Rate of improvement

[insaneoctane]

I'm not an AI expert. But I know AI has shown rapid improvement towards the later part of its training on its way to mastery in other domains (ChatGPT, image recognition, Go, etc). Is that expected or is the size of Tesla's dataset problematic? I think Tesla's dataset and ability to collect data is likely order(s) of magnitude beyond other AI success stories and suspect that compute limits them more than other examples (hence Dojo). It might be an issue with how long iterations take? Just wondering and interested in thoughts from those with more expertise...

 

[daniel]

No expertise here, but when did that ever stop anybody on the internet?

AI has shown remarkable improvements in game-playing, where the rules are clear and unambiguous and the "universe" is free of exceptions or moral judgement. It's also demonstrated remarkable improvements in chat bots, producing very human-like speech, but where also it has shown a propensity for lying. (A lawyer used it to create a brief, in which the AI cited case law that did not exist!) In other examples, chat bots have created "news" stories citing real people saying things they never actually said.

Autonomous driving cannot tolerate the invention of untrue "facts," and it exists in a "universe" where there are no rules governing the behavior of pedestrians and other drivers. It is therefore massively more difficult than games with clear rules, or chat bots that can get away with making up "facts."

Five years ago I thought a fully autonomous car was five to ten years away. I no longer expect to see one in my lifetime. (I'm old.) My Model 3 drives itself on the highway as well as I could. There are streets in town where it would be reckless in the extreme to attempt to use even the latest iteration of the misleadingly named "FSD."

 

[insaneoctane]

Agree that the solutions with clear rules must help. Wondering if image recognition is a better analogy? Certainly simpler!

 

[TresLA]

Driving in the real world amongst humans (i.e. - not all automated vehicles, human pedestrians, etc.), is a much more social activity than many understood it to be only a few years ago. Therefore it requires a much higher threshold of near perfection (i.e. - “the march of 9’s”) to be acceptable/believable than other areas.

Unlike writing words that sound human where predicting what people will think or say is often a luxury, driving requires “reading” each others’ minds: will you go first or will you let me go first? Even if you don’t legally have right of way? Will you risk squeezing in between my car and that truck? Will you run this yellow/red light or can I turn left in front of you? Are you slowing down to a stop only temporarily or stopping in the driving lane where I should cross the double yellow to go around you?

There are lots of self-driving areas where improvement is just a matter of compute/working on it (e.g. - recognizing potholes, school bus flashing lights, new construction changed the driving path, etc.), but the social aspect is essentially predicting human behavior and thinking. Even humans aren’t great at it sometimes, hence we still have many accidents.

 

[daniel]

Agree that the solutions with clear rules must help. Wondering if image recognition is a better analogy? Certainly simpler!

Image recognition is critical, and they're getting better and better at that. But as TresLA points out, predicting the behavior of others is also critical.

An example I've brought up before: Narrow two-lane road, bicyclist is riding on the narrow shoulder, in or near the driving lane. A car is coming up behind the cyclist and another is going the opposite way. The car coming up behind the cyclist will have to veer to the left to give the cyclist room. A human in the other car would anticipate this and keep to the right to give the other car space. Will the AI be able to make that judgement? I've seen no evidence that any AI is yet capable of that. As long as the street is wide and there's plenty of space for both cars and the cyclist, no problem. But many streets are not ideal. Waymo and the other robotaxi companies only operate where the streets are ideal. But Tesla wants to operate anywhere a human driver would go.

 

[beatle]

But Tesla wants to operate anywhere a human driver would go.

This is their big problem. Instead of focusing on where FSD can offer the most benefit for the lowest investment (the highway) they are obsessed with getting everything right all at once. I'd love to let the car take over for me in stop and go traffic or long stretches of highway. I have virtually no interest in letting my car drive itself on city streets.

 

[daniel]

This is their big problem. Instead of focusing on where FSD can offer the most benefit for the lowest investment (the highway) they are obsessed with getting everything right all at once. I'd love to let the car take over for me in stop and go traffic or long stretches of highway. I have virtually no interest in letting my car drive itself on city streets.

They had highway driving down pat several years ago. My EAP works wonderfully on highways. It's still Level 2, and looks as though it will always be Level 2, but it takes nearly all the stress out of highway driving.

I would love full Level 5 autonomy because I'm old and in a few years may have to give up driving. But I fear that's not coming in my lifetime.

 

[Supcom]

Image recognition is critical, and they're getting better and better at that. But as TresLA points out, predicting the behavior of others is also critical.

An example I've brought up before: Narrow two-lane road, bicyclist is riding on the narrow shoulder, in or near the driving lane. A car is coming up behind the cyclist and another is going the opposite way. The car coming up behind the cyclist will have to veer to the left to give the cyclist room. A human in the other car would anticipate this and keep to the right to give the other car space. Will the AI be able to make that judgement? I've seen no evidence that any AI is yet capable of that. As long as the street is wide and there's plenty of space for both cars and the cyclist, no problem. But many streets are not ideal. Waymo and the other robotaxi companies only operate where the streets are ideal. But Tesla wants to operate anywhere a human driver would go.

No. The car coming up behind the cyclist should slow down until it is safe to overtake the cyclist.

 

[JB47394]

Will the AI be able to make that judgement? I've seen no evidence that any AI is yet capable of that.

That's exactly the sort of thing that neural nets will be able to handle well. They're really good at making instant appraisals. Apparently the techniques to handle longer horizon stuff isn't there yet. So just as ChatGPT can continue a sentence, so too will autonomy systems be able to take a driving scenario and "continue" it.

Note that FSDb does do prediction of other agents on the road. I don't have the clip, but a recent video showed FSDb juggling multiple pedestrians and cars, finding a clean path through the whole thing in a very human-like way. As with all things FSDb, it doesn't succeed everywhere every time, but they're working on it.

 

[spacecoin]

Driving in the real world amongst humans (i.e. - not all automated vehicles, human pedestrians, etc.), is a much more social activity than many understood it to be only a few years ago.

No, not really. Perhaps if you say the last ten years it would be true but probably not. The use of game theory in SDC is as old as the field more or less.

This is a random article from 2017 on pedestrian behavior:

Radware Bot Manager Captcha

I would also argue the the rate of improvement in real world AI has an S-curve. Fast in the beginning and slow as hell to get those “nines”. People want to think it’s exponential, but nothing could be further from the truth.

This is why Comma AI with a 5 person team can have similar supervised performance as Tesla (with several hundred people working on FSD) on highways and why it takes a almost a decade for the robotaxi companies to get from supervised to unsupervised.

 

[daniel]

No. The car coming up behind the cyclist should slow down until it is safe to overtake the cyclist.

But that's not the behavior of actual drivers. And on South Kihei Road there are stretches of over a mile where it's never safe to pass the cyclist without edging into the oncoming traffic lane. This is what I meant by roads that are "not ideal" and that at present no autonomous car can handle safely. Tesla recognizes this, which is why even FSDb is still Level 2. South Kihei Road is a particularly bad road, but there are roads like this all over the place. And it is the only road, other than the highway, that goes all the way through Kihei from north to south.

That's exactly the sort of thing that neural nets will be able to handle well. ...

Perhaps they will. I was very optimistic of a 5 to 10 year time line five years ago. I now think that it will take much longer. They cannot handle them now, which is why AP, EAP, FSD, and FSDb are all still Level 2, and all existing robotaxi operators are limited to places with ideal streets.

 

[Hiline]

They had highway driving down pat several years ago. My EAP works wonderfully on highways. It's still Level 2, and looks as though it will always be Level 2, but it takes nearly all the stress out of highway driving.

I would love full Level 5 autonomy because I'm old and in a few years may have to give up driving. But I fear that's not coming in my lifetime.

EAP works well most of the time, but I've had EAP miss highway exits before so it's not quite perfect yet. Furthermore, EAP can totally be Level 4-5 before FSD even gets to level 3, given how much more complex city street driving is than highway driving. I would not mind Tesla prioritizing advanced EAP over FSD, the future of which is still quite uncertain.

 

[daniel]

EAP works well most of the time, but I've had EAP miss highway exits before so it's not quite perfect yet. Furthermore, EAP can totally be Level 4-5 before FSD even gets to level 3, given how much more complex city street driving is than highway driving. I would not mind Tesla prioritizing advanced EAP over FSD, the future of which is still quite uncertain.

I never tried NoAP, so I'll take your word about the highway exits. My use of EAP on highways is limited to driving in a lane, and changing lanes with the turn-signal stalk.

I don't believe that Tesla has any intention of raising EAP above Level 2. And as for EAP being Level 5; by definition Level 5 can drive anywhere with nobody in the car. A Level 4 system could be limited to highways or freeways. Level 5 cannot. And Level 4 can safely park before turning control over to a human. So realistically, EAP would never go above Level 3 (driver at the wheel, ready to take over, but the car must be able to give some advance notice.) But as I said, I don't believe Tesla has any intention of that. They want us to buy FSD, and if they ever manage to get beyond Level 2, I'm sure we'll have to pay for FSD to get it.

It's noteworthy that after all these years, AP and EAP are still "beta" features.

 

[EVdrive]

Five years ago I thought a fully autonomous car was five to ten years away. I no longer expect to see one in my lifetime. (I'm old.) My Model 3 drives itself on the highway as well as I could. There are streets in town where it would be reckless in the extreme to attempt to use even the latest iteration of the misleadingly named "FSD."

Come to San Francisco and try a fully autonomous Waymo, with nobody behind the drivers's seat, and feel how confidently and comfortably it navigates streets full of cyclists, jaywalkers, stalled cars, and narrow streets with paralell parked cars.

Waymo has hit a point where they've solved the engineering problems for autonomous driving, and the next set of challenges ahead of them will be scaling problems - building more vehicles, mapping more cities, etc. The operational expertise for scaling is vastly different from the research and engineering they've been focused on so far.

Tesla does have a unique set of constraints they've chosen to work with, and it'll be interesting to see if self driving is possible with those constraints. But we've seen that the ML bits of self-driving in a complex city are entirely solvable, the only question that remains is whether it's also possible to run this based on a limited # of video feeds and a much cheaper computer.

 

[TarmacSurfer]

The “answer” thus far to virtually any unanticipated situation with the Tesla AP is “slam on the brakes”. While a generally safe course of action (unless you’ve got someone right behind you that’ll end up in your trunk) it’s crude and prone to seriously pissing off everyone else around you - particularly those following you. Add to it the extremely slow to resume normal speed once the hazard is clear and it’s a formula for a clunky solution that is nowhere near ready for prime time.

I’ve observed that if an identified object (vehicle, motorcycle, pedestrian, etc.) is even close to the dividing line or so much as an inch over it, the reaction from the computer is as described above - full stop. Not “drift to the right shoulder and slow to avoid the potential threat” as most drivers would do. Wham. Full stop. And until the other object is determined to be far enough back in its own lane, you ain’t going anywhere - irate motorists blaring car horns behind you notwithstanding. And “unidentified” objects like trash / debris on the road, animals, etc.? Nope - you’re running those over unless you manually intervene.

Safe? Yes - mostly. Efficient? Not at all.

The real promise of autonomous driving is that it can do this BETTER than people can. Take the emotion out of driving, be more efficient, etc. I envision a world of fully autonomous vehicles whizzing along at high speeds more or less bumper-to-bumper in all conditions because computers can handle it and not freak out. They can perform the myriad calculations necessary faster and more reliably than a person. So why aren’t we there yet?

Many answers - some of which have been touched upon here and in interviews with Elon and others. It’s a very difficult problem with many variables and the weight that’s assigned to one factor or another for purposes of coding decision-making is not static - it’s constantly in flux itself. But computers can STILL do a better job if only the inputs and software were better.

As I see it, it’s a matter of when - not if. Computers already do many tasks far better than people simply because the input information is adequate and the software is sufficiently well-developed to handle virtually all possible permutations of those problems. Obviously FSD has a long way to go but is making progress (particularly Mercedes and Toyota). Tesla was the leader in this area for a while but has lost that throne - now it’s somewhere between “expensive gimmick” and “maybe someday”.

 

[TresLA]

No, not really. Perhaps if you say the last ten years it would be true but probably not. The use of game theory in SDC is as old as the field more or less.

This is a random article from 2017 on pedestrian behavior:

Radware Bot Manager Captcha

I would also argue the the rate of improvement in real world AI has an S-curve. Fast in the beginning and slow as hell to get those “nines”. People want to think it’s exponential, but nothing could be further from the truth.

This is why Comma AI with a 5 person team can have similar supervised performance as Tesla (with several hundred people working on FSD) on highways and why it takes a almost a decade for the robotaxi companies to get from supervised to unsupervised.

Yes really. There were those that didn’t think a production FSD (generalized level 5, not just geo-limited level 4) was soon possible, but the general consensus for those advancing the field was much more optimistic. Heck, forget official papers, even I was already doubting it having just purchased FSD Capability package on the cheap. I’m happy with the purchase because I’m happy with the progress. I had pretty low expectations so I wasn’t disappointed. I’ve long said the “rootier” issue is 1) how social the act of driving is and 2) the lack of language or clarity describing the vast expanse between advanced Lane Keep Assist and self-driving (as discussed on this forum ad infinitum, the SAE levels are lacking).

 

[BitJam]

James Douma is certain that FSD progress is exponential or faster than exponential:

It's always been exponential it's just that the exponent increases. [...] FSD is definitely exponential.​

​

OTOH, Elon Musk says the progress is logarithmic:

If you were to plot the [FSD] progress, the progress looks like a log curve ... a series of log curves. [...] It goes up in a fairly straight way and then starts tailing off and you start getting diminishing returns.​

​

For those who are not up on logarithms, logarithmic growth is the opposite of exponential. To slightly oversimplify, many computer algorithms that seem magically fast (like Google search) get their speed because the computation time only grows logarithmically. If exponential growth is fast, faster, FASTEST, then logarithmic growth is slow, slower, SLOWEST.

Chuck Cook (of unprotected left fame) has a lot of experience with FSD and he does not think the progress is exponential. He said if it is exponential then it's down on the flat part of the exponential curve that looks linear.

James Douma said Elon is planning to increase the compute power of Dojo by a factor of 100 in the next two years. IMO this is a testament to Elon's belief that the rate of progress is far from exponential and he is spending perhaps billions of dollars to throw a lot of computer power at the problem to move it along. ISTM if the rate of progress were truly exponential in any meaningful sense then spending that kind of money would be a waste because improvements would soon be snowballing out of control anyway.

Karpathy wrote an interesting article looking back 33 years in neural net development and projecting that ahead to 33 years in the future.

Deep Neural Nets: 33 years ago and 33 years from now

Musings of a Computer Scientist.

karpathy.github.io

​

He seems to think that when looking at these longer time scales, overall improvements in neural nets scale with the improvements in computer hardware. This does not directly refute or confirm the claims above. But the long term progress will be mind-blowing:

In 2055, you will ask a 10,000,000X-sized neural net megabrain to perform some task by speaking (or thinking) to it in English. And if you ask nicely enough, it will oblige. Yes you could train a neural net too… but why would you?​

 

[spacecoin]

Yes really. There were those that didn’t think a production FSD (generalized level 5, not just geo-limited level 4) was soon possible, but the general consensus for those advancing the field was much more optimistic. Heck, forget official papers, even I was already doubting it having just purchased FSD Capability package on the cheap.I’m happy with the purchase because I’m happy with the progress. I had pretty low expectations so I wasn’t disappointed. I’ve long said the “rootier” issue is 1) how social the act of driving is and

You need to understand that L5 is an aspirational level and that it’s zero chance any car in the coming decades will be able to drive in an unlimited ODD regardless of unit budget.

2) the lack of language or clarity describing the vast expanse between advanced Lane Keep Assist and self-driving (as discussed on this forum ad infinitum, the SAE levels are lacking).

What’s unclear? If you understand the concept of ODD and OEDR it should be crystal clear. If the car is performing the full OEDR and the human is not legally driving, and can read a book - it is self driving (L3). The car will will notify you if it leaves the ODD and if you fail to take over, the car will stop safely. If you can sleep in the backseat, thae car is self driving (L4).

 

[spacecoin]

James Douma is certain that FSD progress is exponential or faster than exponential:

James Douma has zero experience with SDC and ML. He is a stock pumper that has no insight nor background to be more of an expert than you and me, He talks word sallad using technical terms which sound impressive if you don’t know what he’s talking about. Like Elon.

It's always been exponential it's just that the exponent increases. [...] FSD is definitely exponential.​

​

OTOH, Elon Musk says the progress is logarithmic:

Elon says… See above + robotaxi on hw3 since 2016.

In 2055, you will ask a 10,000,000X-sized neural net megabrain to perform some task by speaking (or thinking) to it in English. And if you ask nicely enough, it will oblige. Yes you could train a neural net too… but why would you?​

In 2055 FSD might be Level 4. On HW19. Current cars will never.

Disengagement rate is still at about 12 miles not counting accelerator taps. After two years.. When they 2000x that they will still be trailing Waymo…

 

[TresLA]

You need to understand that L5 is an aspirational level and that it’s zero chance any car in the coming decades will be able to drive in an unlimited ODD regardless of unit budget.

“Zero chance” is almost as useless a claim as “100% chance” when predicting future innovations.

What’s unclear? If you understand the concept of ODD and OEDR it should be crystal clear. If the car is performing the full OEDR and the human is not legally driving, and can read a book - it is self driving (L3). The car will will notify you if it leaves the ODD and if you fail to take over, the car will stop safely. If you can sleep in the backseat, thae car is self driving (L4).

If it were so clear, we wouldn’t have the myriad of internet arguments about what they do or do not mean. Doing a quick search on just this forum for “SAE levels” and I see you’ve already been a part of them. You can reference those instead of rehashing the same arguments here. Hopefully we can keep this thread on topic.