FSD v12.x (end to end AI)

[ArchHamster]

Yeah, I'd like confirmation on this as well. I love this feature on my Model X

Given Tesla's penchant for removing features to cut costs, I'm worried about when we get Tesla Vision. Elon might remove code support for the ultrasonic sensors that allow the doors to open fully automatically. Especially since he has made no such promise about this feature returning to ussless cars.

 

[kabin]

There's FSD/China news today about Baidu providing local mapping and data collection. Not sure how they will train data given restrictions on NVDA GPUs and China limiting data exporting. Otherwise China has okayed FSD use on Chinese roads. Assume it all comes to pass this might be a quick boom or bust for FSD and Tesla depending on how Chinese TSLA owners and media respond.

 

[JB47394]

My first drive on 12.3.6 showed no changes in basic driving behavior, as I'm sure we would all expect from a point relese. The differences I noticed were:

1. Vision Park Assist on my USS Model 3. I love the visuals, but it failed on the first parking attempt. I didn't care for the creaking and groaning as it regularly cranked around the wheels while stationary, and I won't be using it to park until they sort that out. In general, the assist is unwilling to consider driveways as useful space for maneuvering. I think the failed first attempt gave up because it couldn't move forward enough, despite there being driveway space it could have used. Later, when I got home, I noticed that the park assist doesn't even show driveways, presenting them as yard space with a curb right across the opening.

2. At the end of an FSD drive, the car will stop and tell you to press the accelerator to resume (but it stays in FSD). So you don't have to take it out of FSD to prevent it from continuing on its own after reaching your destination. They also added a curious arrow in front of the car in the visualization when you're at your destination. I'm not sure what that's supposed to indicate. My car parked about 100 feet from my house because there was no curb space at my house, and it apparently didn't want to stop in front of my driveway opening.

Oh, and I got a proximity STOP and chimes as I backed into my driveway. I sometimes used to get that when pulling out, so I guess my driveway spacing is just a bit too tight for its tastes.

 

[willow_hiller]

There's FSD/China news today about Baidu providing local mapping and data collection. Not sure how they will train data given restrictions on NVDA GPUs and China limiting data exporting. Otherwise China has okayed FSD use on Chinese roads. Assume it all comes to pass this might be a quick boom or bust for FSD and Tesla depending on how Chinese TSLA owners and media respond.

Maybe the US-trained FSD is the foundation model, and Tesla just needs to fine-tune for the Chinese context. Fine-tuning can require magnitudes less compute than training a model in the first place.

 

[rlsd]

My USS car is downloading 12.3.6 now but won't be able to drive it till Saturday.

You should call Tesla road service.

 

[JB47394]

Maybe the US-trained FSD is the foundation model, and Tesla just needs to fine-tune for the Chinese context. Fine-tuning can require magnitudes less compute than training a model in the first place.

How does one fine-tune a neural network? Do you mean by piling on additional training data that will move the neural network to conform better to the environment on Chinese roads? I wonder how well that would work. Hmmm. I wonder what they'll learn about creating true foundation models which cover the basics of "stay on the road", "go around obstacles" and such, and building on that for a variety of behaviors. For example, how about also training for chill, average and assertive driving styles?

 

[kabin]

Maybe the US-trained FSD is the foundation model, and Tesla just needs to fine-tune for the Chinese context. Fine-tuning can require magnitudes less compute than training a model in the first place.

Or maybe Baidu guarantees the data is cleansed prior to export? It's potentially an endless stream of data to move back/forth in the hopes of getting it right. Lots of different laws, roadway construction, roadway hardware, signage. I would guess it would require a major rebuild and/or very strict ODDs.

 

[willow_hiller]

How does one fine-tune a neural network? Do you mean by piling on additional training data that will move the neural network to conform better to the environment on Chinese roads? I wonder how well that would work. Hmmm. I wonder what they'll learn about creating true foundation models which cover the basics of "stay on the road", "go around obstacles" and such, and building on that for a variety of behaviors. For example, how about also training for chill, average and assertive driving styles?

Yes, fine-tuning is just a matter of continuing to train a model on some context-specific or task-specific data.

For very large foundation models, there's an efficient method of fine-tuning called LoRA (Low-Rank Adaptation) where instead of modifying the weights of the foundation model directly, you train and store matrices that represent the differences in the weights. You can even change the precision of the adapter matrices (if the original model weights were stored in float32, you could represent the adapters as int8, for e.g.), at which point it's called quantized low-rank adaptation (QLoRA).

EDIT: If this is useful context, one LLM that I've personally fine-tuned before was the first version of Mistral-7B. The base model was trained using 8 Nvidia A40 GPUs (total of 384 GB VRAM). I was able to fine-tune it using a single Nvidia V100 GPU with 16 GB of VRAM or about 4% of the original compute.

 

[kabin]

This topic probably needs a different thread but Eunice Yoon of CNBC reports FSD will launch in China 'soon.' TSLA is awaiting Chinese approval to export data for FSD training. Gotta wonder if TSLA/Elon requested opaque FSD statistics reporting.

 

[JulienW]

This morning I had an UPL that FSDS totally misjudged and waited until a car was closing to try and pull out. FSDS seems so good at everything expect inner city multilane selection and crowded UPL. Chuck's latest UPL (this morning) is F'n UGLY and more like a Steven King nail biter movie.

 

[tmoz]

This morning I had an UPL that FSDS totally misjudged and waited until a car was closing to try and pull out. FSDS seems so good at everything expect inner city multilane selection and crowded UPL. Chuck's latest UPL (this morning) is F'n UGLY and more like a Steven King nail biter movie.

Has any aspect improved from 12.3.3 to 12.3.6, other than adding parking? Not good statistical sampling on my part, but I've heard more negatives than positives, but now I can't recall aspects that got reliably better, which makes me ask if there was some aspect that got reliably better.

 

[D Good]

They are realizing that hard-coding actually works in some scenarios. Now that they can no longer hard-code chuck's turn, they have to actually train a car that is "smart" enough to complete it the right way. In a scenario like this, the latter is the much more difficult method for producing decent results.

 

[JulienW]

Has any aspect improved from 12.3.3 to 12.3.6, other than adding parking? Not good statistical sampling on my part, but I've heard more negatives than positives, but now I can't recall aspects that got reliably better, which makes me ask if there was some aspect that got reliably better.

People tend to OVER believe in a minor or bug update. In order to add many features or improvements it will be a 12.x update. So 12.4.x will really be the next hope for meaningful improvements.

 

[Ramphex]

Lane selection has been terrible since the 3.25 update. It keeps jamming me into the right lane behind slow trucks on non-interstate roads. Trying to use the left turn signal to cancel the lane change also no longer works, so have to either disengage or wait for it to finish switching lanes and then indicate for it to go back to center lane.

I’m disappoint.

 

[stealthyc]

I was behind a different company's autonomous vehicle yesterday and we both made a left turn onto a two lane (each way) road. The speed limit on the road was 35 but the other autonomous vehicle was driving at about 20 with both flashers on. My car moved to the other lane as if it were going to pass it (as it should) but refused to go any faster than it, which I found very amusing. After a little bit of time with us both just crawling along in parallel, I used the accelerator pedal to pass it, and then everything was back to normal.

 

[Twiglett]

still struggling with folks calling Chuck's turn an "UnProtected Left turn"

 

[MP3Mike]

Has any aspect improved from 12.3.3 to 12.3.6, other than adding parking?

Yellow light handling is much better in 12.3.4 than it was for 12.3.3, at least for me.

 

[tmoz]

People tend to OVER believe in a minor or bug update. In order to add many features or improvements it will be a 12.x update. So 12.4.x will really be the next hope for meaningful improvements.

Yes, 12.4.x will be the acid test.

 

[AlanSubie4Life]

This morning I had an UPL that FSDS totally misjudged and waited until a car was closing to try and pull out. FSDS seems so good at everything expect inner city multilane selection and crowded UPL. Chuck's latest UPL (this morning) is F'n UGLY and more like a Steven King nail biter movie.

No reason to expect anything different. This is not a regression from prior version. Everything looks essentially the same. Exactly what we would expect.

3/7, standard behavior of stopping in the middle of the road, arguably new behavior of pre-rolling the merge. (Probably just a side effect of other prompt engineering.)

Maybe they’re pre-emphasizing to reduce latency, .

12.4 will be the most consequential release in FSD history. The final test. It has been crafted for months. Essentially flawless, looks nearly like a human driving but looks like they still have to fix the stop pose.

https://twitter.com/x/status/1785007281010028606

 

[diplomat33]

The new auto park looks great. I love the visualization and being able to select the parking spot you want now. And the car executed the parking maneuver very smoothly IMO. The disengagement with the oncoming vehicle, not so great.

https://twitter.com/x/status/1784715428050108903