Reading tea leaves, the market believes that whomever is first to AGI will reap all the benefits, leaving nothing to the runners-up. In this view, even a tiny edge in model quality is decisive.
If "the market" believes this, I'd love for "the market" to explain to me how they think this works in practice.
Like if Anthropic hits AGI on Monday, what stops Google or OpenAI from hitting AGI on Tuesday?
There is no substantial difference in capabilities between any of these models, they are all doing the tiniest little leapfrogs over each other from release to release.
Even if the market thinks this leads to AGI, there is no reason to believe any secret magic exists that makes one of them reach AGI without the rest hitting AGI nearly immediately after.
Of course, things like jetpacks and flying cars also have had that feeling. As well as Stephenson's metaverse.
I'm somewhat inclined to give it to them at this point, I mean it's artificial, general in the sense that it can complete tasks it wasn't programmed for, and intelligent enough I guess. I mean if there was a planet with creatures at the level of LLMs we'd probably call it intelligent life—used in that way we're not really implying the creatures are particularly smart. It's as much simulated disembodied intelligence as you could reasonably hope for. I wouldn't at all expect that simulated intelligence: the computer algorithm would have any relation to actual intelligence except at the minimum superficial level needed to complete tasks. With AGI out of the way we would at least have to be more specific on what the sci-fi dream actually entails. The computer from Star Trek is certainly intelligent but you would never say it has a brain or anything close to it.
AGI would mean something with sentience or consciousness. Data is AGI. Even Exocomps were AGI. LLMs don't come close, aren't getting any closer, and there's no reason to think they ever will.
And I don't really think it's entirely unreasonable because there does actually seem to be a real gap in terminology. Even Star Trek didn't really imagine a computer system that's general but exists without consciousness, the whole message of the Exocomp story is exactly that—it's a huge breakthrough that they invented a general system and it turns out to be conscious. And the show makes a point that while the computers they have are incredibly advanced it's ultimately the product of human programming. We seem to just have assumed that any general system would be conscious and the term AGI was born of that assumption.
"Still"? I mean if someone believed that LLMs would result in AGI, why would he change his opinion? Did something happen? I mean the progress generally exceeded expectations so far so your P(AGI) should increase, not decrease. This rant doesn't apply if you did not believe that LLM can result in AGI from the beginning of course.
I like the way Kevin Roose put it: if you were to bring Astra or Opus 5.5 back to 2017 and show it to the leading AI researchers at the time and ask them if it were a form of AGI what would they say? The goal posts shift too fast to keep up with. Some people still worried about it counting Rs in strawberry as the models are making progress in frontier math and physics.
If Opus could autonomously create Fable, and Fable made Mythos, and so on, it would be AGI. The fact that people need to be involved, means it isn't.
If you gave Mythos infinite energy and compute, it still would not autonomously create the next coherent model.
It's a good bar, or at least I don't see your counterargument. The resources to make Mythos were obviously available, but the previous gen models could not manage to build it on their own regardless. It's the bar we have set (solving increasingly complex novel problems in perpetuity), and the original bar for AGI.
https://www.scientificamerican.com/article/what-does-artific...
> I think most people are vastly underestimating just how far this technology can go
It's easy to say this when you don't define "this technology", isn't it? We can easily point to IBM Watson and say this is the same technology, or point at a transistor and say we'd be nowhere without those so we "vastly underestimated" them…
Now, if we're talking about just how far sophistry can go, there does appear to be no limit, and apparently we can scale it.
It really feels like a lack of imagination to think that this technology that surprises even the researchers working on it all the time will just stop improving in the fashion skeptics keep saying it will. I've been hearing it's "hit a wall" on every update since GPT 3.5 and the rate of improvement just keeps accelerating. Every single time someone says it'll never be able to do X (math was the big one I heard a lot - how can a language model do math?) it tends to do X not only well but better than us in time.
Sophistry goes both ways. My guess is if you're a skeptic and you're able to go back and see your thoughts from 2022 they were much different arguments than what they are now. Arguing that LLMs aren't going to be more intelligent than humans because of what they can do now could be like arguing the industrial revolution will stop at the steam engine. How can we possibly know either way?
My personal bet comes from the rate of change I'm seeing and how it's making many of my friends (even the high paid ones, in fact especially the high paid ones) obsolete overnight.
They'd say it obviously wasn't. If you went back before the marketers started moving goalposts most people would laugh in your face for even calling an LLM AI because they know what AI is. They've seen what AI is in 2001 A Space Odyssey, I Robot, The Terminator, Short Circuit, etc. and they'd very rightly see LLMs as being a fancier ELIZA (https://en.wikipedia.org/wiki/ELIZA) and not even remotely close to Data from star trek.
“Twenty years ago, I think people would have happily agreed that systems with the ability of GPT-4 or (Google's) Gemini had achieved general intelligence comparable to that of humans,” Hinton said. “Being able to answer more or less any question in a sensible way would have passed the test. But now that AI can do that, people want to change the test.” - Geoffrey Hinton[1]
[1] https://www.clickorlando.com/business/2024/04/04/tech-compan...
AGI a couple decades ago would be the T-800 from the terminator, Data from star trek, and Johnny 5 from short circuit. What we have now isn't even remotely close. Most people wouldn't even call what he have now AI by the standards of the time.