Comment by Closi
1 month ago
Just comes down to your own view of what AGI is, as it's not particularly well defined.
While a bit 'time-machiney' - I think if you took an LLM of today and showed it to someone 20 years ago, most people would probably say AGI has been achieved. If someone wrote a definition of AGI 20 years ago, we would probably have met that.
We have certainly blasted past some science-fiction examples of AI like Agnes from The Twilight Zone, which 20 years ago looked a bit silly, and now looks like a remarkable prediction of LLMs.
By todays definition of AGI we haven't met it yet, but eventually it comes down to 'I know it if I see it' - the problem with this definition is that it is polluted by what people have already seen.
> most people would probably say AGI has been achieved
Most people who took a look at a carefully crafted demo. I.e. the CEOs who keep pouring money down this hole.
If you actually use it you'll realize it's a tool, and not a particularly dependable tool unless you want to code what amounts to the React tutorial.
I built a Nostr web client without looking at code or touching the IDE with Gemini CLI: https://github.com/lucianmarin/subnostr
So it had a tutorial for that api and it reimplemented it
Depending on the task, the tool can, in effect, demonstrate more intelligence than most people.
We've just become accustomed to it now, and tend to focus more on the flaws than the progress.
> If someone wrote a definition of AGI 20 years ago, we would probably have met that.
No, as long as people can do work that a robot cannot do, we don't have AGI. That was always, if not the definition, at least implied by the definition.
I don't know why the meme of AGI being not well defined has had such success over the past few years.
"Someone" literally did that (+/- 2 years): https://link.springer.com/book/10.1007/978-3-540-68677-4
I think it was supposed to be a more useful term than the earlier and more common "Strong AI". With regards to strong AI, there was a widely accepted definition - i.e. passing the Turing Test - and we are way past that point already: ( see https://arxiv.org/pdf/2503.23674 )
I have to challenge the paper authors' understanding of the Turing test. For an AI system to pass the Turing test its output needs to be indistinguishable from a human's. In other words, the rate of picking the AI system as human should be equal to the rate of picking the human. If in an experiment the AI system is picked at a rate higher than 50% it does not pass the Turing test (as the authors seem to believe) because another human can use this knowledge to conclude that the system being picked is not really human.
Also, I would go one step further and claim that to pass the Turing test an AI system should be indistinguishable from a human when judged by people trained in making such a distinction. I doubt that they used such people in the experiment.
I doubt that any AI system available today, or in the foreseeable future, can pass the test as I qualify it above.
1 reply →
Completely disagree - Your definition (in my opinion) is more aligned to the concept of Artificial Super Intelligence.
Surely the 'General Intelligence' definition has to be consistent between 'Artificial General Intelligence' and 'Human General Intelligence', and humans can be generally intelligent even if they can't solve calculus equations or protein folding problems. My definition of general intelligence is much lower than most - I think a dog is probably generally intelligent, although obviously in a different way (dogs are obviously better at learning how to run and catch a ball, and worse at programming python).
I do consider dogs to have "general intelligence" however despite that I have always (my entire life) considered AGI to imply human level intelligence. Not better, not worse, just human level.
It gets worse though. While one could claim that scoring equivalently on some benchmark indicates performance at the same level - and I'd likely agree - that's not what I take AGI to mean. Rather I take it to mean "equivalent to a human" so if it utterly fails at something we're good at such as driving a car through a construction zone during rush hour then I don't consider it to have met the bar of AGI even if it meets or exceeds us at other unrelated tasks. You have to be at least as general as a stock human to qualify as AGI in my books.
Now I may be but a single datapoint but I think there are a lot of people out there who feel similarly. You can see this a lot in popular culture with AGI (or often AI) being used to refer to autonomous humanoid robots portrayed as operating at or above a human level.
Related to all that, since you mention protein folding. I consider that to be a form of super intelligence as it is more or less inconceivable that an unaided human would ever be able to accomplish such a feat. So I consider alphafold to be both super intelligent and decidedly _not_ AGI. Make of that what you will.
6 replies →
I’ve got to disagree with this. All past pop-culture AI was sentient and self-motivated, it was human like in that it had it’s own goals and autonomy.
Current AI is a transcript generator. It can do smart stuff but it has no goals, it just responds with text when you prompt it. It feels like magic, even compared to 4-5 years ago, but it doesn’t feel like what was classically understood as AI, certainly by the public.
Somewhere marketers changed AGI to mean “does predefined tasks with human level accuracy” or the like. This is more like the definition of a good function approximator (how appropriate) instead of what people think (or thought) about when considering intelligence.
The thing that blows my mind about language models isn't that they do what they do, it's that it's indistinguishable from what we do. We are a black box; nobody knows how we do what we do, or if we even do what we do because of a decision we made. But the funny thing is: if I can perfectly replicate a black box then you cannot say that what I'm doing isn't exactly what the black box is doing as well.
We can't measure goals, autonomy, or consciousness. We don't even have an objective measure of intelligence. Instead, since you probably look like me I think it's polite to assume you're conscious…that's about it. There’s literally no other measure. I mean, if I wanted to be a jerk, I could ask if you're conscious, but whether you say yes or no is proof enough that you are. If I'm curious about intelligence I can come up with a few dozen questions, out of a possible infinite number, and if you get those right I'll call you intelligent too. But if you get them wrong… well, I'll just give you a different set of questions; maybe accounting is more your thing than physics.
So, do you just respond with text when you’re promoted with input from your eyes or ears? You’ll instinctively say “No, I’m conscious and make my own decisions”, but that’s just a sequence of tokens with a high probability in response to that question.
Do you actually have goals, or did the system prompt of life tell you that in your culture, at this point in time, you should strive to achieve goals[] because that’s what gets positive feedback?
Your argument makes no sense
2 replies →
> All past pop-culture AI was sentient and self-motivated, it was human like in that it had it’s own goals and autonomy.
I have to strongly disagree with you here. This was absolutely not the case in a very large amount of science fiction media, particular in the 20th century. AIs / robots were often depicted of automatons with no self-agency, no goal setting of their own, who were usually capable of understanding and following complex orders issued in natural language (but which frequently misunderstood orders in ways humans find surprising, leading to a source of conflict.)
Almost all of Asimov's robots are like this, there are a handful of counter examples, but for the most part his robots are p-zombies that mis-follow orders.
Nonhsentient AI with no personal motivation also frequently comes up in situations where the machine is built to be an impartial judge, for instance in The Demolished Man, all criminal prosecutions need to persuade a computer which does nothing but evaluate evidence and issue judgments.
Non-sentient AIs also show up often in ship-board computers. Examples are Mother in Alien, and the Computer in at least most of Star Trek (I'm no Trekkie, so forgive me for missing counter examples and nuance, technology in that show does whatever the writers needed.)
Even the droids in Star Wars, do they ever really execute agency over their own lives? They have no apparent life goals or plans, they're along for the ride, appliances with superficial personalities.
In The Hitchhiker's Guide to the Galaxy, does Deep Thought actually have self-agency? I only recall it thinking hard about the questions posed to it, and giving nonsensical answers which miss the obvious intent of the question, causing more trouble than any of it was worth.
Ghost in the Shell; obviously has sentient AIs, but in that setting these are novel and surprising, most androids in that are presumed to be just machines with dumb programming and it's only the unexpected emergence of more complicated systems that prompt the philosophizing.
I think we’re looking at the same thing in different ways. But regardless I don’t know think a valid interpretation of classical how AI was classically depicted is as a transcript generator or an extension thereof. There’s still some notion of taking action on its own (even if it’s according to a rigid set of principles and literal interpretation of a request like an Asimov robot) that is not present in LLMs and cannot be.
> Current AI is a transcript generator. It can do smart stuff but it has no goals
That's probably not because of an inherent lack of capability, but because the companies that run AI products don't want to run autonomous intelligent systems like that
Charles Stross published Accelerando in 2005.
The book is a collection of nine short stories telling the tale of three generations of a family before, during, and after a technological singularity.