Elon Musk Says Tesla FSD AI Having Exactly The Problem I Thought 2 Years Ago It Was Having

Join daily news updates from CleanTechnica on e-mail. Or follow us on Google News!

Through the Q&A session of the 2024 Tesla annual shareholder assembly, in response to one of many questions on Full Self Driving (FSD), CEO Elon Musk defined one thing about Tesla’s FSD improvement that utterly affirms a suspicion I had two years in the past. Additionally, this is among the key challenges that made me a bit involved about whether or not Tesla’s method will result in robotaxi functionality.

I’ll begin with what Elon said, emphasizing in daring the important thing half I’m highlighting. “Because the system will get higher, it will get tougher to determine which AI mannequin is healthier. As a result of now, like 1000’s of miles between interventions, how can we, as rapidly as potential, work out which AI mannequin is healthier. And once you make these completely different AI fashions, they’re clearly not, like, tremendous deterministic. So, you might have one mannequin that eliminates one drawback, however creates one other drawback.” I’ll come again to the remainder of this phase later and the way Tesla is coping with the difficulty and attempting to get round it, however I simply wish to emphasize right here that it’s not merely the problem of selecting the most effective mannequin, however the problem of 1 drawback getting fastened whereas one other drawback is created. Let’s return briefly to what I wrote more than two years ago after I began noticing this drawback:

“Mainly, I’ve been attempting to grasp the explanation why Tesla FSD isn’t nearly as good as I’d hoped it could be by now, and why it typically will get considerably worse. One potential concern is what I’m calling the ‘see-saw drawback.’ If my concept is right to any notable diploma, it may very well be a essential fault in Tesla’s method to widespread, generalized self driving.

“My concern is that as Tesla corrects flagged points and uploads new software program to Tesla buyer vehicles, these corrections create points elsewhere. In different phrases, they’re simply enjoying software program see-saw.”

I don’t know the way a lot that was occurring two years in the past in comparison with at this time, however with Musk acknowledging that it’s occurring, let’s check out how he says Tesla is attempting to unravel the issue.

“So, we’re attempting to unravel this by a mix of simulation; importing fashions, having them run in shadow mode — so, it’s really sort of useful not everybody has Full Self Driving, as a result of we are able to see … we are able to run it in shadow mode and see ‘what would this new mannequin have finished in comparison with what the person did.’ So, since we’ve received hundreds of thousands of vehicles that we are able to do that with, that provides us a delta between what the AI mannequin predicted it could do and the person would do. And if you happen to sort of sum up the errors between them, you may see, ‘oh, there was an even bigger error stack from this mannequin versus that mannequin’ once you uploaded them into — we every uploaded them into 100,000 vehicles.”

He then goes right into a bit extra element on how they’re attempting to advance from mannequin to mannequin. “However that’s the largest limiter proper now — it’s not coaching, it’s not knowledge, it’s really testing the AI fashions. After which determining intelligent methods to determine if a brand new mannequin is healthier or not. Like we all know like there are specific intersections which might be troublesome. […] Like, all of the fashions shall be nice at driving on a effectively painted highway with no intersection. They’ll all be nice at that. […] So, we’ve recognized a number of thousand intersections in america the place we’ll really take the brand new mannequin and take a look at the automobile in these very troublesome intersections.”

He additionally added that they haven’t but uncorked any of the additional means of FSD “{Hardware} 4,” however says they’ll do this later this 12 months.

General, the large query with regard to figuring out new higher and higher fashions, and overcoming the aforementioned “see-saw drawback,” is how rapidly Tesla can do that to be able to roll out unsupervised full self driving and robotaxis. Will it take a 12 months? 10 years? Clearly, Elon Musk is optimistic they’ll get there quickly, however as even Musk states, he’s chronically overly optimistic. What’s your estimate?