4 ms·

"Tesla has a software problem, Waymo has a hardware problem" - Andrej Karpathi Is it easier to ramp to 1M vehicles a year and cut BOM costs by 80%, OR, install

"Tesla has a software problem, Waymo has a hardware problem" - Andrej Karpathi Is it easier to ramp to 1M vehicles a year and cut BOM costs by 80%, OR, install more GPUs and collect more data?

datadrivenangel 2y ago
If you have to install the GPUs in each car, then you also have a hardware problem.
- delabay 2y ago
  Tesla has been mass producing bespoke AI hardware for four generation. It's not a hardware problem for Tesla. They are a trillion dollar vertically integrated manufacturing powerhouse.
  - datadrivenangel 2y ago
    Not an insurmountable problem for sure. If Level 4 self driving requires more compute in the car, someone has to physically put more compute in the car, which may require redesigning something to do retrofits, etc. A "FSD" car today may not actually have Full Level 4 Self Driving when Tesla finally rolls it out. Thus it's a hardware problem.
ketzo 2y ago
You pose your comment like the answer is obvious, but I don’t know that it is. 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.
- delabay 2y ago
  It's "hardware is hard" versus "scaling laws". 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.
  - ketzo 2y ago
    But the implicit assumption is that Tesla’s problems can be solved by scaling laws. That seems to me like a big assumption, no? 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.
    - delabay 2y ago
      Waymo (and my lived experience with FSD) is proof the end goal is obtainable. 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)

"Tesla has a software problem, Waymo has a hardware problem" - Andrej Karpathi Is it easier to ramp to 1M vehicles a year and cut BOM costs by 80%, OR, install more GPUs and collect more data?

datadrivenangel 2y ago

If you have to install the GPUs in each car, then you also have a hardware problem.

delabay 2y ago
Tesla has been mass producing bespoke AI hardware for four generation. It's not a hardware problem for Tesla. They are a trillion dollar vertically integrated manufacturing powerhouse.
- datadrivenangel 2y ago
  Not an insurmountable problem for sure. If Level 4 self driving requires more compute in the car, someone has to physically put more compute in the car, which may require redesigning something to do retrofits, etc. A "FSD" car today may not actually have Full Level 4 Self Driving when Tesla finally rolls it out. Thus it's a hardware problem.

ketzo 2y ago

You pose your comment like the answer is obvious, but I don’t know that it is. 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.

delabay 2y ago
It's "hardware is hard" versus "scaling laws". 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.
- ketzo 2y ago
  But the implicit assumption is that Tesla’s problems can be solved by scaling laws. That seems to me like a big assumption, no? 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.
  - delabay 2y ago
    Waymo (and my lived experience with FSD) is proof the end goal is obtainable. 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)