> If you have been following their updates, it looks like that they have finally found the correct approach for FSD and it is improving very quickly. At this rate, it seems that it’ll be close to level 5 this year or next year.

Yeah, their perpetual 2-year estimate dropped to 1 year around a year and half ago, after being at 2 years for at least 6 years. So we’re probably four and half years from it either being ready…or dropping to 6 months off for the next several years.

I have the beta (finally). It’s very very rough.

So would you say your experience is very different from the one shown in this video[1], or is it that you'd describe the video as very very rough?

[1] https://www.youtube.com/watch?v=fwduh2kRj3M

I have the beta as well, it's very very rough.

Like, trying to change lanes into a "flush median" (aka yellow stripes across) to turn onto a one way in the wrong direction. Or randomly swerving back and forth (hard) when a lane splits into 2. Or trying to take 90deg turns at 45mph.

There are times it works amazingly, I've had it slow down and swerve out of the way of someone barreling out of a parking lot without stopping. It's also great at finding a gap to fit into between cars to turn, but it also got itself into that situation by not getting over sooner and instead trying to merge ~500ft before the turn.

Also for some reason on a 2 lane road (in each direction) it would constantly try to be in the "passing" lane.

> Also for some reason on a 2 lane road (in each direction) it would constantly try to be in the "passing" lane.

The problem with collecting real-world training data is that there are going to be a lot of assholes in your data.

It has algorithms and heuristics that dictate how it drives, so I think this is more the programmers haven’t gotten to it yet.