Well, there’s our disagreement in a nutshell. I’m already right because all I’ve said is “I don’t know.” You’re the one putting time constraints on it. For some reason you think you are sufficiently well informed to say what won’t happen when.
I think we’re dealing with a variant of this phenomenon:
If an elderly but distinguished scientist says that something is possible, he is almost certainly right; but if he says that it is impossible, he is very probably wrong.
Arthur C. Clarke
-IGU-
I don’t think so - especially if we’re talking about what can be done in short and intermediate time frames, rather than statements about what is physically possible in the universe.
I mean, most elderly but distinguished scientists would say that it’s impossible to build an economically feasible fusion reactor in the next ten years, they would probably be right. Now, it’s certainly true that no one can know for sure whether we can’t build a nuclear reactor in an economically feasible way within a decade. But if someone with a pretty good knowledge of the state of play in the field of nuclear fusion research and engineering says that we’re more than a decade away from the first fusion nuclear plant, there’s no reason to assume they’re wrong (though they very well might be).
You’re starting to see some contrary voices within the AI field stating that we’re actually pretty far away from having AV’s:
“Decades of breakthroughs in the part of artificial intelligence known as machine learning have yielded only the most primitive forms of “intelligence,” says Mary Cummings, a professor of computer science…”
(forgive the weird link - the WSJ article is behind a paywall, but that copy is not)
It is entirely possible that these more pessimistic researchers are wrong, and that we’re not many years away from being able to build an AI that can perform Level 5 autonomy. But it’s not as self-evident that they’re wrong as Clarke’s statement suggests.
And most AI scientists said we were at least a decade away from a decent Go AI, even in the days before a new world champion was born.
I’m going to stick with “I don’t know” and I’m quite certain you don’t know, and neither does anybody else posting here.
-IGU-
This is a bit of a quibble - but given your penchant for criticizing people for minor things, your statement is false. Even as early as 2014, there were “decent” Go AI’s - AI’s that were capable of winning games from solid human Go players. The best Go AI’s at the time were capable of competing at top amateur level by then.
That said, even at a more general level, I think the Go case rather proves the point that 1pg and I were trying to make in the other thread. I know that most AI scientists were skeptical that the methods used to create an AI for chess could produce an AI that could win at Go (though I’d love a link to any such scientist actually saying we were at least a decade away from even just a “decent” Go AI in the days prior to the championship match). That there would need to be a “fundamental breakthrough” in AI before the machine could master Go, given the differences between the games and the inadequacy of then-existing methods of machine-learning to address those differences. Then Google made that fundamental breakthrough, started doing AI differently than they had with chess, and were able to produce a world-champion Go AI.
That doesn’t have much bearing on the sort of, “Will humans ever be able to something?” question that I think was the target of Clarke’s observation. But it does have bearing on whether or not a commercialized application of a specific AI approach to solving a problem can be implemented within a specific timeframe without a “fundamental breakthrough” that may be outside the control over the company trying to implement the commercial application.
You’re right, there were Go playing computer programs that were playing a decent game before AlphaGo. Just nowhere near professional level. And not using any of the AI techniques used today. And with no clear path forward other than waiting for hardware to get bigger and faster. Until DeepMind revealed AlphaGo in 2015.
Well, that’s a nice story, but it’s not what happened. First of all, although Google bought DeepMind in 2014, let’s please stick to crediting the right people. It wasn’t Google.
And second, there was no fundamental breakthrough at that time. That had already been made in the 1960’s originally, but then languished. In more recent times machine learning and the newer concept of deep learning was being used with neural networks for years. But there had been recent successes (see ImageNet - Wikipedia) which DeepMind took as encouragement to try these techniques on Go. You can read about the 2018 Turing award here (Fathers of the Deep Learning Revolution Receive ACM A.M. Turing Award), where Bengio, Hinton and LeCun were honored for their work. You’ll note that the breakthroughs were mostly in the late 1980’s and 1990’s.
Producing AlphaGo involved training it on lots of human games, then having it improve by playing itself. No breakthroughs needed, just applying known techniques to a new problem where they hadn’t been tried before (although that was certainly significant work). The results were indeed spectacular, and stunned the world (or at least the small part of the world that cared, like me), when in March 2016 AlphaGo won its match with Lee Sedol, a top ranked professional Go player. People were still trying to wrap their heads around this when in May, 2017 AlphaGo (slightly tweaked) proved this was no fluke by beating Ke Jie, ranked best in the world at that time.
Oh, it absolutely does. If you pay attention, you’ll notice that the answers were actually just lying around for years waiting for somebody to notice. Sure, the implementation required some hard work and no doubt some new cleverness, but it was already perfectly doable if somebody had tried.
This is underscored by the fact that less than a year later,. in October 2017, DeepMind revealed that they had decided to try something a little different and start the training with no human input at all, just the rules of the game, leaving the AI to entirely train itself through trial and error. Again, no fundamental breakthroughs, just playing around with the tools at hand.
The result, AlphaZero, was again spectacular. Not only did AlphaZero absolutely crush AlphaGo, but it also happily trained to play chess (and shogi) and absolutely crushed the best chess playing programs available. This 2018 blog post describes what they found: (AlphaZero: Shedding new light on chess, shogi, and Go - Google DeepMind). It’s an easy read.
So, yeah, who is to say which of the many techniques already invented might show stunning results when applied to autonomous driving? Not me. Not you. When will it happen? No telling. Could be next week. Could be never. My opinion is that predicting when requires a certain arrogance that is very hard to justify.
Clarke’s observation is really just that old experts are likely to think in terms of tools they know being applied in ways they know. They know their limitations. It’s the new guys who see that maybe the same stuff that is being used to recognize cat pictures on the internet can (if you squint at it just right) become the Go champion of the world. And maybe drive a car too.
-IGU-
Oh, I completely agree. And that’s why it’s entertaining and interesting to talk about these subjects, even though no one knows for sure how the technology is going to develop - or how fast it will develop.
Good reading. Thanks!
However there is a comment in there that I find particularly relevant when trying to equate the achievements of AlphaZero with the challenges FSD.
AlphaZero’s ability to master three different complex games – and potentially any perfect information game – is an important step towards overcoming this problem.
That phrase, perfect information game, is reference to games with all the rules clearly defined. Nothing can happen on a chess board that isn’t within the rules. Driving is not like that, not even a little bit like that.
(I own a Model Y, for which I paid an extra $10k for FSD. I recently got the FSD beta but I have not used it very much yet.)
Absolutely! Driving is not the same kind of problem. So it will require something different to solve it. But it’s very uncertain how different. For example, how well can AlphaZero do on an almost perfect information game. Can it do almost as well? Or does it just fail?
So maybe the same techniques can be adjusted to handle problems that are similar?
And attacking the problem from the other side, maybe it’s possible to abstract away the imperfections of the driving information somewhat. What happens if we take the driving task and slow everything down 1000 times? It seems it would become much simpler in some ways. Maybe simple enough to be solved? Then the problem becomes speeding things up, which is a very different sort of thing.
So there’s much to be said for using tools that have been shown to work and adjusting the problem to fit the tools. Surprising things can happen.
Anyway, right now nobody knows why machine learning techniques seem to require massive amounts of data, while humans can pick up new things in just a few tries. Since we know it’s possible, because we do it, I suspect the answer to that question will come soon. From a software perspective it may look pretty much the same, but it will change everything. Overnight.
-IGU-
Why do you think it will come soon?
I can understand the confidence that it’s possible - as you point out, we’re able to do it. But as you’ve frequently pointed out on this thread, no one can really know what the development path for AI will be, going forward. Maybe it’s soon - and maybe it’s 20 or 30 years down the road?
After all, learning to do something after just a few tries was also “possible” future development for machine learning in 2002 (for example) - because obviously humans were able to do it then, too. But developing an AI that could do it wasn’t going to happen “soon” from 2002. Machine learning wasn’t close to being able to do that yet.
I think I know why. 
Humans have experienced massive amounts of data before they are faced with the new things we can pick up in just a few tries. We are not starting from nothing, but have our cumulative life experience already integrated.
That is a good description. For example, humans know long before they learn to drive that a plastic grocery bag lying on the street isn’t going to harm them or the car, but a same size rock is to be avoided. You don’t need to drive over 100 different bags and 100 rocks of different types before you figure out how to handle seeing one of them.
Put a big rock in the bag and you’ll likely be a bit suspicious if there is any wind because the bag is fluttering but not moving so you’d want to avoid it.
You’d need another ~99 tries with AI.
AI researchers, of course, know this, and that is why the cars are generally over cautious but still making mistakes.
Mike
Just to clear up any potential confusion: A game with perfect information means that there’s no hidden information. In chess, everything is visible to both opponents. Contrast this with poker for example, where you have hidden information, i.e. players do not know the cards of their opponents.
In other words, all the rules clearly defined (i.e. a game), and also no hidden information.
Sorry, no. Babies can do it too. There’s something about the architecture of brains.
-IGU-
Because trying to solve a problem for which you already have an existence proof of a solution is much, much easier than trying to solve a problem for which there may be no solution. It’s partly a human psychological thing.
But it’s also a very practical thing. We know that we’ve got neural nets that do things a lot like the human brain. They’re just not quite there yet in a few ways: energy usage, training efficiency, size. So we can try variations and observe how they differ in implementation and in effectiveness. Then iterate. Get closer. The field is so new that lots of easy things haven’t even been tried yet. And as our computers get faster, things that we couldn’t try yesterday are easy today.
Evolution has been doing trial and error tests for millions of years. We’re getting better at looking at the details of the architecture of our brain. Can’t be all that long before somebody notices something interesting about which axons connect to which structures and says “nobody’s tried that yet, have they?”.
-IGU-
Which is true…but as I said upthread, it was no less true twenty years ago. And we know now that an AI capable of driving a car was nowhere “soon” back then.
Obviously, we’re leagues ahead of where we were in 2002 in terms of both computing power and knowledge of how to construct enormously complicated machine intelligences. But we genuinely don’t know whether we’re close to what we would need in terms of both computing power and skill/knowledge to generate an AI that can do these things sufficient to operate as a true Level 5 AV driver.
It sure seems like we’re close, of course - we’ve got AI’s that can do a lot of driving skills. But it might still be a long way away before we can get that last bit. DeepBlue beat Kasparov in 1997, but it was nearly two decades later before an AI beat Lee Sedol in Go. Even though we had an existence proof of a solution to Go-playing in 1997, and proof of an AI solution to chess-playing in 1997, neither the tech nor our knowledge base allowed us to do the same for Go back then.
Again, I think we just don’t know whether an AV AI capable of performing Level 5 driving will be developed “soon” or “a longish time from now.” And I don’t think the fact that we have really great Level 2 AI and rapidly developing Level 4 AI in limited areas means we’re necessarily close to Level 5, based on a super-limited layman’s understanding of where the tech is.
I haven’t thought this through yet, but we can perhaps still discuss it. I don’t think Go is a good analogy to autonomous driving. That’s because Go has everything clearly delineated. For example, how would a Go AI react if the table shook and that caused a piece to shift from its spot to an adjoining spot with another piece? Or if the shaking caused a piece to fall off the board?
I agree, Go is not at all analogous to driving.
But that’s not really why I was bringing it up (or, if I may speak for him, IGU either). IGU was noting Go as an example of a problem in AI that seemed like it might take a fairly long time to solve back in 2012-2013, but actually got solved pretty quickly from that point. It’s a good illustration that sometimes a solution to an AI problem that looks really insurmountable might end up being just around the corner.
And I was bringing it up in comparison to chess - another strictly rules-based perfect information game. We had an AI beat one of the best human chess players as early as 1997, but it took another 19 years to accomplish a similar feat in Go. And my understanding (which might be wrong) is that the AI that was able to win top-level Go ended up being created/invented/structured pretty differently than DeepBlue was as well. I’m using it as an illustration that sometimes when it looks like you’ve gotten most of the way towards solving a problem, it might still be a very long way away. It ended up being only about 8 years between computers winning titles in checkers and then to chess, but more than double that - and a pretty different foundation - to go from chess to Go.
We have pretty decent Level 4 AI systems, that can handle highway driving and certain types of suburban driving to a tolerable safety level. We might therefore think that the step from there to Level 5 autonomy might seem close - just figuring out how to do the same sorts of things, but in an environment where there’s a lot more pedestrians and driveways and pigeons/plastic bags and bicyclists and double-parked vehicles and the like. But it might end up being the case that those extra complications present a significantly larger enough universe of edge cases that we’re in a Chess-to-Go situation - that you can’t get from point A to point B with just a faster/stronger/quicker version of the same type of solution.
I know human drivers that operate at Level 4.
Specific circumstances? No driving at night, or in cities, or on highways, or in the snow…
Geofenced? No driving anywhere new.
I have family members - mature family members - who choose to abide by such restrictions.
(To say nothing about human drivers we have all observed, some of whom would struggle to be rated at Level 3.)
The existence proof was true then. All the other things weren’t.
I think we’re pretty close to solving the problem of training neural nets from just a few examples rather than millions. I suspect it just requires a few tweaks to current solutions, but we don’t know what we’re doing so it’s hard to guess how long it will take. That will make every problem, including the all the problems of autonomous driving, look different. Many easy problems will become trivial, hard problems will become easy, very hard problems will become solvable, and impossible problems will become very difficult.
Whether this means autonomous driving will then be a solved problem, with just some implementation details to work on, won’t be known until they get there. I’m sure that at some point we’ll start arguing about how autonomy can’t work because it won’t do the right thing at a sobriety checkpoint or when challenged by a border guard. New impossible problems.
I drive a car that mostly drives itself. It’s an interesting experience, but it is nowhere near autonomous yet. It keeps getting better. Slowly.
Have you ever driven such a car? Do you drive one regularly? Or are you just discussing abstractions for which you have no intuition?
-IGU-
I have, in fact, driven such a car. I do not own a Tesla, so not regularly. But I have a close friend who does (and they paid for FSD), and since they’re an enthusiastic fan of the car, they’ve made sure that I’ve driven it on more than a few occasions. Not recently (ie. not since the beginning of the summer or so), so perhaps not the very most recent version of the software - but close. I agree that it’s a very interesting experience, and one that is both enjoyable and (for a non-regular user) a bit disconcerting.
I suspect not - a car that is capable of handling any driving situation by itself will probably be recognized as a working example of an autonomous car, even if there are a few specific scenarios where a human outside of the car will need to be able to give directions because of their legal authority (ie. not technically a “driving” situation). We’ll just need the AI driver to be as safe (or safer) than a human driver in driving situations.