This is why Tesla is ahead on AI:FSD

[Crissa]

The whole video is good, but...


The synthetic world allows them to create unsafe and crazy interactions to test the AI against. The fleet sampling (about thirty seconds before the time point) shows that their AI knows what things are similar and can be queried. This is great news! Their AI will be able to plot and predict movements over everything else in the scene, and even label stuff it doesn't understand. And Dojo will basically be rebuilding the training data repeatedly, more quickly, so that each car will have a new and smarter AI instead of being stuck with what they have.

It's a huge multi-prong approach.

Still far from robotaxis, but this is why visual data is better than radar.

-Crissa

Sponsored

 

[Ogre]

If nothing else, this video should assure our Canadian and Alaskan readers that Tesla is indeed testing Moose encounters.

 

[electricAK]

Thanks for sharing this video Crissa. It describes their AI and self driving systems pretty damn well in 13 minutes. Not easy to do.

I've spent a little time in the engineering, data processing, and software worlds, and what Tesla is doing here just blows me away. I always knew you could do this stuff, but I'm fully impressed at how well it is working and how quickly Tesla is making it happen.

I've also worked on aircraft doppler radar systems/data. Tesla's visual approach is simply far more powerful. Radar makes it easy to see an object and its velocity, but it's very hard to tell what the object is and where its boundaries are. That extra incomplete data would only confuse the visual model that Tesla has created.

This video gives me a new confidence that Tesla made the right call when they decided to ditch lidar, and now radar.

 

[Crissa]

If nothing else, this video should assure our Canadian and Alaskan readers that Tesla is indeed testing Moose encounters.

Yeah! That's what I posted it.

I keep wondering about those tests where they send a flying waggling mannequin at a car and expect it to stop and two things come to mind: How do they make sure the mannequin triggers are the same stopping distance and why do they think that flapping thing looks like something a driver should avoid?

-Crissa

PS, Tesla isn't the only ones doing this, and their approach has its faults. But it doesn't rely on just simulation or just labeling or just mapping or just machine learning. Which is what I thought this video focused on quite well.

When I saw their AI fleet query, that's when I knew they were ahead. That plus Dojo are two things no one else has. Then being able to build simulated and synthetic worlds with that resolution data so deeply means the AI won't have selection bias. They'll be able to mash that selection bias out of it.

 

[John K]

Observations:

Tesla can pass a “I am not a robot” test
everyone should fear, one Tesla vehicle experiences a situation, all learn from it. Cut off a Tesla, they all swarm for revenge.

”Holy mother of papers!”

The vid is one big, “I told you so!”

Rejecting claim robot is a joke on us, I consider the robot an iron clad promise and waiting to reserve it.

Beyond fascinating.

Sponsored