Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Andrej Karpathy leaves Tesla

twitter.com

Post reply on HN
Print view View original

281–290 of 441 posts

Re: Andrej Karpathy leaves Tesla

#281

Post by bambax »

Animats wrote: post #209
impulser_ wrote: post #8

Earlier quoted context omitted.

Yeah, but Google's vision + lidar tech also doesn't seem any better at solving it either. They have been working on this problem the longest and they aren't even confident enough to produce a product with it. Google is probably the leader in AI and AI research. They are also the leader in data and mapping. They have billions of cash to play with. Yet it seem like they haven't gotten any closer at solving this problem…

Waymo and Cruise routinely have driverless cars on city streets. In California, all collisions, however minor, have to be reported, and DMV posts them on their web site.[1] Most are very minor. Here's a more serious one from last month: " A Cruise autonomous vehicle ("Cruise AV") operating in driverless autonomous mode, was traveling eastbound on Geary Boulevard toward the intersection with Spruce Street. As it appro…

> This is the real version of the trolley problem

An autonomous vehicle has to protect its occupants at the expense of everything else (or, at the very least, appear to do so in a convincing manner), because otherwise no one will step inside.

(At the very least, if a machine is going to sacrifice me or my family to save a third party, I need to know what hierarchy it is following, and how it was decided and by whom.)

But what this incident seems to illustrate is that it's difficult for a self-driving car to share the road with human drivers, and behave like a human -- meaning, allowing human drivers anticipate what it will do.

I drive a motorcycle and a bike in Paris; the reason I'm still alive is because after so many years of this I can tell what all the other cars will do at all times, before they know it themselves.

But an autonomous vehicle that would behave so differently from a human driver as to be unpredictable, would be terrifying.

Re: Andrej Karpathy leaves Tesla

#282

Post by steveBK123 »

Barrin92 wrote: post #10
impulser_ wrote: post #8

Earlier quoted context omitted.

Yeah, but Google's vision + lidar tech also doesn't seem any better at solving it either. They have been working on this problem the longest and they aren't even confident enough to produce a product with it. Google is probably the leader in AI and AI research. They are also the leader in data and mapping. They have billions of cash to play with. Yet it seem like they haven't gotten any closer at solving this problem…

>Any reason why everyone seems to be stuck on this problem? ML maximalism focused on the narrow problem of 'solving driving' while not recognizing that any task as complex as driving requires probably something closer to general intelligence, and theoretically the field has been impoverished in favor of "throw more graphics cards at everything".

Agreed.

People point at statements like "it just needs to drive better than the average driver" and "well lots of idiots drive".

On the flip side, lots of intelligent people with other skills do not have the confidence & comfort to drive outside of their towns local roads, or at all.

There are problem sets for ML like quant trading or phone album image tagging where you just need to be right 51% of the time to make money / be novel & useful.

We've seen plenty of examples showing ML is not appropriate for life&death problems like being the primary reviewer of diagnostic imaging for cancer screening.

The bar for ML to drive a car is much much higher, and I'm not even sure we have the right sensor suites feeding into the models, not to mention the compute capacity to operate at low enough latency.

Re: Andrej Karpathy leaves Tesla

#283

Post by sdenton4 »

aidenn0 wrote: post #113
gruturo wrote: post #51

Earlier quoted context omitted.

> Elon in 2028: “time machine!” Well, to be fair, he only has to hit _that_ goal - at which point he can go back in time at his leisure and fix all the others. And he could hit even the time machine goal as late as he wants, and it won't matter.

This sounds like solid evidence for the fact that he won't ever get his hands on a time machine.

Or it'll be like the time machine in Primer, where you can't go back before the time machine was turned on.

Re: Andrej Karpathy leaves Tesla

#284

Post by kylehotchkiss »

sn41 wrote: post #246
dreamcompiler wrote: post #34

Earlier quoted context omitted.

If Musk actually said this he's an even bigger idiot than I gave him credit for. He makes the best cars in the world (obviously IMHO) and that's worth a lot more than 0. I couldn't give less of a damn if they can't drive themselves.

IMHO, the best cars in the world are still Toyotas - not the bleeding edge, mind you, just more reliability, overall "gets out of the way and does the job" in the most effective manner possible.

With the supply chains this wild, and increased costs everywhere else, I’m also happy to be a Toyota owner. A healthy aftermarket for parts too just in case. And we get CarPlay!

Re: Andrej Karpathy leaves Tesla

#285

Post by elcomet »

dreamcompiler wrote: post #242
elcomet wrote: post #134

Earlier quoted context omitted.

I disagree with you. On the contrary, I think it can only be solved by throwing more data, but much more data that we have available now, to get first a smart AI that has common sense about the world (a world model). Then we can start trying to drive in it.

The counterexample to your point is that brand new 16-year-old human drivers seem to do a pretty good job of driving with very little training data. Yes, they do kill people and they're much worse drivers than 40-year-olds who have much more training data, but the fact that 16YOs kill as few people as they do with so little training data means some other mechanism than raw data quantity is important.

It's not a counterexample. Your 16-year old human has millions of years of evolution into its brain. That's a lot of pre-training data.

Re: Andrej Karpathy leaves Tesla

#286

Post by misiti3780 »

moomoo11 wrote: post #82
dreamcompiler wrote: post #34

Earlier quoted context omitted.

If Musk actually said this he's an even bigger idiot than I gave him credit for. He makes the best cars in the world (obviously IMHO) and that's worth a lot more than 0. I couldn't give less of a damn if they can't drive themselves.

I ended up selling mine because the build quality started getting on my nerves. Misaligned panels that are "within spec" or interior rattles that make the ride completely unpleasant (at only 10k miles). The tech is all right, and I got to try auto pilot at a discount. Unfortunately the phantom braking made AP completely useless with passengers who would freak out and complain. However, when it worked it was quite nic…

you're in the minority ...

Re: Andrej Karpathy leaves Tesla

#287

Post by bambax »

Animats wrote: post #262
dekhn wrote: post #256

Earlier quoted context omitted.

I think the way I would say it is, after a few tens to hundreds of thousands of such events, we'll be able to estimate the parameters of the human preference functions associated with a non-fatal trolley problem (the real trolley problem is irrelevant to the car ML folks, they are not aiming to maximize global utility) and that will help guide our confidence in rolling out more. For all the woe and gloom in the news…

I've been reading those DMV reports for years, and there are clear patterns which repeat. One is where an autonomous vehicle started to enter an intersection with poor sight lines to the cross street. Sensing cross traffic, it stopped, and was then rear-ended by a human-driven vehicle that was following too closely. Waymo has that happen twice at the same intersection in Mountain View. There's a tree in the median st…

> Humans tend not to do that

AI should absolutely mimic the behavior of real (good) drivers.

Although that's not what you're describing here, another problem for AI could result from it knowing more than an average driver; for example, if a high-mounted LIDAR were able to see around corners and let the car decide it's "safe" to do a turn that no human would attempt for lack of visibility, that could cause problems.

(Also, it's surprising that an autonomous car doesn't detect that another car is following it too closely, and slows down appropriately in anticipation. How is this not taken into account.)

Re: Andrej Karpathy leaves Tesla

#288

Post by steveBK123 »

giantrobot wrote: post #190
espadrine wrote: post #43

Earlier quoted context omitted.

Waymo has superior performance based on their historical statistics. It makes sense, since their lidar sensors capture more of the environment, and directly in 3D. Their AI also seems better QA’ed. The Tesla AI Day[0] surprised me as it showed they only had a simple architecture for a very long time, simply feeding barely processed camera pixels to a DNN and hoping for the best with little more than supervised learni…

> After all, sure, we humans only take two video feeds as input… No, we use both our perception and proprioception when we drive or walk for that matter. We have two accute visual sensors mounted in a housing with six degrees of freedom and a sensor feedback mechanism. We also have fairly sensitive motion and momentum sensors all throughout our bodies. Additionally we have audio sensors feeding into our positional mo…

All the vision-only maximalists also ignore how insufficient Tesla cameras are compared to your human eyes in terms of resolution, dynamic range and refresh rates.

If you ever review your Tesla Dashcam footage you'll see that you can rarely even make out a license plate. The cameras are not even HD, let alone UHD.

Further the refresh rate is a paltry 36fps.

Drive the car at high noon or night and you realize how poor the dynamic range is with highlights and/or shadows get lost.

Re: Andrej Karpathy leaves Tesla

#289

Post by elif »

codechicago277 wrote: post #3

Tesla has bet the company on robotaxis, but their vision only tech stack doesn’t seem capable of solving it, which is a problem because Tesla has repeatedly promised FSD is right around the corner, or less than a year away. It’s hard to believe Karpathy would step down if he felt they were close to solving the problem anytime soon. This announcement comes after a 4 month sabbatical where Karpathy said he wanted to ta…

FSD beta tester here. I think they are minimum 3 years away from anything exciting beta.. but the localization, mapping, and visualization are not the reason. I don't think LIDAR would contribute substantially to improvement.

The fundamental flaws are in the decision-making being based upon 10-30 second feature memory, ignoring features outright, and only depending on visible road features instead of persisted map data.

For instance, near my house there's an intersection where it will try to use a turn-only lane with a red arrow when it's trying to go straight thru a light. 100% of the time. Even if I'm in the correct lane with no traffic around.

That's because the turn arrow on the ground is worn off. It is not a perception problem, it is

a) ignoring the obvious red turn arrow signal's significance for lane selection deliberately b) makes no attempt to persist or consult map data for 'which lanes go where' c) it completely disregards painted lines on the ground in the "no drive here" striping.

Also one block from my house, FSD will stay still indefinitely waiting for trash cans (displayed as trash cans) to clear the leftmost lane so that it will turn left.

None of the failures I encounter are due to lack of perception.

Re: Andrej Karpathy leaves Tesla

#290

Post by pilaf »

harles wrote: post #237
codechicago277 wrote: post #3

Tesla has bet the company on robotaxis, but their vision only tech stack doesn’t seem capable of solving it, which is a problem because Tesla has repeatedly promised FSD is right around the corner, or less than a year away. It’s hard to believe Karpathy would step down if he felt they were close to solving the problem anytime soon. This announcement comes after a 4 month sabbatical where Karpathy said he wanted to ta…

Monocular vision only seems pretty clearly not capable of solving the problem. Stereo/multi view systems have a shot (humans are proof), but Tesla bet against that long ago. I wonder what could’ve been in a proper multi view setup.

Humans are perfectly capable of driving on racing simulators using a flat screen though, where binocular vision makes no difference.

And Tesla cars have more than one camera on them. The front-facing camera is actually an array of 3 cameras (the two farthest ones are at about human eyes distance), but they're also equipped with forward and rearward looking side cameras, and back cameras.

I think Tesla underestimated how hard vision-only FSD is, but having a single camera (they don't) is not the reason.

Post reply on HN