> People's inability to accurately assess how easy an inherently easy problem is has no bearing on people's inability to accurately assess how hard an inherently hard problem is.
Really? It seems to me that if people over-estimated the difficulty of an "easy" problem like mastering go, then they're even more likely to over-estimate the difficulty of a hard problem like self-driving. In fact, the over-estimation could scale up faster than linearly, if estimating two problems of size X is easier than estimating one problem of size 2X.
> with the algorithms and computational models we have today or in the near future
That's the thing. When predictions were being made about the difficulty of mastering go, people didn't have the algorithms and computational models that we have today. Similarly, predictions made today about the progress of self-driving cars may be lacking critical information about the algorithms and computational models that will be available in the near future.
I have an opposite view: it's not that shocking that AI has advanced a lot. It's a lot more shocking to learn that humans aren't as great as we hope to be. Also, our prediction sucks. I mean, who said that go is such a difficult problem that it would take a lifetime to solve? Sounds like intellectual arrogance to me. Sure, the problem space is huge, but it's well-defined and homogeneous. There was a time that reciting a long text or multiplying large numbers is considered a humanly intelligent thing, only that it wasn't. Alan Turing used to think that AI is good to humans because it teaches us to be humble, and I think we're kind of getting there (for certain domains). On the other hand, things like self-driving will remain unsolvable because the problem is fundamentally ill-defined; we don't even know what is a good driving.
(Edit) To those who think self-driving is a well-defined problem: it can be in some remote areas, but imagine driving in bustling city streets with kids, bicycles and dogs. The driving problem becomes a communication problem.
> You're saying that for humans it requires perhaps 30% additional learning effort but for computers 9900%, and I'm supposed to accept the 30/9900 discrepancy because… what?
Because AI is not even remotely close to “solved”, and driving is a combination of a million edge conditions, social knowledge and communication, common sense knowledge, vision understanding, planning, and force feedback adjustment.
What’s unclear to me is why you’d expect these things to be so easy. To even say “for computers” doesn’t seem to recognize this isn’t just about executing basic math, but is rather trying to represent intelligence. There is no single computer here, these are implementations of ML models, which are each unique with their own architecture, parameter space, and capabilities.
> I think a car is going to need a Theory-of-Mind to navigate complex social driving environments
As someone with a lot of experience in self-driving cars, my opinion has changed over the course of the last decade from "we can create smart enough models of these separate problems to create a statistically safer product" to "first you need to invent general AI."
It becomes immediately obvious as you encounter more and more edge cases, but you would never even begin to think of these edge cases and you have no idea how to handle them (even when hard coding them on individual cases - and you couldn't anyway as there are too many) so you realize the car actually has to be able to think. What's worse, it has to be able to think far enough into the future to anticipate anything that could end badly.
The most interesting part of self-driving cars is definitely on the prediction teams - their job is to predict the world however many seconds into the future and incorporate that into the path planning. As you can guess, the car often predicts the future incorrectly. It's just a ridiculously hard problem. I think the current toolbox of ML is just woefully, completely, entirely inadequate to tackle this monster.
> In ten years, probably no human can compete with AI drivers anymore.
That's what they said 10 years ago. Sooner or later people will say it and be right, but the last few percent of any problem is a lot harder than people give it credit for. It may not be that hard to stay in a lane or write a little code, and that may look like it's doing most of the job, but those common tasks are just the easy part.
I mean, your entire counter-argument is linking a single person's opinion piece. He says that in general more computation is "good" and that search/learning "seem" to scale with computation. That's about it. It doesn't refute the key ideas at all.
He also gives the stereotypical horribly flawed trope about how "some people in the past didn't think computers could beat them in chess, and they were wrong, then some people thought computers couldn't beat them in go, and they were wrong, so now what they say about machine learning today must be wrong too".
Which is a completely illogical line of reasoning. By that reasoning, I present this same argument: When cars were first invented some people said that they'd never be able to reach 50mph, and they were proven wrong, then some people said they'd never be able to reach 150mph, and they were proven wrong, and therefore anyone that doubts my claim that we'll have 1,500mph cars on our streets next year is obviously wrong, because look, some people in the past made bad predictions.
>I think you're overselling the extent of the difficulty here. People from out of state usually manage to get through New York, pedestrians and all, without getting into an accident.
At this point, "it's easy for a person to do it" is probably better evidence that it's a hard problem in AI.
In this case, I believe it's an inverse learning problem. We observe other agents acting in an environment, deduce their policy, then implement that policy ourselves.
A self-driving car that can reconfigure it's software to mimic the behavior of other cars it observes, sounds difficult to me.
But the AI drives slowly and gets confused easily. Regular drivers routinely have to go around self-driving cars. Not to say they won't improve, but it seems like current AI is assistive to the point where it might be harmful when drivers rely on it in speeds and situations where they shouldn't. I'm sure it will keep improving, but I feel like this is one of those situations where the amount of data and training required, and the amount of iteration on the software required to handle edge cases is not impossible but is exceptionally difficult.
They've barely started trying. We'd be reaching the limits of AI if self-driving cars were an easy problem and we couldn't quite solve it after 15 years, but self-driving cars are actually a hard problem. Despite that, we're pretty darn close to solving it.
There are problems in math that are centuries old, and no one is going around saying we're "reaching the limits of math" just because hard problems are hard.
probably the long tail of unique problems that any autonomous system in an open environment faces.
It's kind of the crux of all these ML learning based systems. By definition you can't learn what's totally new and absent from data, but that is precisely where the intelligence lies when it comes to human drivers. Which is why I think genuine self-driving without fancy geofencing or whatever is almost an AI-hard problem.
> A driver can respond to a visual stimulus in a few hundred milliseconds, and decide an action, such as making a turn. So the computational depth of this behavior is only a few tens of steps. We don't know how to make such a machine, and we wouldn't know how to program it.
IMHO, that makes no sense. Assuming we have a really large state machine with all many possible situations in a state of driving on a straight highway with a few vehicles with more or less constant relative velocity, determining if we need to make a turn is definitely doable in a few computation steps.
>how many years of experience is needed to drive on public roads without supervision
If, as I strongly suspect, full self-driving requires artificial general intelligence, Waymo's algorithms will not get there no matter how long they run simulations or even how many real road tests they do.
> if the CV algorithm fails against these examples when humans don't, then the CV algorithm is too brittle and should not be used in the real world.
This is the tricky bit.
Night-time driving, bad weather, icy roads, bumper-to-bumper traffic: these are all situations in which some algorithms can outdo humans in terms of safety. Faster reactions, better vision (beyond what human eyes can see), and unlimited 'mental stamina' can make a big difference in safe driving.
But then there will be the occasional situation in which the CV screws up, and there's an accident. Some of those are ones where many/most humans could have handled the situation better and avoided the accident.
So how do we decide when the automated car is 'good enough'? Do we have to reach a point where in no situation could any human have done better? Must it be absolutely better than all humans, all the time? Because we may never reach that point.
And all the while, we could be avoiding a lot more accidents (and deaths) from situations the AI could have handled.
> A human can learn to drive a car in 10's of hours, while AI's still require millions of hours of training, and they still end up inferior.
I think it's a bit unrealistic to compare these numbers. A human learning to drive has usually started with 16-18 years of learning how the world works before the 10 hours of transfer learning.
> the idea that more is always better is the common misperception that i'm addressing
It's not a misperception in the case of machine learning. The system can easily discard data that isn't useful at a given point in time. When it is determined to be useful, it's great to have it there.
I'd suggest trying out some machine learning yourself to get a good understanding of why more data is better. Tutorials for Kaggle's Titanic competition can be a good primer.
> if you've ever tried to design a system with multiple, sometimes conflicting inputs, complexity is a real cost to think about (not to mention financial cost).
Not when the system is using machine learning, particularly neural networks.
All the self driving car systems are using deep convolutional NNs, so the complexity of more data is helpful, not a hindrance.
> machine learning is cool, but it's not a panacea. there are plenty of meticulously trained AI's that fail comically in novel situations (like the example in the original article)
The failure in the article was due to lack of diverse sensor input, not because the system was built using machine learning.
You're right that machine learning isn't a panacea. The limits aren't well defined. Still, in the case of self driving cars, deep learning is state of the art. Nobody is hand-coding rules for how to drive the car. It's all about having as much good data as you can gather.
Explicitly programming scenarios in is not how it really works.
>How much of it have we mapped, and how well does the machine cover it?
Enough for Google cars for example to have logged over 300K miles in actual conditions with no incidents.
Also enough to have such things (besides Google's) in pilot operation in several cities the world other, for things ranging from cargo transport to mass transportation (self-driving busses).
It's not like "programming the scenario in" is some huge switch/case statement that needs to cover all possible arrangements of things on the street -- it's machine learning algorithms with several rules and invariants to check and various corrective responses when those are off, and the smartness comes from the combinations of such rules.
>machines will remain shitty drivers because they can't respond with creative judgment to a new situation.
The think is, with the appropriate machine learning algorithms they can both add experience and respond with creative judgement to new situations -- they don't have to have hardcoded responses to them from the start.
About a decade ago winning GO or self-driving cars were seen as pipedreams many decades away. Yet here we are.
The author is making the mistake of thinking that just because he can show some areas were we aren't as far as we thought he has made an argument against AI.
Thats not how it works. We don't get to decide what is the right metrics. All we can see is that we keep making progress sometimes large leaps sometimes slow.
I always find it fascinating that we have no problem accepting the idea that human consciousness evolved from basically nothing but the most elementary building blocks of the universe and once we became complex enough we ended up being conscious yet somehow the idea of technology going through the same just in a different media seems to many impossible.
I know where my bet is at least and I haven't seen anything to counter that neither the OP's essay.
> Self driving is probably the most sophisticated domain besides chat, and there too it's the exact same problem. They can drive a vehicle in some complex scenario with super human performance, and then they randomly drive straight into a highway divider at full speed on a clear day with no traffic.
Yes, very good point. Self-driving maximalists who believe that self-driving will be solved with more data need to realize that ChatGPT was trained with ALL the data possible and is still deficient. This defect is probably inherent to existing neural net models and a leap forward of some sort is necessary to solve this.
Another scary thought: just as each ChatGPT session is different, and you never know whether the agent is going to get angry, overly apologetic, or something else, every self-driving drive may be different due to emergent properties in neural networks that even the best in the field do not yet understand.
> I don't think one necessarily follows from the other though. We might be able to make an AGI with the intelligence of a human teenager. That doesn't necessarily mean it would be better at driving a truck than a more specialised algorithm.
I agree that an AGI wouldn't necessarily be good at driving trucks. I meant it in the sense that an AGI would be capable of producing the "truck driving algorithm" (if we humans can do it, the AGI can do it too, almost by definition).
> I see a lot of people complaining that various machine learning techniques won't lead to AGI and would never create a system which could pass the Turing test.
Most of the complaining seems geared towards media and outsiders who portray X technique as being the "Solution to AGI", not towards the techniques themselves.
I have no idea whom this guy is, and have no thoughts here on his opinions or their falsifiability (happy to agree with you on that aspect if it matters), but just to reply to this particular point here:
> This is well beyond what was considered possible at the time. We're certainly not at 'everything' but shrug. Maybe in ten years.
This feels a lot like what people were saying about self-driving cars being imminent circa... 2015 or so? [1] The skeptical folks rolled their eyes at suggestions that we'd have self driving cars everywhere in a few years, but lo and behold, they were right. Just because we had a massive amount of progress in the years before, that doesn't mean we were on the cusp of achieving the goal. Turns out going from 99.9% accurate to 99.99% accurate (or whatever the numbers were) is harder than all the believers wanted to admit.
This feels like the same thing all over again. Yes, there's been a lot of progress in LLMs. No, that doesn't mean we're anywhere close to deep learning being able to do 'everything' in 10 years, whatever that means.
Really? It seems to me that if people over-estimated the difficulty of an "easy" problem like mastering go, then they're even more likely to over-estimate the difficulty of a hard problem like self-driving. In fact, the over-estimation could scale up faster than linearly, if estimating two problems of size X is easier than estimating one problem of size 2X.
> with the algorithms and computational models we have today or in the near future
That's the thing. When predictions were being made about the difficulty of mastering go, people didn't have the algorithms and computational models that we have today. Similarly, predictions made today about the progress of self-driving cars may be lacking critical information about the algorithms and computational models that will be available in the near future.
reply