Project Dojo

[EVNow]

In parallel, it runs a simulated city with 3,500 instances of the current NN where any instsnce can go rouge at random and purposely drive poorly at various levels from mild deviation to wrong way on the expressway to aiming for other vehicles.

They don't need to go rogue, as best as I can tell

 

[mongo]

They don't need to go rogue, as best as I can tell

Gotta test those humans' reaction times. Both other car's and the driver's.

 

[DrDabbles]

I suggest you listen to the AI podcast by Lex Fridman's interview with Yann. Hopefully you will consider Yann, recipient of the Turing Award for his work on deep learning someone who understands what's going on.

Yann LeCun: Deep Learning, Convolutional Neural Networks, and Self-Supervised Learning | MIT | Artificial Intelligence Podcast

You should watch that again and pay closer attention. I watch all of Lex's material, as well as autonomy and AI conference talks, and read research papers.

 

[EVNow]

Gotta test those humans' reaction times. Both other car's and the driver's.

If drivers don't intervene, AP naturally goes rogue once in a while.

 

[EVNow]

You should watch that again and pay closer attention. I watch all of Lex's material, as well as autonomy and AI conference talks, and read research papers.

So, what do you think about self-supervised learning for FSD and how long it would take ?

He didn't definitely say this ...

The bleak reality that we're decades away from having anything resembling reasonable AI, and that we may never have a system complex enough to actually handle level 5 driving. While you see pessimism, I see realism. I think you're vastly overoptimistic and it's likely because you don't understand the problem at hand.

 

[DrDabbles]

So, what do you think about self-supervised learning for FSD and how long it would take ?

He didn't definitely say this ...

To do it with vision neural networks responsible for driving? We don't even know if it's possible to use self supervised learning. Which is what I said already. This is one of the criticisms I have with Dojo which otherwise could be an extremely powerful tool. But what it means is that Tesla isn't just trying to have a breakthrough with neural networks and vehicle autonomy, but also with self-supervised learning systems.

Like Lex mentioned in the video, even humans performing self supervised learning often get it wrong. Think of the saying "practice makes permanent" rather than "practice makes perfect". Without proper coaching, no amount of practice will improve anything anybody does.

 

[EVNow]

To do it with vision neural networks responsible for driving? We don't even know if it's possible to use self supervised learning. Which is what I said already. This is one of the criticisms I have with Dojo which otherwise could be an extremely powerful tool. But what it means is that Tesla isn't just trying to have a breakthrough with neural networks and vehicle autonomy, but also with self-supervised learning systems.

Like Lex mentioned in the video, even humans performing self supervised learning often get it wrong. Think of the saying "practice makes permanent" rather than "practice makes perfect". Without proper coaching, no amount of practice will improve anything anybody does.

Yann talked about this. SO looks like you are the one who needs to listen again.

ps : My position is we don't know how long it will take. So, I think anyone saying it will take 1 year or 3 decades are just BSing.

 

[DrDabbles]

Yann talked about this. SO looks like you are the one who needs to listen again.

ps : My position is we don't know how long it will take. So, I think anyone saying it will take 1 year or 3 decades are just BSing.

Did you hear the conversation he and Lex had about it? Yann said we can, Lex pointed out several place where ML can not, as well as biological systems that can not. Other interviews Lex has had have described the folly of self-reinforcing ML. I agree with your position, but so far there is nobody doing real self-reinforcing ML on vision based autonomous vehicles. That's not because nobody's trying.

This also gets back to the simulation response that Elon gave on Autonomy day. Several facilities are using machine generated simulations to give random scenarios to vehicles for training purposes. But Elon hit the nail on the head- If you could build a simulator complex enough, you'd effectively be building something indistinguishable from reality. My point here is that if you could build a ML system that corrects and retrains itself, you'd effectively have solved the problem of general AI, which absolutely nobody believes we're anywhere close to doing.

 

[electronblue]

This is why it's still a research problem and not an implementation problem.

Absolutely.

 

[electronblue]

That has nothing to do with the topic of the thread. You continue to be a concern / short troll. I wish this sub-forum was moderated to take care of people like you.

If you are not interested in this topic, why even bother to reply ? Why bring up FUD ? Why resort to slander ?

It is absolutely relevant to the topic to consider the range of possibilities on how truthful or realistic Tesla was about their ”Level 5 no geofence feature complete in 2019” status on Autonomy Investor Day — and thus how that relates to realisim of Dojo helping to fast-track things. I do not know whether they were truthful or realistic and that is why I said it is an open question.

It would have been absolutely relevant to consider that also in October 2016 when Tesla made the original announcements based on which many of us bought EAP and FSD and bought into the original timelines (full EAP shipped in December 2016 as said in Design Studio, coast to coast demo end of 2017, FSD pack differentiating features in 3-6 months in 2017 and so on).

We have no original FSD buyer differentiating features in September 2019. We have no coast to coast demo in September 2019. We haven’t even reached full EAP in September 2019 (Enhanced Summon). All things which Tesla at some point claimed for 2016-2017. 2020 may well roll in without some of these happening, that is a full lease life of a car late — or more.

We Tesla buyers did not sufficiently consider their truthfulness or realism back then, I would argue — with 20-20 highsight. Yes, we were fooled then. I would just prefer to not be fooled again. I would say that is a perfectly reasonable wish.

 

[Buckminster]

 

[Buckminster]

Pretty interesting discussion about the DOJO project.

There was not much new information, but here are the key takeaways:

• DOJO is Tesla's investments to get most computation per economical cost in the next few years . Cheap Computation is essential in terms pf cost and time.
• Geometric Video data for labelling Improves productivity of labellers by 100-1000 times. You don't need neural network to do the labelling. Leveraging labellers is the strategy here.
• SLAM is computationally very expensive.
• Designing HPC infrastructure is not simple. Requires lot of specialized tech. That's why DOJO project is probably taking longer than they thought.
• Key is the ability of the systems to querry the relevant/similiar data from the fleets enabling self supervision
• As the edge cases become rarer, the networks ability to detect an edge cases becomes simultaneously more powerful
• Karpathy's Data cycle seems to be 1-2 weeks long at the moment.
• Neural networks weight updates only happen with relatively big pushes.

 

[Buckminster]

Could Dojo be using Cerebras chips?
Cerebras’ wafer-size chip is 10,000 times faster than a GPU