Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register
This site may earn commission on affiliate links.
If true, will self driving ever be cheap? There could be many more training computer updates needed in the future.

View attachment 1037249

https://insideevs.com/news/715366/tesla-10-billion-self-driving/

The majority of the cost is setting up new servers (buying chips, installing in racks, etc), once they finish scaling to the desired size the ongoing costs drop quite a bit.

On top of that the servers can be repurposed for the humanoid robot training if FSD becomes a solved problem moving costs to another project

On top of that if somehow FSD and the humanoid bot are both solved the hardware can be re-tasked to other companies (rented out AWS style).

Some of that may not come to pass, but its clear the ongoing costs will be less than the ramp up costs even if every existing server works on nothing but FSD from now until the end of time.
 
  • Like
Reactions: zoomer0056
I wish FSD would have a way for the driver to alter routing decisions on the drive. I want to be able to say things like 'skip the next exit' and have FSD skip that exit because I'm familiar with the traffic patterns following that particular exit.

Yes being stuck with the path the map has created sucks if you know an alternative path that is physically longer is actually faster timewise or comfortwise.

This is my #1 reason for disengagements. Not liking a certain part of the route.
 
Yes being stuck with the path the map has created sucks if you know an alternative path that is physically longer is actually faster timewise or comfortwise.

This is my #1 reason for disengagements. Not liking a certain part of the route.

Mine too! My #2 is not executing a requested lane change - I made damn sure to report to Tesla every time this made me disengage FSD.

Edit: Actually, lately my #1 has been: I'm already at destination.
 
Mine too! My #2 is not executing a requested lane change - I made damn sure to report to Tesla every time this made me disengage FSD.

Edit: Actually, lately my #1 has been: I'm already at destination.

I think my #3 or #4 is not getting it into autopark mode (me still in the car) in a parking lot (or anywhere).
Admittedly I haven't "Read The #$%#$% Manual" so I'm pathetically flipping things on and off trying to get the "P" symbols to appear so I can tell it to autopark.
 
Last edited:
  • Funny
Reactions: Hiline
Interesting. Wonder if there will be any evidence of retraining in this one?

So far they seem to be holding off on that. Haven’t seen anything in 12.3.x, just minor adjustments which don’t seem to appear like retraining.

I am not sure what retraining would look like, it’s just my sense that it has not been done yet. I think it would look and feel different.

Probably will take corners, etc. exactly the same way - but hopefully not!

Would be nice to know, in any case.
 
Being a 12.3.X it can't be too much and defiantly not a major retraining. So no way is it a "ChatGTM" moment we are hoping for with 12.4.x. Also seems odd is it is still on 24.3.x. The poor people on 24.8.x and waiting on "next weeks" trial week after week after.....are suffering some Elon blab off promise that didn't happen.

incoming-9C167B92-84B7-4DBA-8C6E-1532546E3268.jpeg
 
  • Funny
Reactions: FSDtester#1
Interesting. Wonder if there will be any evidence of retraining in this one?

So far they seem to be holding off on that. Haven’t seen anything in 12.3.x, just minor adjustments which don’t seem to appear like retraining.

I am not sure what retraining would look like, it’s just my sense that it has not been done yet. I think it would look and feel different.

Probably will take corners, etc. exactly the same way - but hopefully not!

Would be nice to know, in any case.

You mean retraining as in feeding the same data into a different architecture? If so, we'd have to wait for V13, I think.
 
The employee(?) vehicle that shows up on TeslaInfo seems to be on 2024.3.15 but release notes only mention "Minor Fixes" so unclear if this is the same 12.3.4? Tesla 2024.3.15 Release Notes

Given Tesla's previous release patterns, 12.3.1, 12.3.2.1 and 12.3.3 were relatively unchanged in FSD driving behavior compared to 12.3, but each of those versions did add more support for more regions (Canada), hardware (vehicles and computers), software (those on 2024.2.x). So 12.3.4 might be focusing on adding more support such as portrait S/X?
 
You mean retraining as in feeding the same data into a different architecture?

New video showing more correct behaviors, etc. whatever their iteration process is to adjust the weights and tweaks to the architecture if needed. Seems like hasn’t been done recently.

I fear HW3 will never solve the 1.5sec latency issue
FSD v12.x (end to end AI)

I think it can be as low as 0.5 seconds (reaction time here).
 
  • Informative
Reactions: kabin
Interesting short drive today. I drove on a 2-lane per side stroad where the right lane also has parked cars. It's quite narrow so I don't usually like driving in that lane, since you only have about 1-2 feet of clearance to the parked cars. If there's a big pickup or a truck, you have to cut over into the other lane.

Anyway, FSD decided to grab that lane and I was by myself in the car, so I white-knuckled the steering wheel and supervised it to see what it would do.

* It held its position in the lane very well. As I said, I would normally not drive in this lane because it's too close to the parked cars, but if you're going to take this lane, well it did as good of a job as I would have. When it had to move over, it overlapped the left lane by a couple of feet without cutting anyone off and did so very smoothly (generally, non a-holes in the left lane know you might need space so also won't stay right next to you).
* When we had to turn right later, it took the unmarked right lane (really, the right lane becomes as wide as two lanes where the left half continues straight, but with zero lane markings) no problem.
* After that turn, it started slowing down for a pedestrian walking to the median from the other side. I would not have slowed down since it's the middle of the road and not a safe place for anyone involved, but still, I let FSD do its thing. The moment FSD saw that the pedestrian was going to wait at the median, it started speeding up again. Like, maybe 200-300ms of reaction time, just like I would have done.

Of course it fell down in other ways:

1) It probably should not have stayed in that lane with the parked cars. Might be my fault as I always enable "minimize lane changes". Honestly, if we don't have a turn coming up, I want it to ask me for permission before it even puts the blinker on. It's a source of stress for me wondering if the car will decide to change lanes or not.

2) Slowed down for a stop sign that was only for the side road. It's a tough one because the two roads come together as a Y, and the stop sign is in-between the two roads, so the stop sign looks like it's for us. There are other hints, like the lack of a stop line on our side and continuous lane markings, that tell you it doesn't apply to us. I think that's why FSD only slowed rather than jamming the brakes.

3) The thing NEVER respects green arrows. It also sometimes tries to turn on red even though it's not allowed.

Still, overall I rate the drive an 8/10.

It's strange to me but it seems to handle more complicated driving situations better than simple ones, as long as there is traffic to follow. I think I would still prefer to just drive myself on most of these drives, but this makes me more optimistic for robot-taxi after all.
 
  • Informative
Reactions: JB47394 and rlsd
V12 is mostly good but after several days of use my biggest pet peeve has become its indecision around turn lanes, especially dual turn lanes. Not every intersection causes this behavior but when it does it will oscillate back and forth between lanes sometimes 2-3-4 times, or just kind of straddle the line taking up both lanes before committing at the very end.

Also I'm convinced the FSD driving profiles have zero effect on the e2e stack, including the "minimal lane changes" option, and only affects the legacy highway driving.
There's a couple double-left turns around here where there are zero or very faded lane markings through the intersection. It does the same behavior like you noticed and doesn't take a good line. I let FSD try it but I usually end up disengaging and sending my feedback.
 
  • Like
Reactions: tmar2323
I agree with both of your observations, I see that too.

The former will get better with more training. I believe for the profiles to have an effect (I too can find no difference whereas I did see a difference on v11) they need some sort of C++ guardrail code on acceleration. But I also admit to not knowing how else they could accomplish this using e2e.
Without being an AI expert myself, the way I envision it is that they would have a supervisor that implements rules-based logic and limits the outputs of the NNs. If the NN is asking for acceleration > some threshold, then the rules based can examine the outputs and see if it's doing it for some safety-based reason or if it's just accelerating too fast. The same if the NN suddenly wants to turn the steering wheel 90 degrees when driving on the highway. If so, it can just "disregard" the NN's command to the motors and clamp the max acceleration. There's a continuous feedback loop where the NN is updated on the current state of the car, so it would continue to update based on the new acceleration profile.

At some point, there needs to be some things done via rules, where a supervisor is telling the NN "yes you can do this" or "no you can't". How exactly that would work I couldn't say; it could be its own NN or something more procedural, but clearly you can't leave it all to some giant blob of a black box.
 
I don't know of anyone who was offered the free trial when on 2024.8.X. I don't believe anyone with the trial is trialling V11.
HellsKitchen on 4/2 reported that while on 2024.8.7 he signed up for the free trial and FSD 11.4.9 was enabled.

My curiosity got the better of me and I subscribed. I have 11.4.9
A couple of posts later he confirmed he was on 2024.8.7. Like I wrote, "someone reported..."

I noticed only that one report, so these question certainly deserves confirmation: Do 2024.8.x downloads include FSD 11? And, can someone on those versions start the free trial?