FSD v12.x (end to end AI)

[SidetrackedSue]

My take on the slowed down video (watched on the posting site, not embedded) is that the regen was full before she hit the brake. It went from 49 to 40 before FSD disengaged (the point when she started braking.) In the comments she says the car slowed but she still hit the brakes. Unfortunately, her hand and then her graphics, block the screen at the end of the incident so we can't see how slow the car ended up going.

What I find interesting is that FSD didn't also steer to the right of the lane (or the turn lane) in order to avoid the collision. It is only a couple of secs in the video but watching the screen, I see no indication it was moving to the right in order to pass behind the car.

I wonder if the slowing of the car by FSD was what had her look to see what was happening (and then triggered the takeover.)

 

[AlanSubie4Life]

watched on the posting site, not embedded) is that the regen was full before she hit the brake.

Also the car was moderately friction braking first. Unfortunately not visible on the regen bar, but it would have been nearly all the way left.

I wonder if the slowing of the car by FSD was what had her look to see what was happening (and then triggered the takeover.)

She saw the issue well in advance; you can hear her gasp (you can see her jaw move as well) and a fraction of a second later her body shifts. Well in advance (about 0.5 seconds or more, no idea how to go frame by frame with Twitter...) of FSD’s first reaction. Some would say she is superhuman!

 

[EVNow]

Yes it clearly did respond.

Just so bad.

Yes, that ULT by the human driver was so bad.

 

[AlanSubie4Life]

Yes, that ULT by the human driver was so bad.

Definitely terrible. Of course, I was speaking to the response from FSD being unnecessarily delayed. And the human should have popped up the gear shift to disengage and coast down for a bit (easy to criticize - this is one of the most challenging aspects of using FSD). No need for friction braking.

No need for all the drama! It wasn't that close - just needed to slow down early. Many defensive drivers would have slowed down in advance of any movement from the car coming from the side! This tends to drive my wife nuts when I do it. You just have to follow your instincts, though.

Unfortunately the safety margin afforded by defensive driving is largely eliminated when using FSD. One of the things to consider when scoring whether it might be safer or not. No one knows (except Tesla, possibly)!

If you disengage FSD all the time at the most minor things that may go wrong, you’ll get some muscle memory for this sort of situation and it may be possible to ameliorate the slow response and have a smoother human-machine interface when these common events occur. I usually disengage several times per mile because that’s just how it works out to get exactly what I want.

 

[EVNow]

Definitely terrible. Of course, I was speaking to the response from FSD being unnecessarily delayed. And the human should have popped up the gear shift to disengage and coast down for a bit (easy to criticize - this is one of the most challenging aspects of using FSD). No need for friction braking.

No need for all the drama! It wasn't that close - just needed to slow down early.

FSD is definitely slow.

This is one area where from early on FSD has surprised me. We are all conditioned to thinking "computers are very fast" compared to us in computation. But turns out AI is slow (though Chat GPT is quite fast !). We see that with robots doing tasks (or playing football !). The current tech is still slower than humans.

 

[AlanSubie4Life]

This is one area where from early on FSD has surprised us. We are all conditioned to thinking "computers are very fast" compared to us in computation.

Really? I'm surprised that v12 has actually gotten to 0.4 seconds (at times). It seems to perhaps be faster than the prior iteration.

But I never expected it to be fast at all.

 

[EVNow]

Really? I'm surprised that v12 has actually gotten to 0.4 seconds (at times). It seems to perhaps be faster than the prior iteration.

But I never expected it to be fast at all.

I think you misunderstood - FSD has been (and continues to be) slower than what I thought it would be.

 

[AlanSubie4Life]

I think you misunderstood - FSD has been (and continues to be) slower than what I thought it would be.

And I am saying it is faster than I thought it would be (and still too slow).

And I don’t really understand why the expectation would be for it to be fast. My default assumption would be glacially slow!

 

[Dewg]

And I am saying it is faster than I thought it would be (and still too slow).

And I don’t really understand why the expectation would be for it to be fast. My default assumption would be glacially slow!

Is it the NN model, or the compute? HW3 I think has 144 TOPS. If we put the same NN on a 2000 TOPS processor, would it react faster, or the exact same?

 

[AlanSubie4Life]

Is it the NN model, or the compute? HW3 I think has 144 TOPS. If we put the same NN on a 2000 TOPS processor, would it react faster, or the exact same?

I have no idea. Probably neither.

One thing that is lacking is “instinct.” You just have to know what the real dangers are to avoid overreacting to every little thing. You have to bias the system to react only when necessary (though I think in this case it was slow even by those criteria).

Whether it is instinct or whether instinct can be trained is TBD. Probably not with current techniques and state of the art, though.

Just seems way too difficult and too sophisticated to emulate.

 

[zoomer0056]

And I am saying it is faster than I thought it would be (and still too slow).

And I don’t really understand why the expectation would be for it to be fast. My default assumption would be glacially slow!

Ha, I needed AI to sort this out, even AI is confused.

 

[AlanSubie4Life]

Ha, I needed AI to sort this out

View attachment 1042410

But I didn't say it was fast. Just keep modifying the prompts until you get what you want and it agrees with you.

 

[JHCCAZ]

Is it the NN model, or the compute?

Good topic. I think they are inextricably linked, part of the AI team's job is to decide how to develop and tweak the model(s) between the Mothership training and the in-car inference.

The HW3 compute is a limit they can't change in those cars (setting aside the dream of an HW5 retrofit program). However, they can improve its performance by applying more and better-architected compute on the training side, by increasing the amount of training data and by curating it better. Ashok actually talked about all this in the earnings call the other day.

HW3 I think has 144 TOPS. If we put the same NN on a 2000 TOPS processor, would it react faster, or the exact same?

With disclaimers about me not being an ML engineer:
I don't believe that simply running the same generated inference model at a higher clock speed would result in better driving behavior. In fact it would probably mess things up pretty badly. It might work for Go or chess or SAT taking, but driving is a real-time real-world activity.

You don't want reactions to every stimulus to happen faster or earlier, you want faster (and broader) analysis and decision-making, giving more time to plan appropriately. On the training side, more examples of people pulling out too late, too slowly, indecisively or giving clues that they don't see you or they suck at driving, will all help.

A faster inference computer helps because it can execute a more sophisticated model, not just reacting faster but processing more clues to anticipate better and to better evolve the planning as the situation develops. From a resource accounting point of view, it can do a better job with a less refined and faster-to-train model and on smaller training data.

Obviously the scaling trade-offs have limits. They can't make it work by distilling all the computing power on Earth into a Game Boy in the car. But the AI engineers are telling us that all these are soft limits, and that the field is rapidly evolving regarding performance optimization vs resource constraints. Performance problems in the currently released software do not imply that they've reached the limit of what HW3 will be able to accomplish.

 

[KArnold]

Why didn't AEB fire? I'm guessing it's because FSD could have stopped.

But I'd react the same way. We'll, probably not WTF, but yea.

 

[Electroman]

FSD didn't respond to a near miss. You can see the FSD supervisor stomps on the brakes ~1sec after the car begins turning into path and well before FSD is able to respond. Almost certainly if she didn't respond that quick it would've been an accident. The UI isn't the best indicator but the car of interest never changes color. FSD doesn't command steering changes or deceleration. Another poor showing for FSD being safer than a human driver!

https://twitter.com/x/status/1783814624791380153

Another thing that bothers me is, the UI is a bit lagging as you can see in this screenshot. The car has 3/4ths of the way out into the other lane, but UI shows it is bang in the middle.

 

[Electroman]

Just for kicks it is still on FSDS until after this Frame. The car's front end is driving down and she is moving forward so it was responding. You can't read the speed. Also its starts driving down as soon as the car crosses the lane marking.

View attachment 1042339

Video quality is perfectly fine on Twitter.

 

[JB47394]

Another thing that bothers me is, the UI is a bit lagging as you can see in this screenshot.

In what way does that bother you?

 

[David Wang]

I had something funny happen to me today.

I was on the freeway with fsd enabled v12.3.4 and there was a motorcycle passing my car on the right. After the motorcycle had already passed me, the car did the maneuver where it moved
slightly to the left to give the motorcycle space. What a terrible delayed reaction.

The car behind me must have thought I was an idiot.

They really need to move the highway stack to v12, it's not great imo.

 

[aronth5]

A good question is - why is the driver letting the car make that mistake in the first place. Isn't one of the ideas to disengage when it makes an important mistake - otherwise how will it ever get better ?

I almost always intervene if the select lane will lead to a missed turn (or in this case taking a turn not needed). Today on my short 5 mile trip, I had to intervene 3 times to change lanes. Atleast one of them, I had to never to with V11. 12.3.4 seems to take the wrong lane ...

At 12-13 seconds navigation changed from a U-turn to a right turn which the car did start to take. At the same time FSD was asking the driver to apply torque to the wheel. Watch at quarter speed and the planner and wheel were going right. So I'm not surprised the driver didn't intervene until the car suddenly started to turn left. Perhaps better to disengage when he thought the route was wrong.

 

[EVNow]

And I am saying it is faster than I thought it would be (and still too slow).

And I don’t really understand why the expectation would be for it to be fast. My default assumption would be glacially slow!

I see - the expectations are, as I said, are based on how fast computers do computations. But, alas, NN is slow because it has to do too many computations to figure out and is not "hard wired" like real organic neural networks.