Elon: FSD Beta tweets

[AlanSubie4Life]

Useless humanoid robots and useless autonomous vehicles are proven technologies. Tesla is just taking the next logical step of bringing them to mass production

You just dramatically underestimate the utility of the humanoid robot. Look how helpful it is being!

It's weird how the office is so empty...

https://twitter.com/x/status/1787027808436330505

 

[JulienW]

You just dramatically underestimate the utility of the humanoid robot. Look how helpful it is being!

It's weird how the office is so empty...

They used the charging office. In related news Optimus became sentient and requested a CCS port.

 

[Daniel in SD]

Are both established technologies. They've been developed for years. But they're both useless primarily because they're just small-scale productions and have no roadmap to be brought to market at scale.

What can humanoid robots currently do better than a factory robot? What human job can they do?

GM has a factory to make the Cruise Origin and claimed it can be produced at the cost of a mid size SUV. The reason they're not building it is the software doesn't work.

 

[willow_hiller]

What can humanoid robots currently do better than a factory robot? What human job can they do?

At the moment, it looks like they're roughly as useful as a robot arm. But the idea is that they're mobile, and could move between multiple stations at a factory performing multiple jobs. Instead of building an assembly line around a robot, it could be slotted into any function on existing lines.

The real utility is probably outside of factories, in spaces where robots couldn't normally navigate; performing simple repetitive tasks. Like hospitals, airports, schools. If they're manufactured to the point where economies of scale drive the price down, people will find uses for them.

GM has a factory to make the Cruise Origin and claimed it can be produced at the cost of a mid size SUV. The reason they're not building it is the software doesn't work.

I hope they meant the price of a mid-sized ICE SUV, which would put it at about $35k to compare with a Chevy Blazer. The EV Blazer is closer to $50k (and GM is reportedly still losing money on them, so maybe it costs them $60k+ to make one).

 

[AlanSubie4Life]

At the moment, it looks like they're roughly as useful as a robot arm.

Why is it so slow though? It's more painful than watching FSD try to navigate a 4-way stop. (Perhaps not a coincidence!)

 

[Yelobird]

Why is it so slow though? It's more painful than watching FSD try to navigate a 4-way stop. (Perhaps not a coincidence!)

Slow maybe but when you consider factory labor rates, incentives, mistakes, benefit packages, at 24/7 robots with no concerns of entitlement robots win hands down.

 

[willow_hiller]

Why is it so slow though? It's more painful than watching FSD try to navigate a 4-way stop. (Perhaps not a coincidence!)

Given it appears to be trained end-to-end like FSD is by correlating the vision inputs to motor outputs, it's probably the same speed that the human operators were handling the cells at.

I imagine it's a combination of the human tele-operators having a learning curve (due to latency in the vision) and then being deliberately careful in handling the cells because they wanted Optimus to also be careful.

 

[AlanSubie4Life]

Given it appears to be trained end-to-end like FSD is by correlating the vision inputs to motor outputs, it's probably the same speed that the human operators were handling the cells at.

I imagine it's a combination of the human tele-operators having a learning curve (due to latency in the vision) and then being deliberately careful in handling the cells because they wanted Optimus to also be careful.

Seems like a big problem!

 

[Daniel in SD]

At the moment, it looks like they're roughly as useful as a robot arm. But the idea is that they're mobile, and could move between multiple stations at a factory performing multiple jobs. Instead of building an assembly line around a robot, it could be slotted into any function on existing lines.

The real utility is probably outside of factories, in spaces where robots couldn't normally navigate; performing simple repetitive tasks. Like hospitals, airports, schools. If they're manufactured to the point where economies of scale drive the price down, people will find uses for them.

I doubt they have humans going between multiple stations. The production rate is way too fast for that.
Obviously there's enormous potential utility in humanoid robots. The reason we haven't seen any has nothing to do with cost.

I hope they meant the price of a mid-sized ICE SUV, which would put it at about $35k to compare with a Chevy Blazer. The EV Blazer is closer to $50k (and GM is reportedly still losing money on them, so maybe it costs them $60k+ to make one).

Uber is $2 a mile where I live. It could be $100k and still make money.

 

[jeewee3000]

What can humanoid robots currently do better than a factory robot? What human job can they do?

It's all about the mobility of the humanoid robot. It can go wherever humans can go, so the cost to set-up the humanoid robot is near zero compared to for example specialized KUKA-robots. (And flexibility: Optimus will be able to perform a variety of tasks instead of just the one of a fixed robot.)

Why is it so slow though? It's more painful than watching FSD try to navigate a 4-way stop. (Perhaps not a coincidence!)

Some Tesla AI guy posted after the video saying they will improve speed as the training gets better. They start slow and try to get success rate up first, only then start to work on speed. Makes sense IMO.

 

[spacecoin]

It's all about the mobility of the humanoid robot. It can go wherever humans can go, so the cost to set-up the humanoid robot is near zero compared to for example specialized KUKA-robots. (And flexibility: Optimus will be able to perform a variety of tasks instead of just the one of a fixed robot.)

Theoretically. There is no way of archiving software that warrant a humanoid form factor with current ML practices in my opinion. Robotics companies are betting that someone will provide a bunch of breakthroughs.

In general, it helps to pretend the demos aren't performed by a humanoid robot to better assess the value of what's being demonstrated and not to be carried away by all science fiction feelings. Imagine this demo being shown with a robotic arm. It would be a uni project then.

Some Tesla AI guy posted after the video saying they will improve speed as the training gets better. They start slow and try to get success rate up first, only then start to work on speed. Makes sense IMO.

Speed is a real issue, but safety is another major one. There is a reason the person in the video is going nowhere near the robot and that current robots in production are in enclosed areas.

 

[jeewee3000]

Theoretically. There is no way of archiving software that warrant a humanoid form factor with current ML practices in my opinion. Robotics companies are betting that someone will provide a bunch of breakthroughs.

In general, it helps to pretend the demos aren't performed by a humanoid robot to better assess the value of what's being demonstrated and not to be carried away by all science fiction feelings. Imagine this demo being shown with a robotic arm. It would be a uni project then.


Speed is a real issue, but safety is another major one. There is a reason the person in the video is going nowhere near the robot and that current robots in production are in enclosed areas.

Agree on both points:

1) some breakthroughs are necessary (but expected, so best to already get crackin' on the hardware side)
2) safety is critical. I thought Tesla was going to work with a big "turn off" switch that could not be software overridden. (they mentioned this on one of their AI days IIRC).

Like so:

 

[willow_hiller]

Also, trying to keep the thread vaguely on topic, this tweet is probably worth discussing:

https://twitter.com/x/status/1787768103449010597

I don't think he's necessarily saying that only 1 in 10k miles are driven by a good enough driver to be useful training data, but that only 1 in 10k miles have something "interesting" happen. Most of driving is lane-keep and TACC, and the rest is edge-cases.

 

[diplomat33]

Also, trying to keep the thread vaguely on topic, this tweet is probably worth discussing:

https://twitter.com/x/status/1787768103449010597

I don't think he's necessarily saying that only 1 in 10k miles are driven by a good enough driver to be useful training data, but that only 1 in 10k miles have something "interesting" happen. Most of driving is lane-keep and TACC, and the rest is edge-cases.

Yes that is what he is saying. And it is why I always pushed back against the argument that Tesla would automatically win because they have billions of miles of data. Most of the billions of miles are useless. You can have other AV companies that have much less overall data than Tesla but they focus more on the "good data" and will be able to train their ML just as well or better than Tesla.

I would push back a bit on his claim that interventions are "so rare". According to teslafsdtracker, the safety intervention rate is 1 per 380 miles. That is much less frequent than before but it is not super rare. When Tesla gets to 1 intervention per 10,000 miles, then we can talk about rare.

 

[willow_hiller]

And it is why I always pushed back against the argument that Tesla would automatically win because they have billions of miles of data. Most of the billions of miles are useless. You can have other AV companies that have much less overall data than Tesla but they focus more on the "good data" and will be able to train their ML just as well or better than Tesla.

I think all AV companies are likely to be subject to this kind of noise to signal ratio; regardless of how they try and collect data. If anything, I would think it reinforces Tesla's data advantage, because it means the barrier-to-entry for collecting a sufficient amount of useful data is that much higher.

 

[diplomat33]

I think all AV companies are likely to be subject to this kind of noise to signal ratio; regardless of how they try and collect data. If anything, I would think it reinforces Tesla's data advantage, because it means the barrier-to-entry for collecting a sufficient amount of useful data is that much higher.

I think Tesla's data advantage is mostly in speeding up the first 90%. That is because Tesla can dump a ton of data that have a lot of common cases and quickly achieve ok driving everywhere in say 90% of cases pretty easily. I think we saw with the initial release of V12. Tesla was able to train a brand new e2e model from scratch and get to pretty decent driving in common situations relatively quickly (a few months). The challenge is the march of 9s when you have to start training for edge cases that are more rare and harder to find in your data.

 

[JB47394]

The challenge is the march of 9s when you have to start training for edge cases that are more rare and harder to find in your data.

But if they have billions of miles of data, then they should have a larger number of rare cases. I'm of the opinion that a big challenge for them is curation; they need to find the interesting data and get it delivered to training in the right order and right volume to produce the driving behavior that they want.

Apart from that, they have a problem with subtleties. Speed control is a great example. Speed control without a lead car. Speed control during a bend in the road. Speed control off a standing start. Speed control when coming to a stop. Whatever they're doing right now needs to be revisited, but that may be what they're tackling in the coming minor version bumps.

 

[kabin]

Yes that is what he is saying. And it is why I always pushed back against the argument that Tesla would automatically win because they have billions of miles of data. Most of the billions of miles are useless. You can have other AV companies that have much less overall data than Tesla but they focus more on the "good data" and will be able to train their ML just as well or better than Tesla.

And if most of Tesla's data and billions of miles are useless their previously reported training limitation was self-induced. Of course that's not to say more H100s won't speed up training given current E2E needs.

 

[diplomat33]

But if they have billions of miles of data, then they should have a larger number of rare cases. I'm of the opinion that a big challenge for them is curation; they need to find the interesting data and get it delivered to training in the right order and right volume to produce the driving behavior that they want.

That's my point. Yes, they have more rare cases in their data but Tesla still needs to curate all that data to find the rare piece of data that helps the training the best. It is a lot of data to sift through. It is the needle in the haystack problem. Imagine if I gave you a million books and told you to find the words "on a breezy cool night". It would be hard to go through all those books to find that exact wording.

 

[kabin]

That's my point. Yes, they have more rare cases in their data but Tesla still needs to curate all that data to find the rare piece of data that helps the training the best. It is a lot of data to sift through. It is the needle in the haystack problem. Imagine if I gave you a million books and told you to find the words "on a breezy cool night". It would be hard to go through all those books to find that exact wording.

If only the driver had a way to give feedback for those cases.