Yet it doesn't detract from the fact that 20 years ago this was purely sci-fi; nobody - or a least very few - then envisioned that we'd have the capability we have currently have. And of course we continue to have the same vision now that X won't be possible for a great many years, if ever. And we also continually carefully refine the generally accepted definition of "intelligence" so that it specifically excludes whatever the current capabilities are, so we can indefinitely continue to say "this isn't intelligence".

> Yet it doesn't detract from the fact that 20 years ago this was purely sci-fi; nobody - or a least very few - then envisioned that we'd have the capability we have currently have.

I agree with that, but I don't think anyone is moving the goalposts as you imply later in your post.

I think if you showed me LLM AI twenty years ago, I'd be like "what's the trick, what's the catch, how does it work" and then the statistical nature of it would be explained to me and I'd have the exact same reaction as i had when i learned how it worked in ~2021: oh, that's very clever, and maybe even very useful, but idk if it's "intelligent"

Machine learning wasn't unheard of 20 years ago, and statistical text engines were hitting consumer use (iPhone autocomplete probably what 2008-2010?)

I don’t think it’s any more facile to use the term “intelligence” to describe the synthesis engines that we call “AI” when we use the same word to describe the gradient-seeking behaviors of slime molds.

It’s not general intelligence, but it’s a system that’s able to produce novel outputs from its inputs in pursuit of a goal. The fact that the goal is always externally-provided is more related to consciousness than intelligence.

It depends on how you define intelligence. There's a straightforward functional definition: if a system can solve a range of problems that require complex reasoning and it meets objective standards of success across multiple disciplines, then that system is intelligent.

Search algorithms used to be considered a part of AI research. The whole point of the field, from the Turing test onward, has been that mimicry is in some sense all you need. Maybe there's a coherent philosophical position where intelligence is defined as some intrinsic property possessed by conscious agents, but I find it remarkably hard to come up with a precise definition along those lines.

Agree, you're technically correct, and AI it's solving Erdos problems, even if we don't know exactly how or why. More mech interp & other research will help. https://x.com/OpenAI/status/2060451757818601808