Hacker News Reader

Comment by Workaccount2 | Hacker News Reader

I have lost enormous money shorting Tesla over the years. I'm even short right now.

I can tell you though, Tesla is the only company doing two things:

Making full self driving car that doesn't need an expensive and expansive sensor suite. Just off the shelf cameras and a GPU.

Making full self driving that can drive anywhere, no pre-mapping needed.

Waymo probably could do without pre-mapping, Mercedes probably not. Right now though Tesla is the only player with a car that you can buy and will "self drive" anywhere in the country. The other car companies cars only work on select pre-mapped roads, and IIRC Mercedes only works on highways and over 40mph.

The bet against Tesla is that they will not be able to pull this off. Not that competitors will beat them to the punch. If Tesla did get FSD with cameras on par with Waymo, Waymo would likely be unable to compete.

View on HN

Part of me agrees with what you said, but my primary view is:

There's no indication that Tesla has solved the fundamental challenge of self-driving - making the system incredibly reliable and safe and able to handle myriads of rare situations. Tesla is at a point where multiple self-driving companies were years ago. The general experience from this field is that scaling is extremely hard because improving reliability and safety by orders of magnitude is extremely hard.

I think people are too hung up on the sensor suite. Once you solve the fundamental problem, a sensor suite is kind of an implementation detail. Waymo's system is robust to the removal of any sensor type / map. There's a good chance that Waymo in a camera-only mode would perform better than the Tesla robotaxi.

Mapping is not an issue, it's less expensive than people expect and you only need to do it once, so the cost per mile is approximately zero. And in Waymo's case, there's a lot of room for streamlining the process. The more you map, the easier and faster it gets. BTW, MobilEye has a very lightweight and crowdsourced mapping process, they're also camera-centric and have a cheap hardware package.

Waymo's long-term goal is a a system that is cheap and works everywhere. They've spent 15 years with a focus on solving the hard problem. Now they seem to be close to the finish line and they're shifting their focus and money into the easier stuff, optimizing hardware and expansion.

Mapping isn't a one time thing. You need to constantly be constantly updated. Bridges collapse, roads get washed out, lines get repainted and the list goes on. Even when you have it remapped regularly it can be wrong. Cedar road in Sonora California is a good example. They thankfully no longer have it connected, but they have a section of it mislabeled. If you look at the map, there's a section of Cedar road in the parking lot of Avalon Training. That's not actually Cedar road. Cedar road is actually the dirt road to the East of that. It used to be properly mapped to the dirt road. The problem with that is it quickly narrows and has major erosion. At one point there is a two foot drop off that is not safe to cross. Not sure how FSD would have handled it.

Map updates are not a problem, it probably takes something like 1% of the original work per year. The system is designed to be robust to map errors by the way.

This comment is incredibly wrong but I'll just say there is nothing easy about expanding from 2000 cars to several million.

It's relatively easy if you've solved self-driving. It's no longer a monumental research problem. Waymo has been scaling 5x annually since Oct 2020 and they say it was mainly enabled by increases in reliability.

They haven't been scaling by 5x annually. They had like 600 units and now they have maybe 2000

They have, in terms of paid rides per week. 100 to 250,000 in less than 5 years.

See this logarithmic chart: https://x.com/brianwilt/status/1827219050197610624

Ok. I'm talking about cars

When I said "Waymo has been scaling 5x annually" I was talking about paid rides.

clearly the rides are limited by the number of cars

I know, doesn't change anything about what I said. Cars were not the limiting factor. To sustain the 5x growth rate, they will need more cars, which is why there's over 2k in the pipeline and why they've recently built a factory capable of making tens of thousands per year after reaching full capacity.

They're also the only car company gleefully making die-hard political enemies around the world.

> Right now though Tesla is the only player with a car that you can buy and will "self drive" anywhere in the country.

But you can't, Tesla is very clear about this. It is 100% your responsibility to check everything the car does, and intervene if it's taking the wrong route, or doing something illegal, or doing something dangerous.

Both Waymo and Mercedes actually understand the limitations of their cars. It's not a problem, but an enormous point of confidence that they can say exactly in which conditions they can or can't operate safely (in certain localitirs for Waymo, on certain very limited types of roads and driving conditions for Mercedes).

The fact that Tesla is trying to do "self driving anywhere, anytime", and actually has 0 clue when it might or might not be actually safe, is a huge red flag, not a thing to be confident about.

> Making full self driving car that doesn't need an expensive and expansive sensor suite. Just off the shelf cameras and a GPU.

This is the other enormous problem with Tesla's approach. They are trying an approach that could only be as safe or safer than human driving if they make a superintelligent agent. This is unlikely to happen anytime soon, and impossible to work on a small GPU inside of a car anytime in the next 50 years (good luck running a multimodal ChatGPT in your car). Adding many more sensors that humans just don't have is the only way to get a leg up on humans cheaply and efficiently.

Reply on news.ycombinator.com