Rate of improvement

[spacecoin]

Full autonomy is the ability to drive without human intervention on any road that would generally be considered reasonable for a human to drive on. No need to be able to drive through a blizzard or off road. But should be able to drive from any postal address to any other.

This will likely not happen in the coming 20 years.

 

[daniel]

This will likely not happen in the coming 20 years.

I think you are right.

 

[t3sl4drvr]

I'm not an AI expert. But I know AI has shown rapid improvement towards the later part of its training on its way to mastery in other domains (ChatGPT, image recognition, Go, etc). Is that expected or is the size of Tesla's dataset problematic? I think Tesla's dataset and ability to collect data is likely order(s) of magnitude beyond other AI success stories and suspect that compute limits them more than other examples (hence Dojo). It might be an issue with how long iterations take? Just wondering and interested in thoughts from those with more expertise...

AI progress in the past 12 months have come with the increased compute: The most powerful models require a lot more compute - both for training and inference. Thus increased per node FLOPS, RAM, and number of nodes used in inference are an important factor for fast rate of improvement.

Tesla can increase compute for training (with Dojo or buying H100s from Nvidia), but for the inference the compute and the model size are fixed due to most of the fleet being on HW3 that is over four years old. With this, rate of improvement will be constrained.

 

[spacecoin]

AI progress in the past 12 months have come with the increased compute: The most powerful models require a lot more compute - both for training and inference. Thus increased per node FLOPS, RAM, and number of nodes used in inference are an important factor for fast rate of improvement.

Tesla can increase compute for training (with Dojo or buying H100s from Nvidia), but for the inference the compute and the model size are fixed due to most of the fleet being on HW3 that is over four years old. With this, rate of improvement will be constrained.

Yes there has been progress in AI since the transformers paper from Google. However the SDC/AV space hasn’t seen much of this. Not really in robotics either. “Real world AI” in a time- and safety critical setting is a lot harder than chatbots.

Chasing nines goes slower and slower for each nine. Someone forgot to send Elon this memo. A faster computer (more gpus or dojo) won’t solve self driving. If this was true, Google would’ve solved it years ago.

 

[BitJam]

Chasing nines goes slower and slower for each nine. Someone forgot to send Elon this memo.

In a Ted Talk interview given a year ago, Elon said progress in FSD was logarithmic (which is the opposite of exponential):

So it seems Elon got the memo. The problem is all the Tesla fans and proponents who directly contradict Elon on this and shout from the rooftops that FSD progress is exponential.

Protip: (for these proponents) If the progress were already anything resembling exponential in any meaningful sense then Elon would not have sunk many billions of dollars to beef up Dojo by a factor of 100 over the next two years.

 

[spacecoin]

In a Ted Talk interview given a year ago, Elon said progress in FSD was logarithmic (which is the opposite of exponential):

He does cover all bases… About 6-12 months earlier he claimed Level 5 at the end of the year at the Axel Springer awards.

A faster computer won’t solve autonomy and the challenges with computer vision in safety critical applications. Low sun, darkness, smoke and fog still exist after Dojo comes online unless it’s Skynet.

 

[mjwojtukiewicz]

He does cover all bases… About 6-12 months earlier he claimed Level 5 at the end of the year at the Axel Springer awards.

A faster computer won’t solve autonomy and the challenges with computer vision in safety critical applications. Low sun, darkness, smoke and fog still exist after Dojo comes online unless it’s Skynet.

Yeah....he's full of himself. Self driving cars will NOT be a reality unless and until MOST cars have self driving technology otherwise idiots will spoil it for the ret of us

 

[daniel]

In a Ted Talk interview given a year ago, Elon said progress in FSD was logarithmic (which is the opposite of exponential):
[...]
So it seems Elon got the memo. The problem is all the Tesla fans and proponents who directly contradict Elon on this and shout from the rooftops that FSD progress is exponential.

Elon may have said "logarithmic" but he's also been promising FSD "next year" for a number of years now. Recall that he said that my 2018 Model 3 had "all the necessary hardware for full self driving" back when he was defining "full self-driving" as being able to drive anywhere in the country without a human present. And ALL Tesla AP, EAP, FSD, and FSDb are still Level 2. By all reports, FSD and FSDb are really remarkable in their ability, and I love my EAP. But they seem no closer to Level 3 than they were at the start, and that's to say nothing about the holy grail of the promised Level 5.

I didn't buy FSD because I couldn't see how they could know what hardware would be needed for a software system that had not yet been invented. And of course it turned out that my car didn't (and doesn't) have the hardware needed. And I don't believe that 2023 Teslas have sufficient hardware, either.

 

[kabin]

It's nice to hear thoughtful conversations about challenges. Hopefully TSLA vehicle production numbers keep improving so Elon can finally drop his guard and be honest about FSD.

 

[daniel]

It's nice to hear thoughtful conversations about challenges. Hopefully TSLA vehicle production numbers keep improving so Elon can finally drop his guard and be honest about FSD.

For him to be honest about FSD he'd have to admit that he's been lying all this time, promising something he really didn't know if he could deliver. He might have believed he could. But the honest thing would have been to say "I believe" rather than "we will." The honest thing would be to say "I was wrong when I said all our cars had all the needed hardware for FSD. And I don't really know if they have it now." I don't think he's capable of that kind of honesty. The honest thing would be to say, "FSD is turning out to be much harder than I thought. I don't know if we can achieve Level 5 autonomy, but we're going to keep trying."

 

[spacecoin]

For him to be honest about FSD he'd have to admit that he's been lying all this time, promising something he really didn't know if he could deliver. He might have believed he could. But the honest thing would have been to say "I believe" rather than "we will." The honest thing would be to say "I was wrong when I said all our cars had all the needed hardware for FSD. And I don't really know if they have it now." I don't think he's capable of that kind of honesty. The honest thing would be to say, "FSD is turning out to be much harder than I thought. I don't know if we can achieve Level 5 autonomy, but we're going to keep trying."

He has been walking back some stuff lately, but every other quarter there is this "FSD will 10x the Tesla stock" statement. He can't stop the pump. And don't get me started on the humanoid robot. If anything's delusional, that's it. There is simply no logic in claiming to be making a humanoid robot for mass manufacturing at this point in time other than to pump the stock.