FSD v12.x (end to end AI)

[TK211X]

Judging from the *mostly* negative reviews and regressions, I'd wager that 2024.3.20 (FSD 12.3.5) won't roll out wider. But who knows, maybe they can tweak some behind the scenes weights and knobs...

12.3.5 had two small waves and died. I saw on one YouTube video it went over one curb and insanely close to 3 others. So no fix for that. Hope 12.4 or 12.3.6 has a real fix instead of whatever they’re doing now.

 

[AlanSubie4Life]

Hope 12.4 or 12.3.6 has a real fix instead of whatever they’re doing now.

If they don’t retrain it seems really unlikely this sort of thing would be fixed. At least at this point it seems clear the prompt engineering doesn’t have the reach to address these deeply buried behaviors.

12.4 will be the most hugely significant release we have ever seen. It determines whether there is any hope for FSD improvement. Arguably it should be called v12. Will be interesting to see whether any behavior can be changed for the high disengagement areas (I imagine people are very frequently disengaging for corners).

 

[MP3Mike]

12.3.5 had two small waves and died. I saw on one YouTube video it went over one curb and insanely close to 3 others. So no fix for that. Hope 12.4 or 12.3.6 has a real fix instead of whatever they’re doing now.

We don't even know if that is something they were trying to fix in 12.3.4 or 12.3.5. (But I haven't seen it get that close to curbs in my driving on 12.3.3 or 12.3.4, so it may be largely dependent on particular road designs.)

I know that 12.3.4 greatly improved the yellow light handling over 12.3.3. It now taps the brakes and then decides to go through when appropriate, rather than slamming on the brakes at a yellow light. (I'd prefer it didn't tap the brakes first, but that may be a limitation of end-to-end NN and it is the "guardrail" code that forces it through the yellow light, and that code reacts to, and overrides, the NN command.)

 

[Todd Burch]

With 1 billion miles driven on FSD and growing quickly do we know if there have been any fatalities?

If there were, I'm 100% sure you'd hear about them.

 

[Twiglett]

We don't even know if that is something they were trying to fix in 12.3.4 or 12.3.5. (But I haven't seen it get that close to curbs in my driving on 12.3.3 or 12.3.4, so it may be largely dependent on particular road designs.)

I know that 12.3.4 greatly improved the yellow light handling over 12.3.3. It now taps the brakes and then decides to go through when appropriate, rather than slamming on the brakes at a yellow light. (I'd prefer it didn't tap the brakes first, but that may be a limitation of end-to-end NN and it is the "guardrail" code that forces it through the yellow light, and that code reacts to, and overrides, the NN command.)

On some roads for me, 12.3.4 is bouncing over rumble strips it’s so far over the yellow lines

 

[Todd Burch]

Odd experience with 12.3.4 today:

Stopped at a light. While waiting for the green, the wheel started gradually turning left. It continued to slowly turn all the way to the left--almost completely to the stops. When the light turned green it immediately straightened and it was fine, but I think I left 16 ounces of Continental's finest rubber on the road back there.

Also, FSD crashed while I was getting off the interstate (not the car--the software). Red hands of death, told me to take over. Turn signal was on and I saw the repeater camera go black while the system rebooted. Came back in about 25 seconds and I reengaged.

I think FSD (software) crashes are 100% reported to Tesla immediately so I'm sure they have a record of what happened.

 

[Todd Burch]

2021 data from IIHS:
"The death rate per 100 million miles traveled ranged from 0.71 in Massachusetts to 2.08 in South Carolina"

So we know Tesla has well over 300 million miles of v12 driving, and there have been zero fatalities. (We'd know if there were).

So FSD v12 supervised has saved anywhere from 2 - 7 lives. (I know sample rate is still relatively small for this statistic to draw conclusions, but it's encouraging).

I agree with Elon. Tesla can easily prove x interventions or y accidents or z fatalities per N miles or driving. Given a larger sample size of many billions of miles, once we reach zero disengagements it won't take long for them to convince regulators that FSD is many times safer than a human. At that point I expect regulatory approval would come swiftly.

 

[kabin]

Judging from the *mostly* negative reviews and regressions, I'd wager that 2024.3.20 (FSD 12.3.5) won't roll out wider. But who knows, maybe they can tweak some behind the scenes weights and knobs...

Or they need to add more bat brains and ghost poo.

 

[kabin]

2021 data from IIHS:
"The death rate per 100 million miles traveled ranged from 0.71 in Massachusetts to 2.08 in South Carolina"

So we know Tesla has well over 300 million miles of v12 driving, and there have been zero fatalities. (We'd know if there were).

So FSD v12 supervised has saved anywhere from 2 - 7 lives. (I know sample rate is still relatively small for this statistic to draw conclusions, but it's encouraging).

I agree with Elon. Tesla can easily prove x interventions or y accidents or z fatalities per N miles or driving. Given a larger sample size of many billions of miles, once we reach zero disengagements it won't take long for them to convince regulators that FSD is many times safer than a human. At that point I expect regulatory approval would come swiftly.

I'm betting FSD has caused well over 7 heart attacks and strokes.

 

[TK211X]

We don't even know if that is something they were trying to fix in 12.3.4 or 12.3.5. (But I haven't seen it get that close to curbs in my driving on 12.3.3 or 12.3.4, so it may be largely dependent on particular road designs.)

I know that 12.3.4 greatly improved the yellow light handling over 12.3.3. It now taps the brakes and then decides to go through when appropriate, rather than slamming on the brakes at a yellow light. (I'd prefer it didn't tap the brakes first, but that may be a limitation of end-to-end NN and it is the "guardrail" code that forces it through the yellow light, and that code reacts to, and overrides, the NN command.)

My ass, they better be trying to fix FSD trying to end the cars life on curbs.

This is like a red hot panic issue. The car drives over a curb in a turn and obviously lifts up. Guess what’s suddenly exposed. The battery pack. Imagine totaling the car because the car lifted on a curb and either grinds on or settles on something that deforms or damages the battery pack. Unimaginably stupid.

 

[PianoAl]

I found a workaround to the problem of lane change refusals.

Hold the turn signal. Force the signal until the lane change completes.

I just went out and tried it. I had a couple lane changes that looked to be head fakes, where it will normally complete after wandering, but I had one that was a clear refusal as the car sat hard against the lane line for a half second or so. Normally, the car would just not make that lane change, but holding the turn signal throughout made sure that it completed the change.

Something else I noticed is that the car voluntarily moved into the right lane (of two). I had no destination, so this was the car getting right, as it should. My lane change head fakes and the one attempted refusal were, I think, going against FSD's lane choice, which is why it does those things. I think we're telling FSD to change lanes, but a moment later it's internally reasserting its own lane choice, and that's the source of the head fakes and the outright refusals. The difference may be down to timing, or just random chance of how strong the "stay in lane" signal is.

I tried this and holding the stalk down did not work for me. Canceling and requesting it did.

Also, I've seen it refuse for a while but eventually change lanes.

 

[Tronguy]

In 2022, PG&E's Diablo Canyon nuclear power plant produced around 18,000 GWh of power. The problem with nucs is that they are not able to throttle down when they are not needed, such as during peak solar production. Trying to throttle down is what blew up Chernobyl, and also what melted down three reactors at Fukushima. Anyway, that produced some of the power which was in "excess" of what was needed. Also some large, remote solar farms, which also load the transmission lines.

Um. At least get your facts straight. What blew up Chernobyl was a bunch of inadvisable tests run simultaneously with inadequate planning: Using the plant's steam turbine to power the coolant pumps with a simultaneous loss of external power and a coolant pipe rupture. That was entirely idiot-human generated.

I'll give you Fukushima, in the sense that a natural disaster exposed some truly inadequate planning on the part of the operators.

As far as shutting it all down in a hurry: Both sets of plants would have been relatively OK had a scram been excuted early on, but that was bad operation.

But shutting down a nuke? They practice that.

 

[JulienW]

Yea way OT but there is no big disadvantage to over producing with a nuclear reactor since there is almost no fuel costs. So you don't need to power down like a gas generator or peaker plants to save money.

 

[willow_hiller]

Odd experience with 12.3.4 today:

Stopped at a light. While waiting for the green, the wheel started gradually turning left. It continued to slowly turn all the way to the left--almost completely to the stops. When the light turned green it immediately straightened and it was fine, but I think I left 16 ounces of Continental's finest rubber on the road back there.

This has happened to me twice as well. I think the first was on 12.3.3, and a second on 12.3.4.

Could be the model predicting that the traffic ahead of it won't move and it needs to turn the wheel a lot to make it around. Or maybe it's FSD trying to tell us that it's conscious.

 

[enemji]

Stopped at a light. While waiting for the green, the wheel started gradually turning left. It continued to slowly turn all the way to the left--almost completely to the stops. When the light turned green it immediately straightened and it was fine, but I think I left 16 ounces of Continental's finest rubber on the road back there.

Were you behind another large vehicle?

FSDS will do that maneuver when it can’t see the traffic light and identifies that situation as a stuck vehicle and attempts to move around it

 

[JulienW]

Odd experience with 12.3.4 today:

Stopped at a light. While waiting for the green, the wheel started gradually turning left. It continued to slowly turn all the way to the left--almost completely to the stops. When the light turned green it immediately straightened and it was fine, but I think I left 16 ounces of Continental's finest rubber on the road back there....

Here is post I made just yesterday:

At a Red light had a strange one. I'm sure most have had V12 decide to turn the steering wheel a few ºs. Today mine did it to the right/towards the curb. Then it turned it more and then again it turned past 180º. Then it turned it even more and was almost 360º. Started to worry a little since on narrow city street and curb probably <1' away. When light turns greens ready to grab but it spun back around.

Is FSDS So f'n confident in it abilities now that it is playing Beach Buggy Racing while at a Red Light?

 

[Todd Burch]

This has happened to me twice as well. I think the first was on 12.3.3, and a second on 12.3.4.

Could be the model predicting that the traffic ahead of it won't move and it needs to turn the wheel a lot to make it around. Or maybe it's FSD trying to tell us that it's conscious.

This would be a good case for that C++ guardrail control code: if the car’s not moving, don’t allow any steering input.

 

[Todd Burch]

Were you behind another large vehicle?

FSDS will do that maneuver when it can’t see the traffic light and identifies that situation as a stuck vehicle and attempts to move around it

No, just behind a compact car at a light. The light was clearly visible—but that’s a theory as to why it was turning the wheel.

 

[mgs333]

Small push of 12.3.5 on Teslafi, my wife's car got it while on the free month trial, my car not yet.

 

[lascavarian]

Yes, it seems a couple hundred pending for now.