FSD v12.x (end to end AI)

[lzolman]

So she has v12 and never uses it? Do you have v12 too?

I said FSD, not V12! I'm in New England lol

 

[aronth5]

I can totally understand people posting for the engagement because this was not something v11 could really do. And that is fine.

But it just is not somewhere you’d ever want to use FSD, since parking lots are one of the most dangerous places for vehicle damage (they are perhaps the most complex environment), and FSD is certainly not going to help in that regard. Useful (potentially - I have no idea) for the reasons mentioned earlier.

Well with the high number of recent videos which are including parking lots I'd say V12 users disagree with you. Since your going slow as long as you're really careful shouldn't be much of a problem. I know once I get V12 I'll try some out.

 

[TK211X]

Huh, called friend stuck on holiday .8 and he got no update.

So a group with USS got .14 then moved to latest in unison. FSD subscribers without USS are still on Holiday .8?

 

[enemji]

Millions of variations say 10^6? Some estimate Chess might be around 10^50 board configurations while Go 10^170, so maybe this is "easier" in some sense.

The big difference is that in chess everyone is trying to harm everyone whereas in our case the core assumption is that everyone is trying not to hurt anyone.

 

[Supcom]

Fair warning?

https://twitter.com/x/status/1765920801440858129

 

[Mardak]

100x better intervention rate than v11 - No

I suppose it doesn't really matter what you might consider an intervention, e.g., too gradual acceleration (requiring pedal), too rapid deceleration (tricky to intervene with lots of scrollwheel down), turn signal dis/engage, or even "tolerated" steering (as otherwise correct non-jerky steering would be a disengagement). People have reported much improved comfort and smoothness with 12.x that would address various types of interventions, but the inability to get to 100x better would seem to mean you think there'll be interventions that 12.x just won't be able to figure out.

Do you have specific intervention examples of what would be more difficult for end-to-end? Or is it more general in that you don't think 12.x will ever drive like you, so you'll always be intervening anyway?

 

[powertoold]

Nobody cares but I finally got FSD 12.2.1 back after calibration, my observations from last few days:

1) It's hard to go back to normal AP / NoA
2) Found my best setting for ADAS driving: average with auto-max, I just press the accelerator when I feel like it's going too slow
3) Overall very impressive performance and smoothness from 12.2.1, I still think it will eventually be a very compelling ADAS and will blow your minds once in wide release; general robotaxis is likely impossible, especially for HW3, geofenced robotaxis with specialized locale models more likely if at all
4) I love that U-turns are possible now, and it will still attempt the U-turn even if it says, "cannot handle maneuver, please assist"
5) I'm still disengaging for moments when it gets too close to a curb or car when pulling out of parking spots / tight parking lots / turning out of atypical driveways
6) It drives better than most Uber drivers most of the time; it just needs that further refinement for all the little "too close" to curbs / car edges / etc.
7) Tight lane changes can be close to expert level with 12.2.1

Overall the future of FSD is very bright. There is definitely hope that the little things will be resolved over time, but the limit and speed of progress are still unclear.

 

[AlanSubie4Life]

I know once I get V12 I'll try some out.

I will too, probably more than one! But that is very different than using it in parking lots.

Do you have specific intervention examples of what would be more difficult for end-to-end? Or is it more general in that you don't think 12.x will ever drive like you, so you'll always be intervening anyway?

I don't think anything will necessarily be more difficult for end-to-end, as compared to v11.

As you say, since we're talking about a rate reduction it probably doesn't matter what sort of intervention we're talking about. But I'm talking/thinking about just safety interventions, so things like intervening to avoid slowing down too quickly, not allowing sufficient following distance on a lane change, not signaling, incorrect lane choice, avoiding honks from other drivers, etc. Also on things like unprotected lefts (the cross-traffic variety, not the oncoming traffic variety), the intervention rate is very high, and virtually all of those are safety interventions.

Definitely doesn't have to drive like me. But without basic competence and consistency, then there will be a lot of interventions.

100x reduction in rate in v12 vs. v11 is what I don't think is possible. That's a huge, dramatic, incredible reduction. I would be shocked but pleasantly surprised by 10x. It seems like it could be possible, though I don't expect it. But 100x? My opinion is that there is just no way with v12 with current hardware. Bookmark this post.

 

[powertoold]

12.2.1 is a beautiful thing, I've been asking for this ever since I got FSDb

https://twitter.com/x/status/1765880305989325183

 

[Kablooie]

V12 not yielding to a car when taking a left at a stop sign. I thought it interesting because it also involves a bus temporarily blocking V12's view of that car.

Look like another training issue. They need to find hundreds of cases where something blocks the view before a turn so it learns to wait and assess the situation before proceeding.

 

[AlanSubie4Life]

general robotaxis is likely impossible, especially for HW3, geofenced robotaxis with specialized locale models more likely if at all

You are definitely starting to sound less confident about these robotaxis. I guess we will see about these specialized locale flavors.

Oct. 2021:

I'm pretty confident FSD will be reliable and safe enough for geofenced robotaxis (easy routes, no construction, etc.) within the next 6 months. Whether or not Tesla will deploy them is another story.

 

[powertoold]

You are definitely starting to sound less confident about these robotaxis. I guess we will see about these specialized locale flavors.

Oct. 2021:

Yea general robotaxis that operate like an Uber USA-wide seem very far off, but geofenced ones seem within reach for Tesla.

Even Waymo doesn't operate like an Uber within its cities of service. Waymo uses highly curated routes and fences off certain areas and whatnot.

 

[johnm]

But unfortunately it is now over. You know when James “Doom” Douma starts talking about how FSD v12 is going to change everything, it’s all over, and a v13 (Beta?) is basically guaranteed. Reliable indicator! Then repeat the hype cycle.

Tesla hasn't been super consistent in version numbering. I kind of hope they copy Apple.

Mac OS 8 came out in 1997. Mac OS 9 came out in 1999. They then stuck with Mac OS 10.x releases for the next TWENTY YEARS. And then 11, 12, 13 and 14.

Are you sure FSD 12.85 in 2044 won't be amazing? But maybe v13 in 2045 will be!

 

[zoomer0056]

I kind of hope they copy Apple.

Good one. Tesla is not like any other company, Tesla only does Tesla.

 

[Bladerskb]

LOL. Interestingly both can be true - v12 could drive better than Waymo in certain circumstances and worse than even v11 in others.

What circumstances is v12 actually driving better than Waymo? Are there any actual legal tangible examples? Running head fast into a situation isn't driving better.

Yea general robotaxis that operate like an Uber USA-wide seem very far off, but geofenced ones seem within reach for Tesla.

Even Waymo doesn't operate like an Uber within its cities of service. Waymo uses highly curated routes and fences off certain areas and whatnot.

This is simply not true, Waymo literally acts like a Uber and isn't route-based. You can go anywhere within its geofence. Its funn how Tesla fans only point out historical issues. Like if Waymo temporarily blacklists a street or building or parking lot in the past. Its now their personification and somehow nullifies the entire robo-taxi service. Which is like saying if I get 99% on a test, I actually failed and should get an F because I couldn't answer one question. It's ridiculous but makes total sense if your only goal is to prop up Tesla and teardown others.

Earlier in Waymo's launch in 2020, they influenced the routes the car took but that didn't mean that the cars still didn't take routes that it didn't favor. Proven by the fact JJRick racked up hundreds of left turns in like a year. Earlier they also blacklisted construction zones. But they don't anymore, but the problem is, people still will go back 4 years and say see how Waymo struggled at this construction (the famous cone incident).

But when you point out FSD fails from 2 weeks ago, they tell you how its irrelevant because a new version is out. Wait they go back 4 years for Waymo but can't even go back 2 weeks for Tesla. The bias is next level.

So when Waymo would pause for rain, block off construction zones, grey out an area or two, pause for mild fog, influence the routes. You use it against them 4 years later when they can now handle heavy rain, dense fog, construction zones, dead-ends, traffic personnel hand re-routes, all kinds of parking lots, and people like Ken being able to get the car to go basically anywhere they want on the map.

Yet you are still pushing the "its a route service".

You have also pushed the Waymo is just a glorified tram service, a glorified AEB system, now when you realize that Tesla can't do worldwide L5 or USA L4. Now you are referring to the idea of them doing the same thing as a "localized geofence service".

WHAT?

Its crazy the amount of stuff Waymo does that Tesla didn't do and the bad stuff Tesla does that Waymo doesn't do.
But the fans pretended it didn't happen until v12 came out and suddenly they started letting rip just how bad v11 was.
Even Omar was basically saying how horrible v11 was compared to Waymo, but when it was just v11 he wasn't saying that at all, infact he was saying the opposite. It was quite an incredible revelation of the hidden truth.

You fans never ceases to amuse me.

The last thing left on Waymo's DDT is highway.

 

[JB47394]

Look like another training issue. They need to find hundreds of cases where something blocks the view before a turn so it learns to wait and assess the situation before proceeding.

Possibly. It may also be a lack of sufficient persistence. Where does persistent information live in a neural network? Or is V12 simply "living in the moment"?

 

[kabin]

Well with the high number of recent videos which are including parking lots I'd say V12 users disagree with you. Since your going slow as long as you're really careful shouldn't be much of a problem. I know once I get V12 I'll try some out.

And wasn't v12 marketed to be able to take you to your destination? Photon magic doesn't distinguish between roads and parking lots.

 

[kabin]

Possibly. It may also be a lack of sufficient persistence. Where does persistent information live in a neural network? Or is V12 simply "living in the moment"?

Exactly. Maybe a critical unknown. What is v12's scenario recall capacity? 100's of Milliseconds? What's really needed for safe driving?

There might be a forgetting factor where new photon info overwhelms old photon info. Do they recursively feed old object data back into the net assuming/estimating object kinematics forward in time/space?

Recall the team complained about radar having dropouts. Later the team admitted vision only had the same issue. Duh!

 

[AlanSubie4Life]

but geofenced ones seem within reach for Tesla.

Nah.

Are you sure FSD 12.85 in 2044 won't be amazing? But maybe v13 in 2045 will be!

Did not say anything about amazing or not. Current hardware.

 

[EVNow]

What circumstances is v12 actually driving better than Waymo? Are there any actual legal tangible examples? Running head fast into a situation isn't driving better.

Thats why I said "could". People who have watched more V12 (and more Waymo) videos can perhaps give examples.