending infinite money on SMT solvers, quite the opposite.
- roenxi 2y agoBut we are spending more than ever before. I don't see why you think being unable to spend infinity dollars matters here. There are a lot of numbers between larger than current spend between where we are and infinity.
- cjbgkagh 2y agoTo my point, how much do you think a SMT solution costs now?
- roenxi 2y agoCouple of thousand dollars? Consultant turns up on site, gathers data, talks to a few people, sets up a model and solves it. Probably around a weeks work @ say $100/hr. Might be under-calling it for validation and there'll probably be some iteration to do so a project will take a few solves to complete + overheads.
- lherron 2y agoIt’s a question of depreciation of the GPU vs time to deploy AI. If RL and synthetic data creation are all we need for self-improving AI, the large labs have an internal use case to utilize all available compute, regardless of efficiency gains. If they have to go external and deploy in the broader market, time to deploy would make the GPUs worthless before they are fully utilized.
- cjbgkagh 2y agoDid they not already have an incentive to use all available compute? Let’s assume you were buying the GPUs, how would you plan on paying for it? What’s the business case?