[eloff]

I used to think that, but now I agree with George Hotz. Tesla uses machine learning to map inputs to control outputs. Google is using it just to build a model of the world and then use program logic to decide the control action to take.

The problem with the latter approach is the long tail of real life variety doesn't lend itself to a fixed set of programmed rules. You'll always have more edge cases.

I think Waymos approach gets you to 90% faster, but then plateaus.

[jdashg]

I think it's the other way around: Direct mapping of inputs to outputs will work great for a subset of conditions, but the last N% is infeasible without explicit logical programming.

[eloff]

You'll never run out of edge cases if you have to explicitly program the logic. At some point you have to let the machine learning decide it. Maybe there is room for a hybrid approach, but I think anything relying on programmers having thought of every situation ahead of time is doomed to fail.

[jdashg]

I believe relying on machine learning is doomed to fail because I don't think it can feasibly offer the robust safety guarantees required.

I honestly don't expect go-anywhere L5 in the next decade at least, and I definitely expect an engineered approach to hit L4 first.

[eloff]

You're wrong about the requirements. It doesn't have to be perfect, just better than us, and that's easily doable for computers that don't get tired, distracted, or drunk.

I'm not sure how long it will take, but I'm sure we'll get there.

[persistent]

Not sure there is enough evidence in support of those claims about Tesla. Tesla doesn't collect anywhere near enough data from its customer fleet to support the kind of massive training being claimed. In contrast Waymo and Cruise vehicles are regularly in the depot where each car offloads terabytes of data. Who has the bigger training set?

[revscat]

Tesla routinely gets data from the fleet and analyzes it. Here is Andre Karpathy speaking on this subject:

https://youtu.be/Ucp0TTmvqOE

1:49:30. I’m on mobile otherwise I would link to the exact spot.

Tesla has tens (hundreds?) of thousands of vehicles on the road right now that are continuously feeding training data back to Tesla. What evidence is there that this is insufficient, or even deficient? From what I can gather, Waymo has limited testing geofenced to Phoenix, Arizona. Tesla has cars all over the world with autopilot, all providing training data.

I don’t see a reason for your conviction that Waymo has an advantage here.

[losvedir]

Unless Teslas are uploading gigabytes every night over your home WiFi (do they?) they're limited by a 4G connection.

So you have a lot more Teslas sending a lot less data each. I'm not sure which company has the richer training set.

[jodrellblank]

they're limited by a 4G connection

Are they?

- they could store data to be collected in bulk when the car is in the garage for maintenance or servicing or repair.

- Superchargers could include fast data transfer in their cable for cars using them.

- Superchargers could include a wireless data collection from cars sopped at or near them.

- Cars could connect to public wifi to upload data at any car park, any place, any time.

- Tesla could have vehicles driving around, like Google mapping cars, collecting data wirelessly from nearby Teslas.

I don't know if they do any of them, but none are unreasonable things for companies of 2019 to think about doing.

[tigershark]

Hundreds of thousands. The 200k mark only in US was achieved more than 1 year ago, and it triggered the phase out of the incentives. For the last couple of quarters Tesla has been delivering almost 100k vehicles so it won’t be that long to reach the 1 million mark globally.

[xvector]

On top of this, Waymo has the full force of Google’s CAPTCHA behind it.

It is already full self driving in some cities, whereas my Tesla cannot even detect stop signs.

I love my Tesla but I harbor no illusions about the eventual victor of this self-driving race.

[macspoofing]

>Tesla uses machine learning to map inputs to control outputs. Google is using it just to build a model of the world and then use program logic to decide the control action to take.

I suspect that isn't true in so far that engineers will build what works. Building a complicated system like an autonomous car will necessarily involve a messy combination of a bunch of different strategies and with the result converging on a common solution.

[eloff]

I don't know if it's still true, but that was the state of things as of a year ago.