The real issue we are facing is that everything that we thought was not going to be pattern matching and tree search has turned out to be pattern matching and tree search. I remember my father telling me computers were never going to be able to play Chess, because it required creativity for example. Nowadays a neural network with tree search plays chess that looks remarkably human. A lot of problem domains have fallen to what is basically pattern match and tree search.
Extrapolating the trend of the last 30 years, there is evidence that computers will be able to solve every task a human can using pattern matching. If that isn't AGI, it might turn out to be better than intelligence.
The technological future is unknowable, so believing AGI is certain is too much. But believing it certainly isn't around the corner is also too little. If computers can do anything a human can intellectually, they have reached AGI. The list of discrete tasks (games, decision making once the parameters are defined) a computer can't do is a very short list.
If someone finds an objective function for deciding what decision parameters are important AGI could be upon us very quickly. As a postcript, I think people radically overestimate human intelligence.
AGI is Data in Star Trek TNG - trying to be human, making decisions to want to be alive, eventually dreaming and finally using an emotion chip. Another alternative here would be Moriarty or the various doctors in Voyager.
AI is the Ship in TNG - lots of heuristics to figure out what the user is trying to do. Past usage of commands and relating major events outside the ship with algorithms for battle, life support, etc.. Events categorized by importance and automatic handling to save lives when necessary. Basically an extremely advanced Siri that doesn't really misunderstand you - while at the same time not really caring about you or knowing anything about being alive other than the priorities built into its software.
I think for the next 100 years we're going to have AI progressing like the ship in TNG. I don't think we'll have AGI until maybe 100-200 years.
Then again when I was born no one had a fucking clue eventually we would have something like the iPhone and talk to someone in China with <1 sec lag. So my estimates could easily drop to half.
It's 5 years old now (coincidentally the time span quoted to develop a solution), but recognizing a bird was already considered a solved problem 3 years ago, less than 2 years after the publication of the cartoon.
Surprisingly, recognizing a bird is much harder when you count rare species and running birds. Thus, sparse data. Does it recognize penguins too?
AIs today still fail at it. Some folks were trying to train one to match endangered species and they had to pull mighty tricks to have some 70% accuracy. I think it was here on HN some time ago, but can't recall a link.
AGI as discussed by people working on it is closer to the Ship than to Data. The defining thing about AGI is that can figure its way around arbitrary challenges just like humans do, but not necessarily the same way humans do. There's a concept called "orthogonality thesis" that tells you that intelligence and values are orthogonal. That is, there's no reason why a powerful enough AI would have to develop values similar to those of humans (like Data, trying to be humans) - you could have an AGI that's smarter than humans in everything, but is "not really caring about you or knowing anything about being alive other than the priorities built into its software".
> not really caring about you or knowing anything about being alive other than the priorities built into its software".
Is that really "general" intelligence, then? I would argue we don't know today whether consciousness and agency are separable from general intelligence. You could say that "general intelligence" implies some degree of intelligence on any topic, which would include self-reflection and metacognition.
It sounds to me like you're describing something like a p-zombie[0], which we don't know to be able to exist.
I believe it is. I'm also not describing a p-zombie, since I find the whole concept of p-zombies utter nonsense.
The general AI I described would have self-reflection, agency and arguably consciousness, yet - per orthogonality thesis - it may not have anything resembling human values.
Yeah, sure. The orthogonality thesis essentially implies than a GAI randomly plucked out of space of possible minds will likely be considered sociopathic by our standards. That is, if we can comprehend its thinking at all. "Not sociopath" is a very particular set of values.
> a GAI randomly plucked out of space of possible minds
Sorry, but I don't think you have any rational basis for imagining what the "space of possible minds" represents. The only minds with human-level intelligence we know of are human minds that have (with variation) human values.
The claim you're making is analogous to saying "any extraterrestrial life we find won't be carbon-based because out of the space of all possible substrates for life, carbon is a very particular one", but that's an ill-founded supposition because we have a sample of N=1 and maybe carbon-based life is the only kind of life there is.
Maybe a true GAI mind will be "like us", maybe it won't be, but we don't have anywhere near enough data to speak with confidence about it.
Although I will admit to have used/benefited from the myth that chess skill is related in some way to general human intelligence, it’s a myth.
Key differences between chess and real world: perfect information game, finite problem space, well defined rules. It blows my mind that serious people believe that ability to outperform human chess players using massive compute is some kind of major step towards AGI. If only it were that simple.
It’s not that people overestimate human intelligence, it’s that they underestimate the meta-cognitive reasoning that we call common sense.
I suppose it's a subset of "well defined rules", but it's worth calling out explicitly, I think: Chess also has an objective (and trivially verifiable) win condition.
There are few interesting situations in real life where such a thing exists.
In that regard, AlphaGo is a lot more impressive, not just because it is a vastly more complex problem to find a loss function for playing Go (compared to chess), but also because AlphaGo basically learned to play the game by itself without even having a model of the game rules initially (or so I've heard).
That said, I would still not consider it anything like generalized AI, if only because the set of possible (valid) actions at any point in the game is tiny, while in real-world problems it is basically infinite.
> perfect information game, finite problem space, well defined rules
Chess isn't reality, but the things you list are all mostly things that humans can't deal with either. Take imperfect information - humans can't make decisions using information they don't have any more than machines can. Humans certainly can't deal with the infinite (and their approximations to do so are probably measurably worse than those a computer uses, because a computer can use honest-to-goodness probability formulas).
Operating without rules is not clear cut, but most people do invent a whole heap of funny rules because they can't operate without clear rules either. Humans often literally hate and fear things that look different or don't follow all the funny rules they come up with.
These are the same arguments as are deployed against self driving cars - if a computer doesn't have the information needed to make a decision then neither will a human in the same situation.
The threat, opportunity and potential of AGI is very real. Once technology settles down and stops changing then we'll know that the situation has stabilised. But even as it stands what we have now would easily pass muster as AGI for the 1910s and it is still improving extremely rapidly.
I remember when I was a teenager: a moonless night in the woods and my bicycle light was broken. It was so dark that I literally couldn’t see my hand in front of me. I could drive (slowly) because I knew that part of the street by hard and I knew there is a gravel patch on each side of the road. So I every time I entered gravel from the left I just turned a little more to the right and vice versa. And I knew I entered gravel by ear.
There was also a faint red light from a memorial site candle that I could use as a orientation point.
The thing was that I have never done this before (or after), nor did I think I ever would ever find myself in such a situation. I drove through that very road often at night also often without light because I had shitty broken bicycles, but this night was truly exceptionally dark and the darkest night I ever had seen since.
The question is, what would an AI have done in a similar situation (sensors go dark for some reason)?
My fundamental problem with this viewpoint is that people appear to radically underestimate the amount of training data (all of it incredibly rich in context) that humans have had access to over the course of their lifetimes.
> The real issue we are facing is that everything that we thought was not going to be pattern matching and tree search has turned out to be pattern matching and tree search.
Absolutely false. It just seems that way because you read biased research and articles.
There's a boatload of problems that cannot be solved with pattern matching and tree search. Even really simple ones.
One example is estimating/predicting binomial proportions adequately.
>> The list of discrete tasks (games, decision making once
the parameters are defined) a computer can't do is a very
short list.
> I'd be interested in knowing what coungerarguments the
people downvoting this comment might know of, which I
apparently don't.
I didn't downvote, but I will cite artistic endeavours.
How long until a film crew of computers can shoot, edit and score a documentary or film that would be interesting to humans to watch?
How long until they could develop an AAA computer game worth playing?
How long until we could assemble an orchestra of computers that can interpret sheet music with feeling well enough to impress a human audience?
I could go on and on finding other examples of human endeavours that computers/robots/AI will suck at for an extremely long time, if not forever.
As we've seen recently, we are further from completely autonomous self driving cars than was hyped over the past few years.
In my experience, programmers typically like to underestimate the breadth of human endeavour outside the domain of programming, particularly where the arts are concerned. And larger groups of humans working on larger artistic endeavours will take even longer to be displaced by AI, IMHO.
I don't think that you need to limit creativity to the arts, which is unfair to machines because art really depends on human emotional quirks. That is even kind of the whole point of art.
What about technical inventions? Can AI invent, let's say, the process to produce aluminum? Or planar semiconductors? Or a rocket engine? These are also creative works.
> How long until we could assemble an orchestra of computers that can interpret sheet music with feeling well enough to impress a human audience?
I think AIs making music will come long before AAA games and movies, both of which encompass music as an art and then throw in like another 5 artistic pursuits on top.
I think AIs will be making music inside 10-20 years or less. Honestly I think it could happen in like 1-2 years if a company with lots of AI resources chose to focus on it.
Note that the parent comment wrote interpret music not make it.
I am a musician and there were already impressive algorithmic compositions in the 70s using Markov chains, and today we use machine learning to create something that resembles the works of J.S. Bach.
But how is that creativity? Creativity means coming up with new and interesting things (and fair enough many musicians make quite uncreative decisions all the time).
These statistcal improvisators can come up with some interesting combinations but they are completely unaware of them and don’t repeat it ever again.
IMO we didn’t even solve composition, because it also requires a good feel for how humans react to a piece and how a given piece or instrument fits culturally and what emotions it envokes for what reasons.
Interpretation is yet another thing, it means interacting with an audience in one way or another.
To think we managed to solve composition by scrambling together some melodies from an input of thousand melodies and thinking we are done with the hardest part is hubris at it’s best.
> To think we managed to solve composition by scrambling together some melodies from an input of thousand melodies and thinking we are done with the hardest part is hubris at it’s best.
To me, that's missing the point. The output of systems like BachBot and DeepBach is musically interesting precisely because of how weird and serendipitous the results of that "scrambling together" are. IOW, it's not just scrambling, but scrambling that manages to learn and preserve at least the short-term structures that we associate with "music". (That's a big improvement over simple Markov models.) It's nowhere near what humans would make, same as a picture of an actual dog is not similar to what the DeepDream network outputs as "dog-like" - but it's already interesting in its own right.
Creativity and purposeful manipulation of human emotions are orthogonal concerns, even if often bundled together under the term "creativity" in context of arts.
To disentangle them I propose a simple test: take these 'impressive algorithmic compositions that resemble the works of J.S. Bach' and play them to people, telling them they were composed by a gifted human. Then ask what they think of their creativity, and of emotions the author intended to communicate.
That's creativity. As for purposeful manipulation of human emotions, this is harder and would require an AI with a theory of (human) mind, or some equivalent of that. Doesn't sound insurmountable though.
> Creativity and purposeful manipulation of human emotions are orthogonal concerns, even if often bundled together under the term "creativity" in context of arts.
No, absolutely not, I disagree. If we know that a piece of music is generated by software algorithms, it instantly loses any real meaning.
> To disentangle them I propose a simple test: take these 'impressive algorithmic compositions that resemble the works of J.S. Bach' and play them to people, telling them they were composed by a gifted human. Then ask what they think of their creativity, and of emotions the author intended to communicate.
The problem here is you are lying to people in order to achieve an effect. Deception has been known to cause a significant backlash amongst music fans [1]
There are enough music fans out there that will honestly want to know whether the music was created by a human or a computer program. If it ever becomes common place that procedurally generated music is misrepresented as the works of a human or group of humans, then there is a large segment of the population that will turn their noses up, and only listen to live music, where it is apparent that the music is performed by humans.
The rest of the audience, meh, if they want to hear meaning in music composed by machines, that's their prerogative I suppose. Most people I know who actually play instruments or sing are appalled by the idea, or at best, slightly bemused.
I will assert, those who have never stepped away from their computer keyboard long enough to have taken a deep breath, stood in front of a microphone, plucked an electric guitar at high volume or smashed drum skins with sticks in front of a live audience of 100 to 1000 cheering people may just be incapable of understanding this. The feeling in the air can be electric.
Humans connecting with other humans through the creation and reception of music will never be correctly emulated or simulated through silicon, no matter how good the facsimile.
> No, absolutely not, I disagree. If we know that a piece of music is generated by software algorithms, it instantly loses any real meaning.
Which is kind of my point, so I feel we're at least partially in agreement.
The point I'm making is this: creativity in software is either easy to achieve, or near impossible, depending on what you really mean by the term "creativity".
If you have a piece of performance in front of you, and your judgement of whether or not it's "creative" changes when you learn whether the author was a human or a machine, then that "creativity" is impossible for machines by definition - and, frankly, it's also not worth talking about, because it does not depend on the author, but whether or not you consider the author human.
If, however, your opinion wouldn't change upon learning whether the author was human, that kind of creativity is trivial to achieve for machines - it involves relaxing the constraints of whatever algorithm is used to create the performance, and injecting some randomness into it.
Interpretation - better than adequate, not necessarily world class - is a solved problem.
The heuristics for dynamics and tempo changes aren't particularly complicated. You can do a lot with fairly simple phrase recognition. You don't even need a full harmonic, melodic, and structural analysis.
Composition is a much harder problem - especially at the Bach level. And the state of the art is nowhere close to being able to produce satisfactory Bach-level compositions. (In spite of what David Cope says about his work.)
I think music will be the first of these milestones, and before long AI will be writing enjoyable compositions and pop chart hits.
Next AI will write passable stories, and then decent ones, and then marketable novels with a coherent plot.
AI assist tools will be able to help animators create scenes, backgrounds, and characters using just keywords descriptions. Then with the ability to write scripts will come the ability to create movies. It will take some editing to filter out the nonsense but we'll probably have AI film productions within 30 years.
That's my guess, purely based on what's been demonstrated so far.
I actually think AI will never really make great "pop chart hits", for the mundane economic reason that once we figure out how to make good music with AI, it will be possible to absolutely flood the market with that genre of music, making it impossible for any potentially great hits to stand out.
That said, I absolutely believe that AI-generated music can soon and will easily easily replace any sort of background or generic license-free music uses, where they simply need to be "good enough" rather than great.
> How long until we could assemble an orchestra of computers that can interpret sheet music with feeling well enough to impress a human audience?
As others wrote, this will be the first to go. In fact, I believe current breed of NNs can do this already.
Artistic creativity is a trivial problem compared to all others. All you have to do is inject a bit of randomness to the process and then not tell people that the work was done by a computer program.
The "interpret sheet music with feeling" actually happens within the brain of the listener, who tries to connect the music and feelings it evokes in them to a vision of a human being who created that music. In other words, emotions are actually projected. The mechanism works somewhat well if the artist is a human with clear intent of creating emotional impact. But when the artist doesn't intend to create that impact, the audience will find one in there anyway. And so will they if the artist is actually a matrix multiplicator running on a stack of GPUs.
The same phenomenon happens in writing[0], but writing (and similarly, painting) is harder, because the sentences have to have at least some semblance of sense[1]. In music, anything that's not just pure white noise can get accolades for creativity if you insist hard enough that it was composed by a gifted human.
--
- [0] - Ever heard of the stories about people building these whole towers of interpretations of a literary work, and then when someone asks the author whether they meant any of that, it turns out the whole edifice is just a pile of bullshit, and the author really just wanted to write a story they liked?
Yes exactly. You can even get a head start by training on award-winning performances by Izhak Perlman and such, comparing them to the sheet music. Then your NN will be able to read sheet music in the style of Perlman, the same way that GPT2 can write in the style of Tolkien.
Kasparov remarked that Deep Blue seemed to make insightful moves in a way that wasn't machine-like, and he suspected it was human assisted. I don't know if he was right or wrong, but I'm sure today's chess software will feel at least as insightful as Deep Blue did to a chess grandmaster, especially if they think they're playing against a human.
> Extrapolating the trend of the last 30 years, there is evidence that computers will be able to solve every task a human can using pattern matching. If that isn't AGI, it might turn out to be better than intelligence.
Only if the trend lasts and is applicable to every human task. Those are pretty big assumptions.
> If someone finds an objective function for deciding what decision parameters are important AGI could be upon us very quickly.
And if that function doesn't exist, because the real world (not a board game) is messy, dynamic and complex?
> As a postcript, I think people radically overestimate human intelligence.
Individually maybe, but as a group we're pretty damn impressive. I think you're radically underestimating the species. But this has been the case for strong AI proponents since the 1950s. AGI is always 20 years and one good algorithm away.
Agreed (with the analysis, not necessarily that we're anywhere close to achieving GAI). Nice to finally see proof that John Searle was 100% wrong with his assertion that intelligence couldn't possibly be algorithmic though.
I don't know when your dad used to say that, but any researcher from the 50's would quickly tell you that Chess is just a giant tree search problem.
We have a huge number of unsolved pattern matching problems around, so AI still has a lot of value to bring. But we don't have that much evidence that it suffices for everything.
Extrapolating the trend of the last 30 years, there is evidence that computers will be able to solve every task a human can using pattern matching. If that isn't AGI, it might turn out to be better than intelligence.
The technological future is unknowable, so believing AGI is certain is too much. But believing it certainly isn't around the corner is also too little. If computers can do anything a human can intellectually, they have reached AGI. The list of discrete tasks (games, decision making once the parameters are defined) a computer can't do is a very short list.
If someone finds an objective function for deciding what decision parameters are important AGI could be upon us very quickly. As a postcript, I think people radically overestimate human intelligence.