Reply to thread

Message: <blockquote data-quote="Ruin Explorer" data-source="post: 9676977" data-attributes="member: 18">I mean, I don't think there's much point arguing if you just believe this stuff, but hype, hope and outright lies are a major problem here because the insane amounts of money that have been pumped into LLM-based AI. Ed Zitron has been doing a good job covering the depth of outright lying that's been going on.I think it's extremely naive to actually believe these people making these claims unless they back them up with very solid explanations of how it's going to happen, and that's the precise problem - virtually everyone involved in AI at a high level at this point is resorting to vagueness and "well in five years maybe..." (Demis Hassabis, who I used to have a ton of respect for was engaging in precisely this kind of puffery recently. Why? Because if the AI investment bubble bursts, his company and countless others will collapse, and he and his colleagues will personally lose out hugely). Demis Hassabis is exactly a "person who should know" as you say, but the issue is, he's too involved, too close to the issue. And again as Ed Zitron and others have pointed out, a lot of journalists who "should know better" are just repeating hype uncritically (or with with very little critical thinking).More to the point, things don't get "cheaper and better" because it's a law of the universe or the like.They get cheap and better through specific mechanisms, and very often there's a limit - we're seeing that with GPUs right now - they simply can't make the same kind of pure-power gains generation-on-generation they could before, and not only is that making them more expensive - the direct opposite of "cheaper", but they're also having to resort of software tricks which aren't as good as pure power was, which is also meaning that they're not really "better" (framegen for example, which is actually only useful if you already have a high fps - 60+!). I think that's a really good example of why it's downright silly to assume "cheaper and better" is inevitable. It's absolutely not happened with "every other technology ever" - on the contrary with most there's a period that happens, before it reverses, and you start getting smaller and smaller gains for greater and greater cost until a fundamentally different approach is taken. I believe there's good evidence that LLM-based approaches to AI (specifically, not SLM, not AGI) have already hit that point. I don't doubt "AI" in a general sense will be a major part of the future. It's massive LLMs running on gigantic overheated datacenters I'm very skeptical about.I won't argue this further and take us any further off-topic, I mean it's your thread so I thought it was reasonable to respond here, but if you want to discuss it more maybe we should take it to another thread?</blockquote>

Verification