3 ms·
Until I can run frontier models locally on modest hardware, the hardware race will continue. It’s kind of like when 3d graphics came out. We all kind of knew ‘
by bloomingkales 2y ago
Until I can run frontier models locally on modest hardware, the hardware race will continue.
It’s kind of like when 3d graphics came out. We all kind of knew ‘the graphics are good, but not there yet’.
Eventually it does get to Crysis level graphics, and we all take a breather. But then it starts up again with Unreal 5 engine. We will get to photorealistic even if takes a 1000 years.
Same goes for AGI.
- sghiassy 2y agoI think that’s already been achieved https://youtu.be/o1sN1lB76EA?si=4nMZryeOeoVrZewW https://youtu.be/o1sN1lB76EA?si=4nMZryeOeoVrZewW
- leptons 2y agoAGI is a very different level of problem to solve than ray tracing. Photorealism is basically a straight shot, the algorithms for light are well understood, we just haven't had the hardware to do it in realtime, but the hardware was relatively easy to iterate on. AGI is not the same. No, LLMs are not going to turn into an "Artificial General Intelligence" just because we keep improving them - LLMs inherently don't have the capability of thought, interest, reason, motivation, or anything we ascribe to human intelligence. And throwing faster hardware at the problem isn't going to solve that. Sure, people are going to keep trying but I doubt anyone alive today will be seeing a true artificial general intelligence. In 1000 years? That's quite the goalpost.
- HDThoreaun 2y agoWe largely do understand how to replicate the brain though. Obviously not exactly but the brain is malleable so thats fine. The problem has been that we dont have the hardware to multiply matrices fast enough. LLMs are still orders of magnitude smaller than a human brain. Same problem ray tracing has.
- leptons 2y agoSorry but no, "multiply matrices" is not the fundamental problem blocking us from "Artificial General Intelligence", unless you believe in some alternate-facts description of AGI, as Sam Altman recently tried to foist upon the world by describing "AGI" as something capable of making money. That isn't the real definition of "AGI", that's a guy puffing up his company to try to get more money. https://en.wikipedia.org/wiki/Artificial_general_intelligence https://en.wikipedia.org/wiki/Artificial_general_intelligenc... A truly intelligent machine would need the capability to reason, understand, and think, exhibiting cognition, sentience, and consciousness - not merely sift through training data for a plausible response to an input. That isn't intelligence, that's a fancy search engine. LLMs and the entire premise behind them is incapable of producing AGI. They might seem like "AI", but they are nothing at all like AGI.
- HDThoreaun 2y agoThere’s no way to know until we build ones that are as big as the brain
- leptons 2y agoNope. You could build an LLM using every transistor that ever existed of ever will exist, and it would still just be a fancy search engine with no cognition, no consciousness, no actual thought. People will have given up on the LLM hype long before actual AGI is ever close to being a thing. Do you even know how LLMs work?
- HDThoreaun 2y agoIm not talking about an LLM. We largely understand how the brain works. With enough transistors we could emulate it.
- leptons 2y ago>We largely understand how the brain works. No, we certainly do not - we only have a very basic understanding of the biological component, nowhere near enough to replicate consciousness. You're fooling yourself if you think it's just a matter of throwing hardware at the problem.