Hacker News Reader

Comment by andrewmcwatters | Hacker News Reader

George’s insights on self-driving cars remind me of when you get the rare exceptional engineer in your company who knows 100% what he or she is talking about down to the exact implementation level and all of the alternatives, and leadership totally ignores them because they foolhardily believe this kid can’t possibly be right—without asking engineering questions to test to see if they’re right about the engineer being wrong or having not considered something.

Except George seems to have been right to the tune of others spending billions of dollars while this dude hacked on free and open source software.

His biggest bootstrap expenses were probably commodity hardware, before dealing with business purchase orders.

I find that hilarious and extremely gratifying.

View on HN

His insights on self driving cars are total fluff once you realize he was only working on problems several orders of magnitude smaller in scope than the ones spending billions are attempting to solve. Autonomous driving is more than just making lane-keep on highways work with a camera mounted to the windshield and requires more than just "hacking".

His insights do indicate one thing: that he doesn't understand the rigor required to engineer safety-critical systems.

I mean you can say that all you want, but the proof of the self-driving pudding is in the driving, and SAE Level 2 is the most practical self-driving we are ever going to get.

I don't think there's anything wrong with trying to refine that approach as much as openpilot did, and honestly if there was fruit in SAE Level 3 and above, I feel like we would have seen it by now, but we haven't.

I have no idea what you mean by "problems several orders of magnitude." Driving is driving. Your hardware and software can do it, or it can't. What do billions of dollars have to do with anything?

Should people remind you Meta spent billions of dollars on a less popular Second Life?

How would you explain to shareholders that your solution can't compete with openpilot?

> SAE Level 2 is the most practical self-driving we are ever going to get.

So… how much would you bet on this?

$10,000 USD in 2022 dollars that SAE Level 4 is not available in standard economy cars in the next 10 years. My 2022 Corolla LE, a base model compact car, drives itself just fine (SAE Level 2) even on city streets. It would do it better with openpilot, though.

It makes Tesla Full Self Driving look like a joke, in my opinion.

> SAE Level 2 is the most practical self-driving we are ever going to get.

The bet is on this though -- what's impractical about having L4 fleets in major cities that replace Uber/Lyft? Just because you can't personally own one doesn't mean that in 10 years they can't be pretty ubiquitous and practical, particularly for city dwellers, which is most people. I'd bet that your statement is wrong on this basis. L4 cars already exist and are being used in large portions of San Francisco today.

And this is to say nothing of trucking & delivery, where self driving is in some ways easier to solve and a lot of progress is being made.

Well, not everyone lives in San Francisco, and no one wants to take an Uber literally everywhere. So, like most people, I want to own my self-driving car.

And I personally don't think the industry is mature enough to accept the realities and limitations of the technology we work with to just say, "Hey, actually, we can let attentive drivers basically be hands-free for 80+% of their drives."

And to me, that would be great.

Not everyone wants to own a car either, and everyone owning a car is not exactly environmentally sustainable. Replacing Uber & Lyft with robots would add a ton of value and be practical for a very large percentage of the population.

> And I personally don't think the industry is mature enough to accept the realities and limitations of the technology we work with to just say, "Hey, actually, we can let attentive drivers basically be hands-free for 80+% of their drives."

Uh isn't this what Tesla, GM, Mobileye, Comma, and many others are doing? Basically every player in "the industry" wants this & is working on it -- FWIW I think L2 is fine, but there need to be more safety features because the number of folks who've slept while their flawed L2 system drove them is truly alarming. I'm less worried about their safety and more worried about the folks they might crash into.

Even a "selfish driving car" would be a major step up - I'd think about not owning a car if I could order one on an app and have it at my doorstep in 15 minutes; even if after I get into it I take over and drive it. Self-delivering rental cars, if you will.

> Not everyone wants to own a car either, and everyone owning a car is not exactly environmentally sustainable. Replacing Uber & Lyft with robots would add a ton of value and be practical for a very large percentage of the population.

Yeah, no one wants this.

You speak for yourself. Uber and Lyft demand in large cities tell another story.

I'm also mystified: do you really think personal car ownership for every person is at all sustainable?

SAE Level 4 driving is several orders of magnitude harder than SAE Level 2. You are either good enough to drive without requiring a driver or you are not. There is a reason developing L4 systems cost billions of dollars.

George Hotz thinks Comma does the job with only a fraction of the spending compared to the likes of Waymo, Cruise, Zoox, but fails to realize his product also does only a fraction of what the others are trying to do.

"Several orders of magnitude" by what dimension? Engineering labor? Hardware expenses and testing? You can't just throw out a dimension of scale and say, oh it's so-and-so hard that's why it costs billions.

If you were trying to build L4, you'd hire a small team of exceptional engineers, and spend a few thousand on hardware tests. Billions doesn't come close.

You'd try and get Level 2 to be as satisfactory as possible, and see what percentage of the time you could achieve Level 3.

> If you were trying to build L4, you'd hire a small team of exceptional engineers, and spend a few thousand on hardware tests. Billions doesn't come close.

Not sure if you are aware that the self driving (robotaxi) companies have spent tens of billions of dollars in the last decade. R&D expenses are huge, building custom sensors and vehicles is hard and fleet maintenance/operations is costly.

>and spend a few thousand on hardware tests

Okay, that should let you just about afford to certify an Ethernet cable.

Hotz’s product should be compared to what Waymo can do right now, not what it’s trying to do.

The point is Hotz's product shouldn't be compared to Waymo at all. One is driverless and the other is not. It should be compared to the likes of Tesla Autopilot or driver assist systems offered by other OEMs.

Different strategies to bring the same core tech to market. Absolutely comparable.

The difference in capabilities makes the core tech very different.

That's a bold claim against fully-autonomous driving. Does operating a motor vehicle require qualia? If not, what else makes a human uniquely capable of doing it? It may be a hard problem to crack, but asserting that it's impossible is pretty unfounded.

You're not asserting that everything that doesn't require qualia can be done by a machine, are you?

That seems like a reasonable assertion to me. Not that it can be done by a machine today or in ten years but eventually.

Not even geohot is asserting that "SAE Level 2 is the most practical self-driving we are ever going to get", his assertion is that he thinks nobody is going to get there in the next 10 years.

He may be right that we won't get there broadly for 10 years, but I think that's a rationalization of how things have gone for him, and the fact that the methods he has worked on don't have a chance of getting there, rather than the thesis he had when he started comma. Because if this was your thesis, there was really no reason to get into the space when he did.

In particular, Waymo One is already an L4 driver. It operates completely autonomously and passengers do not have the option of taking over. I don't know how often they get stuck and have to call someone to do roadside service, etc, but clearly in some areas it is already feasible. So the question is whether this is possible in dense urban areas or not.

And if you're just looking at SAE 2, the main competitors in this space, Mobileye and Tesla are doing way better than comma.

I definitely thought progress would be faster here, but I don't think we're anywhere near as far away from useful systems as geohot argues.

You don't have to tell me about Waymo. I live where they operate. I'm one of the few people in the United States who can actually use their service. They limit where you can go by an approximate 4x9 city block area, and actually less than that, I'm being generous.

You have no idea exactly how limited the service is. Effectively, you can't go ANYWHERE in the Phoenix metro area unless your A-to-B destination is squarely in that region. That's not Level 4. Whatever pretend system they've got going on, I'm sure works perfectly on those particular pristine roads, but no one is using Waymo.

There are no real, coast-to-coast Level 4 systems.

For a grading analogy, an area like that is level 4 minus, but "coast-to-coast" is level 4 plus. You don't need coast to coast to be "real". If they manage to cover 90% of a metro area, I'd be happy to call it a "real" level 4 service. Right now I'd want an explanation for why they're not adding blocks all the time.

I am aware of the limitations, but covering some predefined area with no human intervention is the definition of L4.

Clearly there is non-zero development risk here, but I don't think it's going so badly that it's definitely going to take 10+ years to launch in a metro area with millions of people.

It may take a decade to get to a coast-to-coast system, but there's not a whole lot of incentive to build that vs serving a few large metros in the near future.

> I have no idea what you mean by "problems several orders of magnitude." Driving is driving. Your hardware and software can do it, or it can't. What do billions of dollars have to do with anything?

We've had the technology for level 2 for 20 years. You just need to track the painted lines and track the distance to the car in front of you. Higher levels are orders of magnitude more complicated.

If you think level 2 is the best we'll get, that's fine, but that's giving up. It's not what people mean by self driving.

> We've had the technology for level 2 for 20 years. You just need to track the painted lines and track the distance to the car in front of you.

Considering many vehicles still do this inconsistently or poorly I’m not sure that’s an encouraging sign.

> he was only working on problems several orders of magnitude smaller in scope

Isn't that kind of the point? Tesla would have L4 if the architecture was built right, but it wasn't, and now they're scrambling.

I think George made it very clear that the architecture is what these companies spending billions are getting wrong. Tech debt is more expensive than you think at the cutting edge.

You know what is total fluff? Tesla FSD. Disgusting. At least comma is honest about their hardware, and that's about all it comes down to.

> I think George made it very clear that the architecture is what these companies spending billions are getting wrong. Tech debt is more expensive than you think at the cutting edge.

And yet the architectures he disapproves are the only ones who have shown fully autonomous driving is possible. There are driverless robotaxis running now, albeit in small areas but expanding.

> You know what is total fluff? Tesla FSD. Disgusting. At least comma is honest about their hardware, and that's about all it comes down to.

I fully agree.

> And yet the architectures he disapproves are the only ones who have shown fully autonomous driving is possible.

They did not.

> There are driverless robotaxis running now, albeit in small areas but expanding.

Running in small areas shows that they can drive in small mapped out areas under specific conditions. Not that we have fully autonomous self driving.

> Running in small areas shows that they can drive in small mapped out areas under specific conditions.

That is fully autonomous driving. It's just classified as Level 4. Geofences and operating conditions can be extended to provide a real, useful taxi service to a lot of people.

Not really. Like Karpathy mentioned in the latest lex episode, keeping an updated exact map of any area isn’t a scalable way to achieve this (which is what these geofenced services do) and isn’t how humans are able to drive.

It’s perfectly scalable and it’s a solved problem. Maps aren’t the bottleneck in scaling. The cars don’t even require an always up-to-date map to work safely (but they are self mapping anyway). I’m not sure how long Karpathy will push this narrative because Tesla isn’t able to afford mapping like the others.

Tesla also uses maps, it’s just less detailed. If humans are able to drive without maps, why does Tesla require a map of lane geometry, stop signs and intersections?

> keeping an updated exact map of any area isn’t a scalable way to achieve this (which is what these geofenced services do) and isn’t how humans are able to drive.

Humans combine low-res maps with extremely good visual odometry & spatial reasoning. Computers don't come close to the way humans reason about space & vision yet, so stronger priors help to make up the difference.

Making it scalable to keep an updated exact map of an area is part of the work L4 companies have been doing. I think it's hard for folks on the outside to really know just how much progress has been made here. Before Google Maps existed folks didn't think you could scalably map the world either.

It's not like companies like Cruise or Waymo want to keep doing this if they can help it. It's pretty disingenuous to say that these companies don't have any incentive to jump off the mapping wagon as soon as the tech to do so exists.

In fact, L4 companies do have to deal with areas they're driving where they see the map is outdated (often these are construction zones). So either they have mapless driving capabilities to handle these cases, or they do something else (seems like Cruise cars love pulling over en masse, lol)

I'm curious what his insights were? Are those collected somewhere?

Asking as someone that has largely not been following the field. Not challenging, genuinely curious.

https://blog.comma.ai/a-100x-investment-part-1/

https://blog.comma.ai/a-100x-investment-part-2/

It's basically a bunch of apples-to-oranges comparisons that don't make sense.

I'm not entirely clear I get those blogs. :(

If you are saying that the blogs are a bunch of apples-to-oranges comparisons that don't make sense, I think I fully agree on that.

Here's a rough summary:

1. There is no point in making human understandable layer between perception and planning. Mapping your surrounding with AI and then feeding it to handwritten planner isn't going to achieve real selfdriving.

2. To make AI you need big data, which means big fleet. Therefore you can only succeed if you have customers that are wiling to pay for you current version, because this way you gain money for each extra car in your fleet, instead of losing it.

3. Robotaxis are bad business. From a customer standpoint taxi with a driver is a selfdriving car, replacing driver with AI will make drive longer and more expensive. The path for self-driving company is to make driver assist and go from there to L5.

4. Vision only, AI powerful enough to be level 5 is powerful enough to retrieve depth from image.

Some of that seems straight forward. That said, I question point 4. "Vision" is doing a lot of heavy lifting in robotics. Everything is sensor reading, such that there is no reason to restrict the sensors to the same visible spectrum that humans have. The only advantage of doing that, is that you can easily get data labeled by a human.

On the first point it looks like Tesla's approach to fixing bugs is much easier compared to Georges approach to fix bugs which is to modify and retrain the model until there are no regressions. I don't understand how this can cover every possible scenario. I am not in to ML and Tesla's approach seems more sense to me. Can anyone in the ML field comment on why George's approach might work.

> spending billions of dollars

This is usually the problem; once you're in the many millions of dollars you find that other things are way more important than actually being right or even delivering a product. Time and time again the story is told.

Reply on news.ycombinator.com