Tesla Full Self Driving Is Now 'End-To-End AI'
In this video I discuss the recent Tesla full self driving 'FSD' demo that Delon Musk livestreamed on Twitter 'X' showing a 45 minute drive around Palo Alto California, driving past Mark Zuckerbergs h...
I feel like the NTSB need to draft a min spec for self driving cars and a testing course that involves some of the worst circtimstances to get approved. I feel like all self driving cars should have to have lidar, and other sensors. Computer vision really isn't working out.
You build a benchmark and tesla will train on that benchmark, says nothing about real world use but gets them signed off.
But yes western society is currently in a hellscape of refusing to do even basic regulation of any new technology so it'll probably be a good 20 years of murder robots on the streets before anything gets written down.
To be fair we already have giant metal murder boxes zooming around on the streets. If AI kills even a single person everyone flips out even though over 40,000 people die every year in the US from car accidents. And that is just the deaths, not including injuries. Yet I don't really see anyone calling for more regulations on driving tests for humans.
People want AI to somehow be perfect when in reality as long as AI is even 1% better than humans that's saving over 400 lives per year. AI doesn't get sleepy, distracted, drunk, etc. so it probably already is at least 1% better in most situations. Humans are horrible drivers.
But yes western society is currently in a hellscape of refusing to do even basic regulation
US regulations are only written in blood or money. the united states was built on the backs of slaves, and then wage-slaves. literal graveyards filled with workers.
im not disagreeing with you, i just found this comically disparate to history... ie, its always been a regulation hellscape.
I don't think mandating lidar specifically by name is right, seeing as computer vision is definitely a software problem. Instead they should mandate some method to detect objects in any light condition + a performance standard, which in practice during certification could mean lidar. Regulations should be as minimal and specific as possible.
Good point. Mandate the ends rather than the means. If they get better functionality with some new tech in a few years, we don't want outdated regulations holding the industry back.
computer vision is definitely a software problem.
No, it isn't.
If it were only software, don't you think Tesla should be the best of them all, being the pure software shop they are?
But it is a real world problem. Recognizing real objects in real world conditions like weather, natural and artificial lights, temperatures (want some ice on your camera?), winds & storms, all kinds of unforeseen circumstances, other bad drivers, police and firemen...
And that's why that pure software shop is so bad at it, while all the real carmakers shrug... they are used to it since forever.
Should a self-driving car face more rigorous tests than actual human drivers? Honest question.
Yes. A human brain can handle edge cases it’s never encountered before. Can a self driving car?
These are a tiny number of things I’ve encountered over the past few years. I’m sure plenty of other drivers can provide other good examples. I’d want to know how a self driving car would handle itself in situations like these.
Yes because each person must learn on their own and have limited experience relative to the general public as a whole.
Self driving cars can 'learn' from all self driving cars and don't get tired, forget, or anything like that. While they shouldn't be held to perfection, they should absolutely be held to a higher standard than a human.
Throw I some pot holes and child pedestrian crossing the street, etc and they'd even come out with a powerful marketing ad.
AND have triplicate back up system that runs in parallel.
Pretty much what the UNECE did.. there are standards for these things. Tesla doesn't meet them, which is why FSD 'beta' is still 'seeking regulatory approval' in the rest of the world.
Everyone would build to pass the test track. This does get at the problem though: the permutations of scenarios an L5 system has to correctly process is a huge number. Trying to build a system that can do that appears to be beyond anyone’s av system right now. This is why the most advanced deployments are all geofenced. That way at least the traffic signs and signals, lane markings, etc all understood and tested. Even then ‘shit happens’. Untested scenarios still occur. Also the maps are always out of date.
The problem really requires AGI, and nobody has one of those, or if they do it’s a secret.
It's not really a sensor issue, as much as having software that can interpret the sensor data and act on it. Cameras and lidar effectively provide same thing, distance to objects in 2d/3d. But u need software to process that data and identify where the road is, where little jonny is, and what to do...arguably, the distance measuring problem has been solved for a while with lidar or with cameras, it's object identification and reaction to that info that's not solved. You can't really solve it with traditional if/else programming, while AI gives you only a probability of what something is or what action to do...so the problem is hard.
But ntsb/dmv whatever needs to come up with a way to test and classify autonomous driving software...probably doing real world test and identifying edge cases where it fails.
bUT thAt WOuLD StiFle proDucT iNnoVatIoN!!!
When I found out teslas don't have LiDAR I nearly shat myself.
Yeah, fun stuff happens when the AI tries to interpret vision of sunlight shining straight into the lens.
Don't you know those things cost money!?
Sadly cost cutting MBAs seldom concern themselves with silly things like function or necessity.
For an autonomous vehicle without radars or LiDAR they do still drive pretty darn well. AI DRIVR makes really good videos about FSD on YouTube and love it or hate it, it's quite impressive how well it does despite the the lack of these sensors.
For an autonomous vehicle without radars or LiDAR they do still drive pretty darn well.
That's a bit like saying "For an old hound dog with only two legs left he is running pretty darn fast"... :-)
Still pushing full self driving even while they are dodging lawsuits by claiming it's just highway cruise control that customers are abusing
People commonly confuse Autopilot and FSD beta. One is the advanced cruise control and comes on all models while the other is supposed to be autonomous driving and costs $15k extra.
Yeah I'm not getting in a car driven by AI. The tech ain't where it needs to be for that.
It needs to be regulated to hold manufacturers responsible when their software isn't good enough. My understanding is that there already probably is enough regulation and government agencies just need to hold Tesla accountable.
Personally, I'm all for cars driven by AI iff it's better and safer than a human driver. Human drivers make a lot of mistakes and driving is the most dangerous everyday activity many people do. But if the AI isn't better than a human, that's a problem. I don't need AI drivers to be flawless, as that's an unrealistic bar. I just need them to be undeniably better than humans. Everything I'm hearing about Tesla's self driving is that they aren't.
Especially Tesla. I am very into computer vision research but I would never trust a vehicle that relies on only that with 0 LIDAR or other sensing technologies in place.
Great news. You don't have to. The other people around you moving thousands of pounds of steel will be relying completely on that, and there is nothing you can do to stop it.
Unfortunately you can still be hit by cars that idiots let the AI drive.
Thats why I drive defensively and look both ways before I cross a street.
I'm convinced that the self-driving AI from competent companies is better than human drivers already. The bar isn't perfection, the bar is the average driver, and the average driver is bad.
Having said that, I'd never get in a car with Tesla's self-driving solution. Musk polluted the term "full self-driving" to cars that definitely weren't.
No, the bar is accountability. If the FSD screws up either:
It absolutely does not matter that it's better than an "average" driver. If the best driver in the world screws up, they are still held accountable.
When a service is willing to take responsibility for collisions and driving violations, then we know it works. If the guy asleep at the wheel (which he allegedly can do in an autonomous car) is still the one held responsible, then were not there yet.
That said end-to-end AI totally sounds like equivocal marketing buzz.
When a service is willing to take responsibility for collisions and driving violations
Devil's advocate: it's kinda hard to pin the responsibility on Tesla when at the end of the day there was a person driving and the driver's always responsible.
I'm not disagreeing with you, I'm on team ban-human-drivers
Ideally, we'd get to the point where the driver merely directs the vehicle to where it wants to go, and then the computer system works out all the pathfinding and maneuvering, so that yes, any instance where a vehicle avoidably collides with another thing can be regarded as a malfunction.
I wonder what happens when the car is on a collision course with a golden retriever and the only way not to hit it would be to damage the car. Or same scenario, but the only way not to hit it, is it to hit an 07 Carolla parked on the side of the road. Not saying humans have superior judgement... just wondering if it will be programmed by the theory of actuarial of philosophical science.
That makes me think- will the AI see a kid that's about to run out from behind a parked car? As a human, if I see a kid run from the house into a row of parked cars, I know he's still there and will slow down before I get there. But would self driving make that same leap of logic? I'm not sure what the range and capabilities of self driving cars are right now in terms of scanning, but hopefully it would be smart enough to take preventative measures
I doubt it. Germany has already implemented (considering implementing) regulations regarding the ethics of autonomous vehicles. As it is, cars are simply trying not to collide with anything and given their reflexes and perception are way faster and more accurate than human beings, they have a better chance of saving both the dog and the other car.
That said, one of the problems we're seeing with smart devices (that is devices that are software run rather than controlled by simple mechanics) is that companies are keen to abuse the power that gives them, hence the whole John Deere tractors debacle and the development of right-to-repair laws. Also, some BMWs require rental of some of their features (such as seat warmers) which seems to me as less than ethical.
So I hope we'll get to a point where not only is it anyone's right to jailbreak their devices (including a self-driving car) but there will be several FOSS options we can choose from. And that means someone who programmed them may actually find a process-layer in which hazard prioritization or victim prioritization is considered.
It is certainly an entertaining idea of speculative fiction that an aggressive driver package is developed, gets popular and then causes a rise in traffic accidents. More likely would be software packages that allow the vehicle to operate despite self-test failures, again leading to a higher traffic collision rate.
There are many such hypothetical scenarios based on the trolley problem, but the real answer is that a good self driving system will never end up in that situation in the first place.
So as a dev, you just program to not let that situation arise, then you won't need to program a solution for that.
You wouldn't let chatGPT drive a car would you?
DriveGPT. Written by chatGPT proompted by chatGPT. Powered ny Nvidia^tm©®.
Gonna take the concept of AI hallucinations to a whole new level.
You mean LLM hallucinations?
Typical ML models are not hallucinating in the same manner.
"There was a ghost! This is ectoplasm!"
Not a fan of Tesla or Musk, but can we differentiate the broad public understanding of the term AI from machine learned control systems? People anthropomorphize the situation into thinking there is an I, Robot style driver enough as it is.
Counterpoint, though, maybe doing so encourages skepticism of Tesla's capabilities.
It's not AI it's VI.
Virtual intelligence is not artificial intelligence.
That's right, all the mistakes it makes are now powered purely by AI. Isn't that fantastic?
What do you think was powering it untill this point?
The current beta is a mix of ai algorithms and human programming. It is also a dead end now as they’ve abandoned that approach.
I guess so?
As a cyclist I really do look forward to the day where good AI is consistently better than the average-to-worst drivers out there; the bar is depressingly low and the stakes are high.
I write (and test) software for a living and my experience with Tesla as a consumer device is that it's many generations away from being something I would trust.
Also, I've seen what happens to product quality when management overrides its engineers in the way elon does- we get pre-alpha quality out there in the wild, being tested on a public that didn't sign up for that shit
That title and thumbnail are pure poetry.
Felon Must speaks again.