But I will say: by the time anyone has “something to show for it,” AGI will, pretty much by definition, already be here. I don’t think we’re gonna be able to see it coming, or know what it will look like, except through very speculative predictions. So I do think while skepticism is warranted, we should still evaluate everything in good faith, and not automatically jump to shilling. (Not saying you’re doing that, btw - just commenting)
The issue with AGI as it's often framed is that it implicitly assumes a general intelligence like us, without any metric that can relate human intelligence and machine intelligence. We have machines that beat us in Go, so if we took those machines and somehow combined them with machines that beat us in Chess, Starcraft and Dota, are we any closer to AGI than we were previously? What about if that same AI can also drive a car? What's the standard, and where does it end? Where do we fall on that same metric?
There are also some fairly deep philosophical questions in this realm, too. For example, to what extent is the human experience (and intelligence) linked to our physical bodies? Cooking by smell and writing by sound are things we do on a daily basis that have no machine analogue. How important is the embodiment principle to AI? Certainly we don't do things by plugging our minds in directly -- we drive cars using the same limbs that we use for everything else, which probably has enormous efficiency advantages. To what extent can you separate the intelligence from the body and the environment?
Frankly, I believe we're a lot further away than we think. The last decade has taught us that relatively simple methods can be applied in surprisingly powerful ways, which is an important start, but it doesn't tell us anything about how close we actually are to a given goal, and how we might go about reaching it.
literally anything
the intelligence of a wasp; the intelligence of a cat; the intelligence of a Downs syndrome person; the intelligence of a 3 month old baby
AI is very far from all of the above
We keep moving the goal post; a common theme was when I studied ML in uni in the AI winter of the 90s, that beating Go surely would mean human level intelligence. And many ML models we see now would be considered human level a few decades ago; we moved our goalposts and definitions, which is fine.
However ‘resembling’ is vague; I find most stuff on social media (most notably tik tok and instagram) not human level intelligence either, or, reverse, people easily would (and do) believe comments and posts done by something like gpt 3 are done by humans. That is because the level is so low of course. I know gpt 3 is not intelligent (for my definition anyway, which is intentionally vague) but resembling in some cases: definitely.
The goalposts haven't moved. If AI researchers thought that general intelligence would be required to beat a human in Go, then they were simply wrong.
AlphaGo is not a general-purpose intelligence. It only does one thing. It plays Go.
I think it's fair to say that the goalposts are moving (as they should).
Sure, point is that every definition we come up with gets beaten and then we were wrong with that definition in the first place; defining what agi means seems hard.
https://en.wikipedia.org/wiki/Moravec's_paradox
One of my favourite demonstration of that was the 2015 DARPA robot challenge, where one robot after the other failed at such difficult tasks as walking and opening a door.
Dr. Marcus Hutter's AIXI is a solid mathematical treatise that reduces intelligence to the concept of information compression, and you can exhaustively construct a logical extrapolation from aixi to any particular feature of intelligence at higher levels, but it's similar to string theory in that it's all-encompassing in scope. It's not useful in narrowing the solution space if you want to build a high level intelligent system.
While I do not think that is intelligence, it does making testing for it slightly broken. It will come down to ‘I will know it when I see it’ by elitists for exactly the above.
So in short ; I do think the Chinese room argument against it is quite a good one where even the elitist people can fall for if the ‘con’ is elaborate enough (as in; if gpt gets more data, more efficient learning and learns to know when it cannot answer and has to look it up, like calculations). Or maybe then it is intelligent? (To be clear; I don’t think so; I think we need a better tests and definitions).
I think the sad state is that the Turing Test is too diluted to be useful as a marker...
An interesting would be to do the Turing test with a timer; the interrogator gets 30s for each session and has to say who is what within 30s and then continue to the next batch. I think computers would go very very far if you do that. But that is the attention span consider normal currently for human/human social media interaction; often even less.
With the benefit of hindsight, this just seems like a bad goalpost. Go is hard for computers, but why did anyone think the smarts required to beat a human at it would transfer to understanding or generating text, or anything else a human can do?
What was "solved" with AlphaGo was using deep learning machine learning which are effectively black boxes. There was a certain assumption in the question for AI researchers academically that it would be an understood algorithm as an AI agent like a Prolog application, not a brute forced model. That's still not the case that we have a "solved" strategy and all we can do is watch it play as if it is a deaf mute player.
So there still is no "tic-tac-toe" known winning strategy to Go or anything.
That doesn't make AlphaGo any less impressive or any less practical, but it even has its own readout issues. It can't even read ladders without hard coding it in, for instance, because it becomes a long enough depth search. This is one of the first things a newbie would learn.
It's just a 19x19 board, so it was always known if you could read all the possible outcomes you could see all the possibilities and win. This is just looking at all possible outcomes and picking the best one, not knowing how to play. Creating models of data that is 2, 3, or even 4+ dimensions is always possible, just depends on how much computing power you can throw at it. The created models are essentially aggregate simplifications to play quicker.
Generalized intelligence is so much different. You have to define the problems themselves that you are trying to solve, figure out what the variables are, and solve it. Then you have to operate and run the machinery to create those experiments. Outside of a scenario that you've taken actual physical territory as an intelligence, I can't see how it would get there (think Terminator or BSG, doesn't have to be malicious but they'd have to be in control of the physical area autonomously).
But the hardest part is defining the problems independently given the sheer number of problems they'd need to define second to second just to solve basic tasks, and they'd likely have millions of variables with millions of possible values.
> It's just a 19x19 board, so it was always known if you could read all the possible outcomes you could see all the possibilities and win.
All possible outcomes is not something you can iterate over in our universe.
> There was a certain assumption in the question for AI researchers academically that it would be an understood algorithm as an AI agent like a Prolog application, not a brute forced model.
I never heard this, and the result is really not brute forced. You can't brute force go.
> That doesn't make AlphaGo any less impressive or any less practical, but it even has its own readout issues. It can't even read ladders without hard coding it in, for instance, because it becomes a long enough depth search. This is one of the first things a newbie would learn.
Only in early versions, AlphaZero didn't have any built in knowledge and can learn different games and the later developments in MuZero went further to make it more generalised as a learner.
Removing the hard coded logic and removing even seeing how humans plan, it got better. It found strategies and ways of playing that experts had missed in an ancient game.
> This is just looking at all possible outcomes and picking the best one, not knowing how to play.
"It's just doing X, it doesn't really know how to Y" is a common refrain. It looks at options, and explores "what if" scenarios in a guided sense with a feeling about how good any particular potential board is. I find it hard to say that it doesn't "know" how to play.
It's more like climbing a mountain and seeing higher peaks. Researchers may have been thinking that creating a human beatable program may get them insights into how to create an AGI, but they just find out more problems to surmount.
To give you an example of how lost we are, we don't even know how human genetics work with the human brain. It is one thing to make a machine that can pretend to think, but to make a machine that not only actually thinks, but also communicates and is able to act in a 'human' or 'intelligent' fashion is exponentially harder.
Just my 2 cents. I hope to be proven wrong, but we haven't cured cancer or discovered the secret to fusion yet, so...
Building planes that fly didn't require us to understand how birds or insects fly.
Getting directions from Google Maps didn't require them to figure out how hamsters navigate a maze.
Now, if you want to build a computer that passes the Turing test, perhaps you need to understand how humans work. Maybe? But it's not clear that this knowledge is necessary to build something smart enough to drown the universe in paperclips.
The latter reminds me of Edsger Dijkstra's aphorism: "The question of whether machines can think is about as relevant as the question of whether submarines can swim."
(Now, it might turn out that we need to understand how humans tick and how genetics interact with the brain in order to build a successful paperclip optimizer. Probably not, but it might turn out that way.
I am just saying that this would be a surprising empirical fact to learn. Not something that we can just assume based on armchair reasoning from analogy.)
Now to get slightly off-topic:
> I hope to be proven wrong, but we haven't cured cancer or discovered the secret to fusion yet, so...
Oh, we can totally build fusion reactors right now!
First, a fusor is a bench-top nuclear fusion device. The main downside is that no one has figured out how to get more useful energy out of it then we put in. So probably not what you had in mind.
See https://en.wikipedia.org/wiki/Fusor
Second, we can build a nuclear fusion device that does generate useful energy:
You take a huge tank of water, some steam turbines, and a supply of fusion bombs.
You take one of the bombs, explode them in the water, and use the turbines to generate electricity. Repeat as needed.
It's a very simple system, and we had the means to make this work since the 1950s. Of course, it's also a completely ridiculous design that approximately no-one would want to use in practice. Especially when you already have more conventional nuclear fission reactors.
But something along very similar lines was seriously considered for spaceship propulsion. See https://en.wikipedia.org/wiki/Project_Orion_(nuclear_propuls...
> Of course, it's also a completely ridiculous design that approximately no-one would want to use in practice. Especially when you already have more conventional nuclear fission reactors.
Right, and that's the point.
There are three logical leaps here, and proof is missing for all three.
- One, that AGI is something we can replicate in a Turing Machine.
AGI might require a specific effect in quantum mechanics to work, for example. Light refraction is completely understood, but still extremely difficult to solve for a single case -- computing power is getting there, but it took about 60 years from it being understood to us being able to compute it reasonably, and even then it's only an approximation -- our best rendering farms about around 5 F-stops. That's nowhere near the human eye's ability.
Another example, the 3-Body problem solved for N-bodies. Or how about, protein folding. Folding@Home is remarkable, but even with the combined GPUs of hundreds of volunteers, it still takes Folding@Home months, years, to calculate the folds of a single protein. The brain has billions of them.
- Two, that we will figure out the solution to Turning Machine simulation in our lifetime.
This can stand on the arguments of one. I'd just like to add that there are many really simple conjectures that are as-yet unsolved within mathematics. The Collatz conjecture is a good example here, but there are hundreds of thousands of others. Despite it being probably the simplest problem to teach, it's been about 90 years and even proof theory machines haven't made too much headway. Erdos, probably the greatest mathematician to ever have lived, stated "Mathematics may not be ready for such a problem". In that, subject experts doubt our ability to solve it within our lifetime.
Why should this be any different for an exponentially harder problem?
Perhaps a reference to the fact that it took about 150 years to go between "neurons are in the brain and influence our actions" to "we can do brain surgery, kind of...". And probably about 200 more years until we can say we actually "understand" the brain in a meaningful way?
- Three. That the resulting device will be practical, useful, and will be able to understand itself enough to replicate a better version of itself, within a reasonable timeframe.
I think you argued this sufficiently yourself! The fact that we can build something doesn't give it use. We can build machines that extract energy off of graphene. At no point does it make it useful to do so, however.
And yet, despite all of these unsolved problems, we are supposed to be able to simulate a machine to replicate the brain? In the next 20 years, no less?! Is it not unreasonable to think that this is a preposterous assertion whatever dimension you look at it through? That so many people have thrown themselves into wishful thinking, and cults at that, is beyond me.
First, what makes you think that solving the Collatz conjecture is easier than AGI? Doesn't that beg the question? Just because a challenge is easy to state doesn't mean that we should expect it to be easy to answer. Fermat's Last Theorem was easy to state; open a random math textbook for many theorems that are harder to state, but easier to prove.
Second, quantum mechanics is totally amenable to simulation on a Turing machine. In fact, all of known physics is. You rightly point out that the question is rather how good modern hardware and software is at the task. (Btw, quantum computers would be really good at simulating quantum mechanical systems. And that would probably be their main use; right now we don't really know much else they are better at than classical computers.)
Third, you bring up the point that figuring out how arbitrary proteins fold is still rather difficult. I agree. So an atom for atom simulation of the human brain would presumably be also rather difficult.
That doesn't mean that AGI is impossible. It only means that AGI via atom-for-atom simulation of a human brain would be rather difficult.
Simulating a human brain atom-for-atom is only one approach you can try to take to reach AGI. There are other approaches people are working on. Only one of them has to work.
So let's be honest here, in the space of all possible outcomes -- the negative outcomes where "it doesn't work" or "it works but is too slow to use" or "it doesn't understand how it works either", is a larger order of infinity than the outcomes where it does work, at least based on what has happened with pretty much every other technology that we have records of predictions for (Is there a technology that we have predicted where it wasn't wildly different or handicapped compared to our dreams? I'd honestly like to know TBH).
There is literally no reason for pontificating more over this. All of this, the best guesses, the worst guesses, is all just a variety of wishful thinking. The only true answer is we have no fucking idea.
Then there are those who think AGI is impossible in principle.
The people I get most frustrated with are those that use the first argument - we're nowhere close to implementing an AGI - as an argument that therefore AGI is impossible. Sometimes they don't even realise or recognise that this is what they are doing.
The argument is often that "we do X and computers don't do X therefore computers can't be like us in that way". Well no computers don't do those things yet because we haven't developed AGI yet. That's not a reasonable argument that developing AGI is impossible.
Yes. I don't believe that it is impossible in principle either, but it is neither trivial to argue that it is possible. An argument that annoys me to no end is this "well, AI continues to get better (there is no regression in our capability), and therefore it must sooner or later exceed any given threshold, for example human level AGI. QED." That argument is just plain wrong. The world record in the 100-metre dash is only ever improving, but that doesn't mean that humans will soon run Mach 1.
Well of course they have no idea how we think. They're making no actual effort to study how we think. That's for cognitive scientists and neuroscientists, and AI people often sort of dismiss it as having too long a time-horizon or overly high compute requirements before it's applicable to "real world" problems.
So there's plenty of scientists and engineers who are trying to understand how the brain works and how to combine that with eg deep learning.
That attitude seems entirely reasonable to me.
> I just reviewed a paper yesterday that was powerful and general, but wasn't really trying to hit SOTA on any one test task its architecture could perform. It was trying for generality. This likely got it dinged by the other reviewers.
That seems like an interesting paper!
It's perfectly reasonable to see that and still be confident AGI will happen - just not with current models.