Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

> Tesla FSD is practically perfected

Absolutely untrue. It's not even close to perfect. I use it a few times a week and while it's seen some significant improvements over the last year, it still makes pretty dangerous mistakes, especially on left turns, right-turn slip lanes, intersections with flashing yellow lights, streets with very worn road lines or where the positioning of the road lines shift from one side of the intersection to the other. I can think of a dozen other similar situations so I'm very sure there are many others I've never even encountered.

It remains to be seen how much juice Tesla can squeeze out of a transformer approach to autonomous driving, but it's by no means a sure thing.

edit: I misunderstood the comment. I see now that "practically perfected" is a reference to the ideal conditions mentioned in your first sentence.



You seem to have (accidentally?) left off the rest of the comment:

"Tesla FSD is practically perfected, with no accidents whatsoever, under such conditions"


You're right. Thanks. I edited my comment to acknowledge that.


Thank you to post your personal experience. One pattern that I have noticed about Tesla FSD stories on HN: it is black and white, with little grey area. Either people live an area where the roads are easy for FSD to navigate, and they come here to say "It is perfect!". And vice versa: People live in a place with a bunch of complex roads and intersections where FSD does not perform well... And they come to HN to share their experience. (Privately, I cannot wait to see self-driving models try driving on Jakarta or Napoli. It will so much fun to watch those on YouTube!) I do think that Elon/Tesla is taking a crazy gamble to release FSD early to gain billions of hours of training data! I can understand where this would make some people uncomfortable, due to the safety concerns. Except Waymo, who else has the training data that Tesla has at this point? Wiki tells me that FSD has been driving on public roads since 2016. That must an astonishing amount of training data accumulated in the last 8 years! I assume he will be tasking X.ai with improving the FSD model using this enormous training data.


The streets of Jakarta and Napoli wouldn't work with the Mercedes Benz system either.

> Who has the training data at this point?

Woven by Toyota has roof pods on rideshare (Lyft) cars collecting data: https://woven.toyota/en/our-latest/20220407/




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: