Hacker News Reader

Comment by iw7tdb2kqo9 | Hacker News Reader

Tesla had more chance to succeed than Waymo. It's impossible for Waymo to match Tesla in real world training and testing data. Somehow Waymo is winning.

View on HN

Waymo chose an approach that seems geared towards actually doing the task.

Tesla tried for the moonshot - They wanted a consumer car with cheap sensor hardware to perform the job. Trusting that computing "smarts" could solve the rest of the problem.

I'm in Atlanta where Waymos have started popping up left and right - the sensor bank on these things is HUGE. You can spot 'em from way far off... Giant sensors on top. Big sensors on the front wheel wells back wheel wells, big sensors on both front and back. Big sensors basically all over them.

I'm of the opinion now that Tesla was just way, WAY off base about what sort of requirements exist for sensing, and that they don't, in fact, have much more real world training data because their data is just garbage from the cameras.

Waymo is winning because Waymo accepted the actual requirements early. Tesla is off in lala land with a dead end solution. Lots of great marketing from tesla... but very little progress now in years. They really seem stuck in a local optimum with the camera-only approach, and it's not close to delivering the promised experience.

> consumer car with cheap sensor hardware to perform the job

Except, they took even that out because a certain someone leading the company thought they don’t need the sensors. It’s like someone trying to figure out how to make a bicycle balance itself and they decide to take the wheels off.

Waymo is winning by losing money as fast as they can?

Have you looked at their financials? They're terrible.

Waymo has ~40 million miles of test data. Tesla has ~4B miles. But for training both quantity and quality are important. Tesla only uploads excerpts and metadata for its 4B miles, and it's video only. Waymo can fully analyze all 40M miles, and it's much richer, with LIDAR and other sensors.

My feeling is that Waymo has the data advantage.

You can also just buy camera data from other OEMs, not to mention the vast amounts available from public sources like YouTube or Google maps vehicles that are accessible to them if they really care. The vast, vast majority of that data is uninteresting though. It's not much better than simulation data, which Waymo has many tens of billions of miles worth.

I’ll trust Waymo’s 40 million over Tesla’s 4B any day. You can make something work most of the times with our vision. But when there is no one to man the car, most of the times isn’t good enough

Waymo has already lapped Tesla. They sell 250,000 rides per week and rising, for a total well over 10 million. That compares to precisely zero rides sold commercially by Tesla. It's possible Tesla could catch up, but first they need to stop losing ground and develop technology that can actually compete with their leading competitor. Plus, right now the Tesla brand is synonymous with the most obnoxious blowhard in America. Meanwhile the Waymo brand is on the way to becoming the "Kleenex" of self-driving technology.

Waymo has been "scaling" down the Peninsula from SF for about 18 months now, and it's still not generally available.

Tesla can scale by simply adding vehicles.

Scaling by applying for expanded operational areas is how all deployments in California have to work, by law. Tesla avoids that by not operating robotaxi fleets in California and not submitting applications or mileage data for the vehicles they have there.

> Tesla can scale by simply adding vehicles

Between Elon having personally alienated voters in cities and Tesla's continued reliance on remote drivers, Tesla's ability to produce these vehicles seems low on the list of scaling constraints.

And I can scale my poop by pooping again. The question is whether Tesla has a thing worth scaling, and the answer to that question is "no".

But, can you shit at web scale? Elon can.

Scale what? Tesla does not have and has never had a commercial self-driving taxi service.

June 17, 2025: "Ride to more places in the Bay and LA. Starting today in SF, with new areas coming to LA later this week. Download the Waymo One app to see our new service areas."

https://x.com/Waymo/status/1935012059806061025

Tesla could have more camera data in sum (that's not even clear - transmitting and storing data from all the cars on the road is no easy task - L4 companies typically pysically remove drives and use appliances to suck data off the hard drives), but Waymo has more camera data per car (29 cameras) and higher fidelity data overall (including lidar, radar, and microphone data). Tesla can't magically enhance data it didn't collect.

This is a crippling disadvantage. Consider what it takes to evaluate a single software release for a robotaxi.

If you have a simulator, you can take long tail distribution events and just resimulate your software to see if there are regressions against those events. (Waymo, Zoox)

If you don't, or your simulator has too much error, you have to deploy your software in cars in "ghost mode" and hope that sufficient miles see rare and scary situations recur. You then need to find those specific situations and check if your software did a good job (vs just getting lucky). But what if you need to A/B test a change? What if you need to A/B test 100 changes made by different engineers? How do you ensure you're testing the right thing? (Tesla)

And if you have a simulator that _sucks_ because it doesn't have physics-grounded understanding of distances (i.e. it's based on distance estimates from camera), then you can easily trick yourself into thinking your software is doing the right thing, right up until you start killing people.

Another way to look at it is: most driving data is actually very low in signal. You want all the hard driving miles, and in high resolution, so that you can basically generate the world's best unit testing suite for the software driver. You can just throw the rest of the driving data away -- and you must, because nobody has that much storage and unit economics still matter.

This is to say nothing of the fact that differences between hardware matter too. Tesla has a bunch of car models out there, and software working well one one model may not actually work well on another.

The Tesla approach was kind of magical thinking; put enough data into the magic box, and eventually a perfect driver will pop out. In some ways it's similar to the "just make a big enough LLM and it will be GAI" thing (a model which has pretty much fallen out of favour even with the most relentlessly optimistic LLM vendors in favour of so-called reasoning models).

In reality, just shovelling data into a box is not sufficient.

"Waymo is winning". One wonders why most people never see any of these videos but only hear about how bad Tesla FSD is?

https://x.com/niccruzpatane/status/1928477936845226469

https://x.com/TeslaCamera/status/1929167075731239226

https://x.com/friscolive415/status/1881181885445063041

https://x.com/LAMultiBroker/status/1885370114054512921

https://x.com/ananayarora/status/1808679192432927153

https://x.com/niccruzpatane/status/1929590206488883246

https://x.com/TheTrailerDan/status/1934618081881370766

https://x.com/JeffTutorials/status/1778188663253574035

https://x.com/ScannerPacific/status/1935016023536844960

https://x.com/greggertruck/status/1864836542402797812

https://x.com/neil_csagi/status/1803229926033858746

I could post these all day ...

Reply on news.ycombinator.com