And Waymo has that whole time to bring down their costs of construction, learn how to solve problems of scale, and so on.
I agree that a CyberCab would have structural advantages if the service were at parity w/ Waymo, but that is a very big hypothetical!
I know which one I'd back.
Yeah, I could be wrong and they're just looking for the lowest risk profile. But, riders seem to be pleased and accident rate has been impressively low.
Is it easier to ramp to 1M vehicles a year and cut BOM costs by 80%, OR, install more GPUs and collect more data?
It's not a hardware problem for Tesla. They are a trillion dollar vertically integrated manufacturing powerhouse.
Thus it's a hardware problem.
If Tesla can solve FSD with only cameras, then yes, their ability to scale is much better than Waymo’s.
But that is a stupendously large “if”!
I’ve seen the Karpathy quote before, and I think it’s kind of hilarious to compare one company which does X0,000 robotaxi rides per day, to another company which has a robotaxi maybe entering production in two years, and say, “well look, they both have problems!”
Bit of a misleading simplification, no? Tesla boosters assume they will crack FSD, it’s just a matter of time. That seems like a big leap to me.
Andrejs point is that software problems are easier to scale your way out of. I think we can both agree there's plenty of precedent for this line of thinking.
A company cannot as easily scale their way out of lowering material costs by an order of magnitude or go from 1000 cars a year to 1M cars a year.
I’ve read The Bitter Lesson, I understand the argument “the human eyes are just cameras, and we do just fine”; I’m not totally unsympathetic to the idea that Tesla is better-positioned than some people think.
But I still think it’s a big leap to suggest Tesla will definitely solve FSD by scaling / software, and Waymo is fundamentally screwed and has no way to innovate out of its own scale problems.
Companies lower material costs by an order of magnitude all the time! Look at battery/solar over the last 15 years — in many areas we see two OOMs of cost reduction. It’s a challenge; but one that seems on par with creating software to replicate the optical capabilities of the human brain.
We're surrounded by AI miracles, this website is full of examples, yet people can't conceive of Tesla reaching this finish line. It's pretty odd.
I've clocked 20k miles on FSD. In the brief period I've used it I've seen step improvements. I guess in my mind it's a foregone conclusion with no breakthroughs required. All I need to believe is on the Tesla earning reports, which is Tesla data center build progress (training) and vehicles sold (data)