This is a perfect example of the 'AI effect' in action:
>"It's part of the history of the field of artificial intelligence that every time somebody figured out how to make a computer do something—play good checkers, solve simple but relatively informal problems—there was a chorus of critics to say, 'that's not thinking'." AIS researcher Rodney Brooks complains: "Every time we figure out a piece of it, it stops being magical; we say, 'Oh, that's just a computation.'"
We really need another name, which refers to people uninformedly quoting 'the AI effect'.
How about "the AI effect effect"?
This happens all the time:
- X suggests that success at a particular task is a good proxy to general intelligence.
- Years later Y solves that task with a clever hack and lots of computation, but no general intelligence.
- The media breathlessly suggests we've made progress on general intelligence.
- Z points out we haven't really.
- Someone says "Oh that's the AI effect!", by which they mean the goalposts have been unfairly moved on AI.
No, they have not. There's something called general intelligence, humans can do it, and Deep Blue cannot, and it's fine to say that Deep Blue beating humans at Chess was not thinking.
It just turned out that Chess wasn't as good a proxy to thinking as we initially thought, and mentioning "The AI effect" just clouds the discussion.
> Years later Y solves that task with a clever hack and lots of computation, but no general intelligence.
While that is true the set that contains instances of Y is only growing and the members are each becoming less narrow.
Deepblue was only chess with many handcrafted parts. AlphaZero does chess, shogi and go with no prior knowledge. Then they went from perfect information, turn-based games to RTS with fog of war.
It's not general intelligence, but at least further down the road in behavioral complexity than jellyfish, despite some crippling limitations (e.g. separation of training and inference)
Well, I think you do have a point, but on the other hand, there have been people saying "Machines will never play chess well because it will require general intelligence!"
And when machines start to play chess (or go, or Starcraft) they just say "Oh well it's just computation."
So I think it's justified to bring up "AI effect" when people say "machines will never do $(insert your favorite activity here) because it requires general intelligence!"
Since the beginning of time, people have anthropomorphized any and every thing that is complex enough to be beyond a person's understanding. Being equivalent to a human mind is the default assumption for anything. If it was a wrong assumption when applied to storms, volcanoes, Babbage's machine, Clever Hans, ELIZA, and a billion other things, then why would you assign any weight to most claims that something requires intelligence?
It makes me think of the saying about atheism, that an atheism disbelieves just one more god out of thousands than a (mono)theist. You disbelieve in intelligence in very nearly all of the things it's been used to explain.
It's also possible that an autistic chess-winning machine is not actually reproducing the full range of behaviors humans mean by the phrase "playing chess".
> There's something called general intelligence, humans can do it, and Deep Blue cannot
There's no such thing. General is a strong word. Humans can't for example handle more than 7 objects in working memory. We have an inbuilt limitation to how much complexity we can grasp. Programmers know what I mean.
General intelligence requires a general (infinitely complex and challenging) environment. Without it it will always be a specialised intelligence. Human intelligence is specialised in human survival, as individuals and part of society.
As you appeal to Programmers: do you believe quicksort is a general sorting algorithm?
Would it still be a general sorting algorithm if someone had just developed it and used it solely to sort numbers between 1 and 100?
It would be.
Why do you think you need to a general environment to develop a general algorithm?
No one is arguing human rationality is unbounded. But our intelligence generalizes, in a practical sense, to a great many more tasks than any computer algorithm I'm aware of.
Mainstream psychologists believe in general intelligence. They even attempt to measure components of it, with good prediction of performance on unseen tasks.
I think there's a big burden if you want to argue it doesn't exist.
> Why do you think you need to a general environment to develop a general algorithm?
Because intelligence is not intrinsic to the agent but the result of the agent trying to maximise rewards in an environment. In the end the driving force is survival - the agent needs only to survive, any method would do. So as soon as it overcomes a challenge it stops evolving and turns to exploiting. And the list of challenges that threaten survival does not scale to infinity. Anytime you think about intelligence you need to also think about the environment and the task otherwise it is meaningless.
You state this as if it's fact but it's a point of view, not necessarily wrong but not proven either. More to the point, it's circular: "intelligence is not intrinsic to the agent but the result of the agent trying to maximise rewards in an environment" is a statement that the human mind operates on the same principles as ML algorithms. If you assume this, then not surprisingly it follows that ML algorithms can in principle do anything the human mind can do. Not everybody agrees and the question has not been empirically settled.
Well, humans can understand any utterance certainly in their native language and there are an infinite number of those, given that natural language, as far as we can tell, is infinite (you can genereate utterances for ever without ever generating the same utterance twice). That is as general as anything gets.
Also, we may not be able to keep some number N of objects in memory simultaneously, but we can definitely reason about an infinite number of objects, like I did above. And if you want to handle more than N objects, you just do it N objects at a time. You can always write things down etc.
I'll need an example of a process we can't explain by cutting it up in chunks though. DNA and the stock market- well, maybe we don't "understand" them as such, but that's not to say we can't, ever.
I think it would be very difficult to find a process that _cannot_ be understood by humans, let alon explain why. It'd be a bit of a paradox really. "Let me explain to you this thing that is impossible to understand".
However stating "can't handle more than 7 objects in working memory" as fact seems a little too assertive to me :)
Humans can do it, but are usually not very good at it as the number of items goes up.
That is a good point about something I hadn't considered. I was going to disagree with you until the last sentence, which then clicked.
As AI is developed, it is also championed as "closer to real thought" and might someday solve the more general intelligence problem, only for those same experts to later realize that would never be the case and then retroactively proclaim "this was never going to solve general intelligence, but it is still AI".
The history is full of this kind of revisionism, so I agree with you.
I agree with you, but just want to clarify for others who don't understand why the parent misused the term AI effect. AI effect is more related to the fact that once you understand the trick behind the magic you no longer believe the task to require intelligence, even though it is still able to do an amazing task that wasn't possible before and had for centuries been thought off as requiring intelligence and something that only humans could do.
That the current AI techniques couldn't work for every imaginable task, and that people believed they might isn't a case of the AI effect. That's just hype, and possibly another coming case of the AI freeze.
The AI effect will be how in 5 or 10 years, CS students will learn neural nets as part of their first or second year curriculum and the whole machinery will be behind a single API to some Apache library. And someone will then use that to some cool effect and say they're doing AI, and others will laugh and say, you're not doing AI, you just used a neural net to learn weights from a big data set silly, that's not AI.
This effect happens for every task, even when AI is not involved. Until you understand how, someone doing something you can't comprehend will have you think they are a genius and maybe of a higher IQ and probably very intelligent. But learn the "how" for yourself and it'll stop being so impressive. Rubik's cube is a good example of this. Once you realize there's a trick to it, it stops being as impressive.
Some relevant quotes from the wikipedia article:
> AI effect is: As soon as AI successfully solves a problem, the problem is no longer a part of AI.
> Software and algorithms developed by AI researchers are now integrated into many applications throughout the world, without really being called AI.
> AI advances are not trumpeted as artificial intelligence so much these days, but are often seen as advances in some other field
> practical AI successes, computational programs that actually achieved intelligent behavior, were soon assimilated into whatever application domain they were found to be useful in, and became silent partners alongside other problem-solving approaches, which left AI researchers to deal only with the "failures", the tough nuts that couldn't yet be cracked
> The great practical benefits of AI applications and even the existence of AI in many software products go largely unnoticed by many despite the already widespread use of AI techniques in software. This is the AI effect. Many marketing people don't use the term 'artificial intelligence' even when their company's products rely on some AI techniques. Why not?
No, it talks about the other AI effect, where AI researchers thinks that generalizing the model will be relatively easy. Like, lets say someone makes an AI to play tic-tac-toe at the same level as the best human and then says that with just a little bit more work it will be able to beat a grandmaster at Go.
Right. Someone linked to "Artificial Intelligence Meets Natural Stupidity" last week, which makes that argument in detail. People in the field have been claiming Strong AI Real Soon Now since the General Problem Solver of the 1960s, which is a simple tree search algorithm. I heard a lot of that at Stanford in the 1980s, when the expert systems boom was about to crash and people thought Symbolics LISP machines were magical.
There is progress, though. Machine learning does do a lot. And it makes money, so effort will continue at a high level. AI used to be a dinky field - maybe 20-50 people at MIT, CMU, and Stanford. Now it's huge.
My point was that machines mastered tic-tac-toe 50 years ago and mastered Go just last year. Making a statistical model to predict a small subset of three body problems isn't very impressive at all, kinda like tic-tac-toe. It is a bit interesting that it works, but not much more than that. Adding another body, or adding starting velocities to these three bodies, would make the problem a lot less tractable, so it is very unlikely the same techniques will work there.
I think the root the issue is that "AI" is a vague term where almost any kind of program, no none at all could go under. The fact that a machine can automatically calculate the sum of two numbers could be considered artificially intelligent by someone who's never seen a computer before. At the same time, no matter how far you go, any program (based on turing-like machines) will always be a deterministic chain of cause-and-effect, which is "just a computation".
But AI in its meaning use contains a certain set of fields in a way that makes sense. It's just the puplic misinterpration nonsense that makes it look so weird.
Are computers meaningfully different to brains in that “deterministic” respect? Both are apparently based on deterministic interactions at the nanometre scale and exhibit simple behaviour that can be explained from first principles (eg. simple arithmetic, or a stroke). Both also exhibit complex emergent behaviour that can’t be explained from first principles (eg. the output of complex ML systems, or most human behaviour), which we sometimes call “non-deterministic.”
I’m happy to concede that software is artificially intelligent just as I’m prepared to concede that most animals are naturally intelligent. It’s just a question of drawing an arbitrary line when things are intelligent enough to warrant the description. The AI effect is the process of that line getting moved closer and closer to “capable of doing everything an average human can do.”
The definition of a modern human is someone who does what machines do not do, so as machines become more capable it changes what humans do.
I guess it seems very likely to me that we will have Skynet/paperclip maximizers before AGI. I mean, I think we kind of do right now in what the internet is growing into.
We're like single cells imagining what a brain would be like but we'll never know because probably we'll be part of something brainless that dominates the ecosystem, or if there is an organism with a brain, we won't be able to perceive it because we're components.
Of course, but that thing is getting more and more computerized and the software is getting more sophisticated. My point is just that if there is a phase change, we probably won't notice because we're part of it and don't perceive it.
The concept of AI to the average person outside of the computer bubble is "human intelligence made artificial"
So if it's not actually thinking, then technically, from the perspective of an average layperson, it is not real AI. However, the "AI effect" is still real, but it's related to a problem local to academics and AI laypeople.
Most non-AI-laypeople/AI academics believe AI is a machine that can think like how a person thinks.
That link honestly reads like more kool aid and deflection of criticism written by AI professionals. Why aren’t mathematicians complaining about their work being largely unknown?
>"It's part of the history of the field of artificial intelligence that every time somebody figured out how to make a computer do something—play good checkers, solve simple but relatively informal problems—there was a chorus of critics to say, 'that's not thinking'." AIS researcher Rodney Brooks complains: "Every time we figure out a piece of it, it stops being magical; we say, 'Oh, that's just a computation.'"
https://en.wikipedia.org/wiki/AI_effect