Lex Fridman explains Tesla Dojo

Please support this podcast by checking out our sponsors:

GUEST BIO:
Jeffrey Shainline is a physicist at NIST working on. Note: Opinions expressed by Jeff do not represent NIST.

PODCAST INFO:

SOCIAL:

[joncardell]

Not only the focused application, but Tesla’s sense of urgency and their engineering bandwidth makes me excited for DOJO.

[jmatt]

Can’t wait for GM and Ford to launch their version of Dojo. Made with union labor of course and with the blessing of Biden.

[Beachtrader]

Tesla can also run a different version of fsd in the background in "shadow mode" over the entire fleet to test its efficacy. We gotta get you into fsd beta. We are seeing the iterations live. 10.2 is out now for those with 100 safety score. 10.3 is due out next week and we should see some more allowed into the fsd beta. This is the city driving, it already did highways.

[miraculixxs]

He thinks too statically about models - these are software models, not hardware. What gets sent to cars are a combination of software + data (eg weights + hyper params) and then loaded unto the car's hardware. It is not necessarily the very same model for each car, eg models could be seggregated by geography, region, city. I don't have specific insight, just noting that's possible.

[parkeranderson]

This is one of the best Lex episodes ever. Extremely interesting and entertaining listening to these guys' conversation.

[seattlevkk]

If it’s truly modular then the representation of the domain should be separate from the training. Is dojo specific to vision or is it more general than that and if so how have they separated the representation of the domain? That sounds challenging

[larrythelobster]

Can't wait for your interview w Elon. We really need someone w your understanding of technology/software to dig deep into FSD and real world AI w him. Going to be great

[dr.doppeldecker]

I thought Elon was crazy when he had his Tesla Robot presentation, but opening Tesla Karate Dojo's is taking things too far imho.

[The_Conspiracy_Analyst]

9:21 The problem with that is with the current approach using things like Neural Nets and such, that's impossible. That's their Achilles heel -- they are very difficult if not to troubleshoot and debug when we're talking about isolating faults. That's the chaotic nature of these types of approaches. And there's nothing that you can just slap on there to make that not true. So when a self crashing car wraps itself around a telephone pole, there's no source code to go through and isolate the fault. It's just a big, opaque binary blob. And there's nothing to pin down to say "ok why did you do it this time and not the other 10, 000 times". Heisenbugs!

[jordanschultz]

The doing architecture of the chip design was underappreciated

[nahiag]

We need to get lex into a video with the FSD beta.

[imrickjamesb]

"True perfection has to be imperfect. I know that sounds foolish, but it's true."


Noel Gallagher

[matthiasdollinger]

Is possible to integrate all edge cases (regarding autonomous driving) into the neural nets over time or will there be a point, where you improve on one problem generating edge cases but you get worse on another one? I have absolutely no AI expertise. But that's one thing I've been wondering for quite some time.

[bukurie]

Thank you for interview for Advanced TESLA❤Dojo🌏

[elenabob]

4:10 that's the standard process to put into production a new tuning except that on other constructors we have our own fleets and we don't let the client test them.

[woolfel]

If you look at the IO design of DOJO, it's really built to handle large video streams. I doubt NL use cases would benefit from this kind of massive IO. If you need terabit IO, there are very few systems that will handle that well.

[multiversehacking]

Self driving cars ubiquitous on the roads has always been "5 or 10 years down the line." 6:50

[balaji-kartha]

Dojo will help bring robocars into the market and make driving itself obsolete so quickly that this is going to surprise everyone

[thosoz]

Edge cases like the tree branch on the road are one thing.
The other drivers are another.
Once the system is proven 10X better than humans and
ALL 'accidents' are caused by group above, the cost of insurance
for human drivers will wipe out most peoples desire to drive.

[yonusa]

Ok, I see it now,
As someone who had had 2 hair transplants. I can see what people mean by there being a deadline. 99% of the time the skull changes it’s form and hair where it used to be just looks ridiculous af.
Learned a bit, hair distracted me more than an it driven by a specific application 😅