FSD v12.x (end to end AI)

[AlanSubie4Life]

We don't know, and while we can see V12 doing amazing things in Omar's videos

Are we not watching the same videos? All the videos I have seen of any significant length are strewn with errors and situations where the vehicle is driving other drivers nuts. It also gets in the wrong lanes routinely, moves at the last moment, etc.

The primary places I see improvement are how quickly it gets up to speed (something that could be done with v11), and an improvement of smoothness - though there are still likely hesitations - it takes an extremely keen eye and careful watching of the video to detect the hesitations and I just don’t have the patience. (It is best for the driver to call them out in real time which of course Omar does not do.)

To me it looks a little better than v11 - if you don’t count the regressions. But that could be due to how v11 is tuned, and it is a low bar in any case.

I expect continued improvement over v11 when/if v12 is released.

 

[Daniel in SD]

Knowing Elon's tendencies, it will be respectively:
Completely FuSD Up and Partly FuSD Up

I'm hoping it doesn't take too much longer for me to get Completely FuSD Up.

Full Self Driving Beta = L2 (secretly beta versions of L3-L5)
Full Self Driving = L3
Fuller Self Driving = L4
Fullest Self Driving = L5

 

[JHCCAZ]

Full Self Driving Beta = L2 (secretly beta versions of L3-L5)
Full Self Driving = L3
Fuller Self Driving = L4
Fullest Self Driving = L5

On your scale I want to be at least Fuller of it.

However, to be a bit more technically clear, I think the only reasonable goal is L4 (meaning situation-based more than geofence-based), so to me Fuller Self Driving would still qualify as Completely FuSD Up for product tier purposes. Going beyond that will remain aspirational with asymptotic progress.

 

[drtimhill]

I don't think the videos are edited in any way. I watch the 1x videos, and in many videos, Omar narrates or has a guest, and unless Omar loves to waste time, I doubt he's re-recording multiple narrated videos and only uploading narrated videos that were "successful"

You can claim he omits the major flaws, but otherwise, I don't think you can claim he edits his 1x videos

Even Omar would admit he picks the video that are more successful than others. (He has mentioned that he is seeing more disengagements done by the car, which matches what he was told about this release being dialed in to highly sensitive.) You simply dont see the really bad ones at all. However, overall V12 is showing great promise even from this sample set.

 

[Daniel in SD]

FSD beta has always been smooth.

https://twitter.com/x/status/1421378318155223041

https://twitter.com/x/status/1427116875494813697

https://twitter.com/x/status/1436838732326338563

https://twitter.com/x/status/1447476012694990852

https://twitter.com/x/status/1460430950651756544

https://x.com/WholeMarsBlog/status/1472153673627148295?s=20
https://x.com/WholeMarsBlog/status/1483563612123774978?s=20

 

[drtimhill]

Full Self Driving Beta = L2 (secretly beta versions of L3-L5)
Full Self Driving = L3
Fuller Self Driving = L4
Fullest Self Driving = L5

You forgot the inevitable final stage:

Bloated Self Driving = L6 (where the car refuses to drive you home if you aren't polite to it)

 

[drtimhill]

I've said this many times on this forum, and here we go again:
If autonomy was solved by "more data" and "more compute" it would have been solved five years ago.

That's the old "if this were a good idea someone else would already have thought of it and done it" non-argument.

 

[spacecoin]

That's the old "if this were a good idea someone else would already have thought of it and done it" non-argument.

No it isn't. Waymo has a fraction of the data that Tesla has, yet drives 2000x safer and also better. How come?

Future preview: GPT-5 won't be meaningfully better than GPT-4. Diminishing returns. Same difference.

 

[zoomer0056]

No it isn't. Waymo has a fraction of the data that Tesla has, yet drives 2000x safer and also better. How come? GPT-5 won't be meaningfully better than GPT-4. Diminishing returns.

Good question. What's the true real (ha, I like how people talk) answer?

 

[spacecoin]

Good question. What's the true real (ha, I like how people talk) answer?

Engineering, talent and a 10 year head start on designing safety critical systems?

 

[powertoold]

Waymo is incredibly primitive compared to V12, that's the real answer

It's basically a micromanaged train

 

[spacecoin]

Waymo is incredibly primitive compared to V12, that's the real answer

It's basically a micromanaged train

That's the Tesla-stan FUD. Reality: There is not a single aspect of Waymo's system that isn't at least 10x more advanced than what Tesla has. Cameras, Radar, planner, safety, object detection et.c.

Now I'm going to watch the game! Have fun!

 

[JulienW]

.....
Fullest Self Driving = L5

L5 = IMPOSSIBLE (that bar is way too high for anyone for a looooooog time)

 

[powertoold]

L5 = IMPOSSIBLE (that bar is way too high for anyone for a looooooog time)

There's also not a good business case for true level 5. It exposes the service to more liability. There are plenty of people who would intentionally try to break a level 5 service.

 

[sleepydoc]

How? There's no legal definition, or meaning, for the word.

There is a definition, just not a codification. There’s a difference, and the lack of codification does not preclude using a definition for legal purposes so your statement makes no sense.

 

[sleepydoc]

I've been the driver in an airbag deployment. While the deployment did rip the watch off my wrist, that was the only issue I had. if you watch the way they deploy in the slo-mo, they tend to bypass the steering wheel and actually press your hands DOWN onto the wheel somewhat (that is, the bas is pressing back down ONTO the wheel). However, for all that I was able to walk away from a very nasty crash thanks to the airbags.

I actually looked at several slo-mo videos before making that post just to see if I could discern any patterns/potential dangers with hand position or otherwise. What I saw was the deployment pattern really depends on how it's configured in the steering wheel. I didn't appreciate the phenomenon you describe, but I was looking more at the initial expansion pattern. What you describe makes sense, though.

Regardless, I'm glad you're ok.

Yikes that would be a pucker moment. Now imagine the same event with your hand on the center horn at deployment. Airbags are an incredible invention with incredible power to make them effective. Glad you survived your event.

Clearly, Elon had all our interests at heart when he made a button for the horn!

 

[Usain]

Nothing of what you're saying is true.
We haven't "learned a lot" in the last five years. The last major break-though was in 2017. Is there anything in particular you are referring to?

"And computing power for training has increased exponentially" doesn't mean anything, but no, it hasn't.

How do you reckon Waymo was outperforming what Tesla has today five years ago, by a factor of 100x or so in terms of reliability... What Tesla has is undeployable as unsupervised, and it's been about the same for 2+ years. v12 won't likely change this.

https://twitter.com/x/status/1085944196186304512

Elons brother said "perhaps in five years, definitely this decade" in a recent interview, which I found funny. so perhaps 2029, definitely 2030?

Edit: Link:

Elon Musk's brother Kimbal is not as optimistic about Tesla's Full Self-Driving

Elon Musk’s brother Kimbal, who is also a board member at Tesla, is not as optimistic as his brother about...

electrek.co

Everything I said was true. You are just being a hater.

We have "learned a lot". Even if you just judge by the number of NN papers published in the last 5 years, it's "a lot".

Just look at how much Tesla's training compute capability has grown in the last five years and will continue to grow in the next several of years. It's clearly an exponential curve.

And to my main point, Tesla could not have done V12 five years ago.

 

[Usain]

That's the Tesla-stan FUD. Reality: There is not a single aspect of Waymo's system that isn't at least 10x more advanced than what Tesla has. Cameras, Radar, planner, safety, object detection et.c.

Now I'm going to watch the game! Have fun!

Like @powertoold said, Waymo is basically a micromanaged train.

Drop off a Waymo vehicle to my driveway tomorrow. It won't even be able to drive itself around my block.

That's not too advanced. It's kind of pathetic.

 

[drtimhill]

No it isn't. Waymo has a fraction of the data that Tesla has, yet drives 2000x safer and also better. How come?

Future preview: GPT-5 won't be meaningfully better than GPT-4. Diminishing returns. Same difference.

View attachment 1017486

HD maps. Next question?

And yes, its well know that NN's improve asymptotically at some point, which is implicit in the nature of NNs and the way they are trained. However, I dont see any way you can know that Tesla is anywhere near that point yet.

 

[sleepydoc]

Like @powertoold said, Waymo is basically a micromanaged train.

Drop off a Waymo vehicle to my driveway tomorrow. It won't even be able to drive itself around my block.

That's not too advanced. It's kind of pathetic.

Waymo has done some remarkable things but it’s got fundamental differences from Tesla’s approach and goals so it’s hard to compare the two. Tesla can’t do what Waymo does but waymo can’t do what Tesla’s doing, either.