I'm renting a Tesla Model Y with HW4 and FSD version 14.2.2.5 to see for myself what it's like and how plausible it is that vision-only level 4 self-driving is imminent.
If I disengage the self-driving but it's ambiguous if I actually needed to, I'll resolve-to-PROB appropriately. In other words, we're estimating the probability of a would-be crash of any severity over 2000 miles of self-driving, almost all highway miles.
1,000People are also trading
I'm over a thousand miles in and ... I have so many opinions I don't know where to begin. But one of them is that it's absolutely uncanny how human-like it is. And zero safety-related interventions so far!
@dreev Now at 1243 miles, doing some in-town miles for a bit, before road trip home. Today it got itself confused in a parking lot, but the good kind of confused, where it just froze not knowing what to do.
Should I subsidize this market? It's wild the depth of ignorance we seem to be at for how close Tesla actually is on vision-only self-driving. I think there will be a lot of chances for me to find out definitively that the answer is "nowhere close" on this road trip. But if it's flawless, I fear that that will leave me almost as clueless as ever. Because 2000 flawless miles in a single experiment is compatible with a very wide range of theories for how reliable the self-driving actually is.
But if this market probability (currently near 50%) is correct, then I guess an intervention-free trip actually tells us a lot!
This kind of feels like a Necker cube. On the one hand, a decent chance of going thousands of miles without a single intervention would mean Tesla FSD is insanely impressive and seemingly on the brink of level 4 autonomy. On the other hand, each additional 9 of reliability (99%, 99.9%, etc) could be a years-long engineering project. And if 2000 miles is touch-and-go then we need a few more 9s before it stops needing human supervision.
(It actually reminds me of progress towards AGI. LLMs and diffusion models keep doing more and more wildly impressive things -- things we were sure would require AGI five years ago. It makes the remaining distance to AGI feel like it must be short. But dotting/crossing those last i's and t's before all white-collar jobs are automatable could be multiple years-long engineering projects. Or not, we just really don't know. See Toby Ord's new essay today on the deep, deep depths of our not-knowing.)
@dreev 16 holders is not a big sample, nor any kind of valid sample. This is basically a random group of people interested in the subject, with unknown expertise on the subject. Are they Tesla owners? Have they experienced FSD? This feels like measuring public perception rather than actual knowledge.
@MarkosGiannopoulos Yeah, that's the argument to subsidize it, to attract more knowledgable traders. Do you have a strong opinion yourself about the correct probability?
From all of Elon's commentary about upcoming FSD versions I'm fairly convinced that the robotaxis aren't significantly different in terms of software and hardware. Seems the biggest difference is automatic camera washers.
Which means that if the true probability for this market is more than like 1% I don't see how that's compatible with a YES in "did Tesla launch level 4 robotaxis last summer?" market.
More thoughts over on AGI Friday.
PS: The car is showing up in half an hour! 🤞
@dreev Congrats on finding an excuse for driving a Tesla! "Look honey, this is a research project" :D
On a more serious note, given that a lof of other people https://fsddb.com/leaderboard#highest-fsd-percentage have done multiple of 2000 miles without issues, you are most likely to not have any issue.
@MarkosGiannopoulos I agree. It's looking quite good for NO in this market. What's your sense for how many miles I'd need to bump this up to before the probability here would be closer to 50%?
@dreev Maybe 10.000? But this also depends on your routes. Per https://teslafsdtracker.com/
city miles are harder for FSD.
