Hacker News Reader

Comment by ximeng | Hacker News Reader

The latest version 12 FSD beta seems significantly better, to the extent that experience with previous versions is less relevant. There's no real reason to expect these systems to get worse, they'll only improve. Today's videos versus those of even just a year or two ago shows huge progress.

Not better than doing yourself today in general, but it will be in more and more situations. https://www.youtube.com/watch?v=fpoXr_z_6a4 is illustrative of the current level of ability. A good example where the self-driving compares favorably with poor human driving is at 10:15 in the video.

There are some pretty major gaps (for example, limited ability to reverse and read signs), but I wouldn't bet on them not being resolved in the next five years.

View on HN

I've tried FSD a handful of times during this month's beta and, while it's way better than when I tried it for a month a while back, it's at the point where I can very much see why the NTSB/etc are concerned about people trusting it too much.

It's generally really impressive and then suddenly very wrong. As in, it navigated a four-way stop with a in the crosswalk flawlessly. And then, later, it was lining up for a left turn into a parking lot and decided that the oncoming truck wasn't relevant, so it spun the steering wheel very quickly and lurched forward until I mashed the brake.

I'm very hopeful that we can replace the mistake-prone meatbags that drive cars now, but I'm incredibly concerned with people on other forums saying that they trust it completely.

It's clearly not at the point of reliability where it should be used without close supervision, but the progress is undeniable.

See 29:14 in the video I linked for an example of it behaving potentially dangerously, in what the video author describes as a regression in its behaviour and a similar situation to what you describe.

I also tried it out during this month's trial, and while it is convenient in certain scenarios, I wouldn't consider it production-ready.

If you treat it as pre-release and truly understand the risks, then it functions adequately. As a side note, I am pleased with the progress being made in this space, when I worked in that specific industry years back, it was nowhere near the functionality you have now.

Will they only improve? Or will they get better at most things while old previously handled edge cases sneak back in without warning? There are so many situations that it needs to handle 100% correctly.

Iirc Tesla has a giant internal validation sets of edge-case-related disengagements they test against using a simulator whenever they’re gonna deploy a new version.

That’s good to hear. It’s a huge problem space, and seems like it’d be rather difficult to test completely.

Reply on news.ycombinator.com