5 ms·

Why would they want to abandon FSD? I look up videos whenever these threads come out and over the last few years it seems to have made astonishing progress. 2

Why would they want to abandon FSD? I look up videos whenever these threads come out and over the last few years it seems to have made astonishing progress. 2 million miles per day of FSD driving is impressive. The resulting training data from every corner of the USA is a bonkers advantage.

NovemberWhiskey 3y ago
Only if you think this is actually a model training problem.
- jdewerd 3y ago
  Their existing system is pretty strong evidence that it was in fact a model training problem -- do you mean to say that you think training will stall out before it gets to the point of being viable?
  - cmeacham98 3y ago
    I'll admit that I am by no means an expert and might be completely wrong, but yes, I believe that model training cannot fully solve driverless vehicles. All models that we train today have edge cases where the model produces wildly inaccurate output. As far as I'm aware, nobody yet has managed to squeeze this problem out. ChatGPT is still an overall useful tool with the occasional hallucination. Those types of edge cases in software driving a car are unacceptable.
    - misiti3780 3y ago
      I agree, but i think it's getting close to be much safer than humans driving on highways.
    - jankyxenon 3y ago
      Right, said another way, is driving a completely deterministic activity or does it sometimes require general intelligence. If it's not completely deterministic, then you might need a human in the loop.
    - jdewerd 3y ago
      ChatGPT intentionally plows through low-confidence scenarios. It just tries to do its best because "I dunno" is reliably low-utility while a hallucination doesn't gamble lives. It errs on the side of hallucination by design: if you alter the RLHF feedback, you can alter the confidence level. In FSD, as you point out, the stakes are higher. It makes a different design decision: it slows down in low-confidence situations. Just like a person. If you watch a FSD video, you'll see that most of the human interventions happen because FSD lost confidence in a tricky scenario and halted the car even though the pathing ultimately wound up being correct. This won't solve every edge case, but neither do humans -- the bar is (or should be) "better than human," not "perfect," otherwise we are sacrificing lives on principle. I have no idea if training will stall out half way through the march of 9s. It might, but if it does, it won't be because there is no high level strategy to control hallucinations. That already exists.
      - cmeacham98 3y ago
        Tesla sold their investors and customers on "full self driving", not "self driving where you must always have your hands on the wheel ready to take over". Customers are willing to tolerate the latter right now because they believe the former is still in development.
        jdewerd 3y ago
        Do you think it isn't? I'm not here to argue that the overly enthusiastic timetables weren't cringe. They definitely were. I'm just here because there seems to be consensus that FSD is doomed and it doesn't remotely seem that way to me.
        cmeacham98 3y ago
        Of course it's in development, but if I'm right that ML cannot solve this problem then a lot of that development is wasted. My point is that in the long run, Tesla cannot solve the ML edge cases by having a human available to take over at a split second notice.
  - minwcnt5 3y ago
    They still haven't removed the driver, which Waymo did in 2018 when AI was much less advanced than today. Therefore it's pretty clear there's some fundamental limitation that probably has to do with hardware and/or inability to build proper guardrails for their AI, as opposed to the capability of the models. In other words: building AI models that can drive impressively in the common case is likely the EASY part of the problem, and they haven't made much progress on the hard part, which is likely dealing with the long tail.
    - jdewerd 3y ago
      Should you remove the human and then expand scope (Waymo)? Or expand the scope and then remove the human (Tesla)? I dunno. I personally think the Tesla approach is fundamentally better because it obtains more data, faster. Millions of miles per day rather than millions of miles per year, with better macrodiversity to boot. However, that's just a heuristic: we've got a Tortoise and Hare scenario and as the parable teaches us, it's not so difficult to imagine a scenario where slow and steady wins the race. Even though I favor the Tesla approach, I could see this going either way. That said, I do think it's uncharitable to read failure into Tesla's high level decisions. They could have chosen the Waymo path and dumped all their effort into maximizing reliability on a small scenario subset. They chose not to. There are Andrej Karpathy interviews about it. They wanted to data-max. They data-maxed. Now they have 300x the data collection rate of Waymo. That's not failure, and it may even be a killer advantage. We'll see.
      - ra7 3y ago
        > They wanted to data-max. They data-maxed. Now they have 300x the data collection rate of Waymo. Citation needed. What Tesla lacks is quality data, which Waymo has due to their high fidelity sensors. Tesla has been "collecting data" for years without much to show for. They still have embarrassing failures so basic in nature [1], you know they're going nowhere. If all it took was data, they would have something meaningful by now. [1] https://twitter.com/elaifresh/status/1727889381309239731 https://twitter.com/elaifresh/status/1727889381309239731
        3y ago

NovemberWhiskey 3y ago

Only if you think this is actually a model training problem.

jdewerd 3y ago
Their existing system is pretty strong evidence that it was in fact a model training problem -- do you mean to say that you think training will stall out before it gets to the point of being viable?
- cmeacham98 3y ago
  I'll admit that I am by no means an expert and might be completely wrong, but yes, I believe that model training cannot fully solve driverless vehicles. All models that we train today have edge cases where the model produces wildly inaccurate output. As far as I'm aware, nobody yet has managed to squeeze this problem out. ChatGPT is still an overall useful tool with the occasional hallucination. Those types of edge cases in software driving a car are unacceptable.
  - misiti3780 3y ago
    I agree, but i think it's getting close to be much safer than humans driving on highways.
  - jankyxenon 3y ago
    Right, said another way, is driving a completely deterministic activity or does it sometimes require general intelligence. If it's not completely deterministic, then you might need a human in the loop.
  - jdewerd 3y ago
    ChatGPT intentionally plows through low-confidence scenarios. It just tries to do its best because "I dunno" is reliably low-utility while a hallucination doesn't gamble lives. It errs on the side of hallucination by design: if you alter the RLHF feedback, you can alter the confidence level. In FSD, as you point out, the stakes are higher. It makes a different design decision: it slows down in low-confidence situations. Just like a person. If you watch a FSD video, you'll see that most of the human interventions happen because FSD lost confidence in a tricky scenario and halted the car even though the pathing ultimately wound up being correct. This won't solve every edge case, but neither do humans -- the bar is (or should be) "better than human," not "perfect," otherwise we are sacrificing lives on principle. I have no idea if training will stall out half way through the march of 9s. It might, but if it does, it won't be because there is no high level strategy to control hallucinations. That already exists.
    - cmeacham98 3y ago
      Tesla sold their investors and customers on "full self driving", not "self driving where you must always have your hands on the wheel ready to take over". Customers are willing to tolerate the latter right now because they believe the former is still in development.
      - jdewerd 3y ago
        Do you think it isn't? I'm not here to argue that the overly enthusiastic timetables weren't cringe. They definitely were. I'm just here because there seems to be consensus that FSD is doomed and it doesn't remotely seem that way to me.
        cmeacham98 3y ago
        Of course it's in development, but if I'm right that ML cannot solve this problem then a lot of that development is wasted. My point is that in the long run, Tesla cannot solve the ML edge cases by having a human available to take over at a split second notice.
- minwcnt5 3y ago
  They still haven't removed the driver, which Waymo did in 2018 when AI was much less advanced than today. Therefore it's pretty clear there's some fundamental limitation that probably has to do with hardware and/or inability to build proper guardrails for their AI, as opposed to the capability of the models. In other words: building AI models that can drive impressively in the common case is likely the EASY part of the problem, and they haven't made much progress on the hard part, which is likely dealing with the long tail.
  - jdewerd 3y ago
    Should you remove the human and then expand scope (Waymo)? Or expand the scope and then remove the human (Tesla)? I dunno. I personally think the Tesla approach is fundamentally better because it obtains more data, faster. Millions of miles per day rather than millions of miles per year, with better macrodiversity to boot. However, that's just a heuristic: we've got a Tortoise and Hare scenario and as the parable teaches us, it's not so difficult to imagine a scenario where slow and steady wins the race. Even though I favor the Tesla approach, I could see this going either way. That said, I do think it's uncharitable to read failure into Tesla's high level decisions. They could have chosen the Waymo path and dumped all their effort into maximizing reliability on a small scenario subset. They chose not to. There are Andrej Karpathy interviews about it. They wanted to data-max. They data-maxed. Now they have 300x the data collection rate of Waymo. That's not failure, and it may even be a killer advantage. We'll see.
    - ra7 3y ago
      > They wanted to data-max. They data-maxed. Now they have 300x the data collection rate of Waymo. Citation needed. What Tesla lacks is quality data, which Waymo has due to their high fidelity sensors. Tesla has been "collecting data" for years without much to show for. They still have embarrassing failures so basic in nature [1], you know they're going nowhere. If all it took was data, they would have something meaningful by now. [1] https://twitter.com/elaifresh/status/1727889381309239731 https://twitter.com/elaifresh/status/1727889381309239731
      - 3y ago