Comment by jdewerd | Hacker News Reader

Comment by jdewerd | Hacker News Reader

Parent Full thread

Should you remove the human and then expand scope (Waymo)? Or expand the scope and then remove the human (Tesla)?

I dunno. I personally think the Tesla approach is fundamentally better because it obtains more data, faster. Millions of miles per day rather than millions of miles per year, with better macrodiversity to boot. However, that's just a heuristic: we've got a Tortoise and Hare scenario and as the parable teaches us, it's not so difficult to imagine a scenario where slow and steady wins the race. Even though I favor the Tesla approach, I could see this going either way.

That said, I do think it's uncharitable to read failure into Tesla's high level decisions. They could have chosen the Waymo path and dumped all their effort into maximizing reliability on a small scenario subset. They chose not to. There are Andrej Karpathy interviews about it. They wanted to data-max. They data-maxed. Now they have 300x the data collection rate of Waymo. That's not failure, and it may even be a killer advantage. We'll see.

> They wanted to data-max. They data-maxed. Now they have 300x the data collection rate of Waymo.

Citation needed.

What Tesla lacks is quality data, which Waymo has due to their high fidelity sensors. Tesla has been "collecting data" for years without much to show for. They still have embarrassing failures so basic in nature [1], you know they're going nowhere. If all it took was data, they would have something meaningful by now.

[1] https://twitter.com/elaifresh/status/1727889381309239731

Tesla's data isn't just more volumnious than Waymo's, it's also higher quality. Macrodiversity is important. LiDAR isn't. Scene reconstruction got good in the last few years, and that's what LiDAR was competing with. Meanwhile, there's no substitute for macrodiversity.

Anyway, you asked for citations. Here they are.

Waymo:

> Waymo was the clear leader in 2022 with 2.9 million miles with safety drivers

https://www.eetimes.com/waymo-cruise-dominate-av-testing/#:~....

Tesla:

> FSD Now Drives 1 Million Miles Per Day

https://www.notateslaapp.com/news/1327/tesla-fsd-milestone-1....

> Graph showing 2M miles per day

https://digitalassets.tesla.com/tesla-contents/image/upload/...

> Tesla has been "collecting data" for years without much to show for.

Except, you know, millions of miles per day of self-driving in situations that Waymo isn't even planning to attempt for years.

Waymo has millions of driverless miles at this point. Tesla has zero. I'm not sure how Waymo's safety driver miles is relevant here.

> Except, you know, millions of miles per day of self-driving in situations that Waymo isn't even planning to attempt for years.

And what exactly have they got to show for it? All that data collection and they can't handle the easiest of situations in San Francisco. You don't need nationwide data not to run stop signs or swerve onto oncoming traffic.

Safety-driver miles are more useful for training. Driverless miles are more useful for proving. Like I said, different paths: remove the human then expand scope vs expand scope and then remove the human. Waymo is undeniably further along path 1, Tesla is undeniably further along path 2. Which path is shorter? I'm pretty sure it's path 2, but time will tell.

> All that data collection and they can't handle the easiest of situations in San Francisco

Unlike Waymo, they haven't limited to and exclusively trained on the easiest situations in SF for the last decade. Just like Tesla hasn't even attempted true driverless miles, Waymo hasn't even attempted unmapped back-alley navigation, let alone unmapped back-alleys in the middle of nowhere. You need scope and reliability to win this race, and it seems to me that the Tesla formula for improving reliability scales better than the Waymo formula for expanding scope. Time will tell.

> Safety-driver miles are more useful for training.

Waymo does most of its training in simulation. They do 20+ million miles of simulation every single day with 25,000 "virtual" vehicles. They don't even have to put out many cars for training.

> Unlike Waymo, they haven't limited to and exclusively trained on the easiest situations in SF for the last decade.

I've got a bridge to sell you if you think driving in SF is easy or that Waymo is only doing it in the easiest situations. They cover the entire city 24/7, and in rain and fog.

> Just like Tesla hasn't even attempted true driverless miles, Waymo hasn't even attempted unmapped back-alley navigation, let alone unmapped back-alleys in the middle of nowhere.

And they won't because maps are clearly a big part of their tech stack and they've found a way to efficiently maintain them. That's what gives them the confidence a.k.a reliability to remove the driver.

> You need scope and reliability to win this race, and it seems to me that the Tesla formula for improving reliability scales better than the Waymo formula for expanding scope.

So far, evidence suggests Waymo is expanding both scope and reliability faster than Tesla. They just announced airport terminal rides in PHX today. Community trackers (absence of Tesla sharing data) show Tesla's intervention rate hasn't really improved much at all and they still seem to make extremely basic mistakes. So the formula clearly isn't working as of today.

Tesla has simulations too, of course. Thing is, the world has fractal complexity and synthetic data only captures the variation that you tell it about. Simulations for synthetic data generation are literally my job -- this is where my intuition is at its strongest.

> if you think driving in SF is easy

I'm not here to trivialize SF driving, I'm here to say that it's a small corner of a big weird world that Tesla has done a much more thorough job of exploring. Google and Waymo are used to having a scale advantage, but Tesla's fleet utterly dwarfs them.

> Waymo is expanding both scope and reliability faster than Tesla. They just announced airport terminal rides in PHX today.

That's nice, but FSD already drives almost everywhere in the USA. It can usually even figure out poorly mapped rural roads in flyover country. What's Waymo's timetable for those?

> So far, evidence suggests Waymo is expanding both scope and reliability faster than Tesla.

FSD has gone from private to "mostly works everywhere in the USA" in the last 3 years. It has taken Waymo 10 years to start escaping the hubs one city and route at a time. They don't even use freeways, last I checked. Slow and steady is a fine strategy, I wish them luck in the marathon, but it's a bit early to call this one for Waymo.

I'm used to holding back the enthusiasm in Tesla circles where we are naturally predisposed to call the race early for Tesla, so this is a fun change of pace.

> I'm not here to trivialize SF driving, I'm here to say that it's a small corner of a big weird world that Tesla has done a much more thorough job of exploring.

Then explain to me why it can’t stop for stop signs in SF or can’t drive in a straight line without swerving onto oncoming traffic. Surely, they have a lot of data from the city with their large fleet? Either the data is helping or it’s not.

> That's nice, but FSD already drives almost everywhere in the USA. It can usually even figure out poorly mapped rural roads in flyover country. What's Waymo's timetable for those?

Yes, with a driver. And still with poor intervention rates after years of development. What’s Tesla’s timetable for getting their reliability up and removing the driver? That’s really the ballgame in self driving.

> FSD has gone from private to "mostly works everywhere in the USA" in the last 3 years. It has taken Waymo 10 years to start escaping the hubs one city and route at a time.

“Works everywhere” is doing a lot of heavy lifting here when you rely on a driver to prevent crashes. Literally does not pass the true test of self driving.

It took Waymo 10 years to completely remove the driver. Now they’re announcing new cities and capabilities constantly, with freeways coming soon. Let’s see how many “rewrites” it takes for Tesla to tell drivers not to pay attention, let alone completely removing them. I wish them luck (as a Model 3 owner), but they’re working with their hands tied behind their backs on this one.

As a vehicle user, and not a stock holder, there is only heuristic that will determine whether I get into either of their vehicles: Which one is more likely to kill me.

I'll take a ride in a self driving taxi through residential streets that might hit a dog.

I will not pay for a vehicle that will hard brake on a highway or drive under a tractor trailer.

Maybe after more deaths and data Tesla will pull ahead in that risk calculation but I will not pay to be part of that.

Of course. I feel the same way. Nobody is asking us to behave differently.

Reply on news.ycombinator.com