FSD v12.x (end to end AI)

[powertoold]

V12.2.1 makes some mind-blowing decisions based on subtle cues

12.2.1 auto wipers are broken as before

Also sentry mode seems broken, keeps recording videos when there's no one around

 

[brkaus]

My observation is that it sometimes/often takes a long time to acknowledge a change in torque. More precisely, it sees the change in torque right away, but doesn’t cancel the blue pulsing until sometime later. So, you think the car didn’t notice your response and you try harder or in a different direction, but the pulsing stays there. Kind of like a remote video game with terrible lag.

I feel like it is dependent on what is going on the the center screen. I see the lags when I am listening to streaming music and not radio for example. At least that’s my perception.

 

[AlanSubie4Life]

If hw3 can’t be replaced with hw4 maybe they will make it so hw5 can replace hw3, if it’s just a simple wiring connection issue

Haven’t they booked most/all of the revenue now? If they have delivered the promised product with v11 except for prior to March 2019 customers, why would they do this? It’s all gravy for the updates now; product has been delivered.

I guess it is arguable whether they have achieved the Autopark and Smart Summon goals, no idea on the exact timeline of specific promises there. But anyway, the key is what revenue they have booked. That determines when they are done, and from there they just have to deliver continuous “gravy” improvements, which can just asymptotically improve things.

Anyway the website says they are done. It does imply by buying FSD you are purchasing continuous improvement but that is clearly so vague it means nothing.

 

[FSDtester#1]

V12.2.1 makes some mind-blowing decisions based on subtle cues

12.2.1 auto wipers are broken as before

Also sentry mode seems broken, keeps recording videos when there's no one around

Your battery and the phantom drain must love the nonstop Sentry recording!!

For the first time ever last week, I got an alarm notification on the app.

Went to the car, nothing. Checked recording, some dude in a loud ass muscle ice car drove in front of it, stopped, and must of revved his engine.

 

[Mardak]

Not familiar with all these AI projects to know if applicable and how they worked exactly… They've been working on this for a long time so suspect they're way up there

The Elo rating system reflects relative skill level such as ability of a chess player to win or tie against a stronger opponent or benchmark, and the charts show that initial thousands of training steps can stay at 0 rating (i.e., loses all games) as the neural network is only starting to understand how pieces move and how important which pieces are. One reference point not shown on the chart is an average human chess Elo might be around 800, and the neural networks quickly get to that and continues well past that with additional data and training to build on that initial understanding of piece value to then learn concepts like capturing the enemy queen is generally good to then tactics and eventually subtle strategies to be superhuman. The progression of AlphaZero neural networks come from it incrementally understanding concepts to then make use of those concepts to discover new concepts when appropriate for its skill level.

12.x has "only" been actively prioritized for about a year now, so that might be a long time or relatively short in the whole Autopilot program, but either way, end-to-end should still be early in training steps in that it hasn't had the full feedback loop of learning concepts then discovering where those are wrong (require disengagements). Neural networks trained on diverse data pick up signals in unexpected ways, so sometimes it can learn the correct behavior for the wrong reason or could fail to learn "complex" behaviors because it doesn't have prerequisite understanding. For example, staying out of blind spots could naturally come from mimicking training data behavior of how people drive, but it could also depend on learning of a concept that adjacent vehicles might suddenly change lanes.

 

[JulienW]

V12.2.1 makes some mind-blowing decisions based on subtle cues.

12.2.1 auto wipers are broken as before

Also sentry mode seems broken, keeps recording videos when there's no one around

Funny that Tesla is supposedly working on "Deep Sentry" and lower the power used by 50%. Probably right up there with "Deep Rain".

 

[AlanSubie4Life]

The progression of AlphaZero neural networks come from it incrementally understanding concepts to then make use of those concepts to discover new concepts when appropriate for its skill level.

Yes I get that basic concept but no idea if convergence pattern shown in trivial games like Go or chess will be matched for FSD task. Just no idea. FSD much more complicated and open ended.

12.x has "only" been actively prioritized for about a year now, so that might be a long time or relatively short in the whole Autopilot program

Seems like a long time for AP program.

Seem like they are probably pretty far along. As I said it seems likely there will be significant further improvements. If they are 90% of the way there there are a huge number of improvements to come.

I really have no idea though. Super speculative and I have absolutely no clue.

I am optimistic that certain very desirable behaviors in mundane driving situations will start to emerge. Like the buffer concept. If they try. My concern is they don’t care.

And of course getting it to be good for more complicated scenarios in nearly all locations seems near impossible. Just mind-bogglingly complex. I can’t even begin to conceive how it will work on current hardware, while it seems pretty clear (and not just in retrospect) that this approach would work extremely well for Go and chess.

 

[sleepydoc]

Semantic distractions are a dead end. There will always be technological/cost/weight/size limitations. We have to work within real world constraints.

More importantly, is FSD using enough sensors and collecting sufficient data to safely respond to the environment? I've yet to see v11 or v12 equipped vehicle respond quicker than 1.5 seconds even though processing keeps up with realtime input data.

But can you define ‘drive?’

 

[sleepydoc]

On v11 you can still change music volume or use speed up or down to satisfy the nag. Word is on v12 that's gone.

@AlanSubie4Life has mastered the art of satisfying nags without disengagements, read some of his tips.

Alan had Elon implant a neuralink in his brain so he is one with the car. The rest of us can only dream of having such a symbiotic relationship with our Teslas

 

[Mardak]

My point is that, apart from training up, there will also be a period of packing, where the Tesla engineers will have to learn how to cram as much competency into the available network capacity as possible

Yeah, for AlphaZero, they wanted to be hands-off without human knowledge/interference, so the neural network size was fixed for a given training run. This also emphasizes that the skill progression in the charts is coming from better training data as self-play found more interesting examples from its current understanding. The early mediocre performance wasn't because the neural networks didn't have the capacity to achieve competency.

Tesla has the flexibility (and additional complexity) to adjust the network size, parameters, architecture, etc. as they collect even more training data to notice the networks' ability to learn certain concepts or detect the inability to reduce losses in other cases. Hopefully the previously reported "excess" compute of the faster 12.x networks provides Autopilot engineers more options in exploring 12.x improvements outside of more training data.

 

[stopcrazypp]

Haven’t they booked most/all of the revenue now? If they have delivered the promised product with v11 except for prior to March 2019 customers, why would they do this? It’s all gravy for the updates now; product has been delivered.

I guess it is arguable whether they have achieved the Autopark and Smart Summon goals, no idea on the exact timeline of specific promises there. But anyway, the key is what revenue they have booked. That determines when they are done, and from there they just have to deliver continuous “gravy” improvements, which can just asymptotically improve things.

Anyway the website says they are done. It does imply by buying FSD you are purchasing continuous improvement but that is clearly so vague it means nothing.

I had to double check, but you are right. The "Autosteer on city streets" is no longer "coming soon". That means Tesla considers everything delivered now (putting aside other issues like non-USS cars not having Autopark and Summons). As such, other than the pre-2019 customers you mentioned (which were promised more on the order page), Tesla considers themselves to have delivered all the functionality as ordered.

So Tesla will not have to do any upgrades it seems for most buyers.

 

[stopcrazypp]

Funny that Tesla is supposedly working on "Deep Sentry" and lower the power used by 50%. Probably right up there with "Deep Rain".

I take it you mean this?

Tesla works to slash Sentry Mode power usage by ~40%, and why that's bigger than you think

Tesla is working on a software update to slash Sentry Mode power usage by ~40% – reducing “vampire drain” on...

electrek.co

I couldn't find "deep sentry" which implies they are changing the detection algorithms. If the change instead is to switch to continuous recording that allows them to shut down more of the processors (and other subsystems), I'm sure a lot of people would be fine with that tradeoff.

 

[mongo]

The quote you posted says "Approach for a left turn shall be made in that portion of the right half of the roadway nearest the center line"

Approach isn't in question as they were designated turn lanes...
That passage continues:
"and after entering the intersection the left turn shall be made so as to leave the intersection to the right of the center line of the roadway being entered."
"To the right", not "to the immediate right " In other words, hug the center of the departing lanes and enter anywhere to the right of center of the entering lanes...

You should read 257.650 of the Michigan Vehicle Code. It pertains to right-of-way, specifically for left turns.
257.650 Right-of-way; turning left at intersection; violation as civil infraction.

I did...

Sec. 650.

(1) The driver of a vehicle within an intersection intending to turn to the left shall yield the right of way to a vehicle approaching from the opposite direction which is within the intersection or so close to the intersection as to constitute an immediate hazard

Which is what I wrote:

I can't find a right vs left yeild priority rule beyond just yeilding to vehicles in the intersection.

The law says nothing directly about left lane selection only that left yeilds to other cars (straight or right). Though, given the lack of similar wording regarding right turn law, left is lower priority.

A left turn onto a 4 lane road should not need to yield to a right turn as the right turn would go to the outer lane.
Two left turns and a right into 2 lanes allows for a zipper merge into the right most lane, but left can't cut right off.
(In Michigan)

 

[TK211X]

This is not the Version we’re looking for

 

[aronth5]

it'll beep. then you have seconds to present torque to the wheel. Then it'll SCREAM and show you a dirty picture, then you're done.

Oh so that is why some people get 5 strikes. Picture viewing

 

[rlsd]

This is not the Version we’re looking for

View attachment 1023392
View attachment 1023393

Poll: When will FSD V12 be in wide release?

I'll try whining at support ;) Don't try this. :)

teslamotorsclub.com

 

[powertoold]

I don't know how Tesla is gonna train V12 for all the thousands of different situations. So far, NNs don't seem to "get it" yet.

I had to disengage today, optionally because 12.2.1 drove into a backed up left turn lane, it wasn't blocking the intersection, so I guess it's fine, but it wasn't ideal

V11.4.9 usually stayed back if it saw the lane was backed up

 

[Todd Burch]

I wish there was another way to let it know you are there because there’s a fine line between apply force to the wheel for nag and disengaging FSD

I would get the blue flash sometimes, apply force to the wheel and it keeps flashing then I would apply more force and disengage it on accident

It’s not really a fine line. There’s a big difference in torque required.

The trick is this: you can apply a pretty small torque, but you have to hold it for about a second. If you do that, you’ll never disengage accidentally unless you happen to be challenged in the motor skills department.

 

[jabloomf1230]

Alan had Elon implant a neuralink in his brain so he is one with the car. The rest of us can only dream of having such a symbiotic relationship with our Teslas

It's like book/movie Terminal Man. Every time Elon disengages FSD, his brain receives a shock.

 

[Todd Burch]

I don't know how Tesla is gonna train V12 for all the thousands of different situations. So far, NNs don't seem to "get it" yet.

I had to disengage today, optionally because 12.2.1 drove into a backed up left turn lane, it wasn't blocking the intersection, so I guess it's fine, but it wasn't ideal

V11.4.9 usually stayed back if it saw the lane was backed up

Millions and millions of video clips. This is going to take a little while. But the great thing is that every new network version should, on the whole, get better. And only Tesla has the capability to gather that much data.