Hacker News Reader

Hullo, another Model 3 driver here.

I mentioned this in another comment, but it's relevant here; OP wasn't using Autosteer or Autopilot, they were using cruise control: a miles more simple system--and that system behaved in a way that seriously reduced confidence in it.

(For the record, I've experienced exactly the same behavior in my M3 as well.)

So, if we can't trust something as simple as cruise control, and we've seen similar confidence-eroding issues in the autopilot features Tesla has already deemed "good enough" for a live release, I think it is 100% reasonable to be skeptical that Tesla is anywhere close to "solving self-driving."

They have a history of overblown claims and faulty software in this realm--as much as I adore my car and the disruption Tesla has done to the industry, Tesla does not deserve the benefit of the doubt in this realm, especially not with safety-critical systems.

Cruise control is exactly that, simple. It errs on the side of safety and makes fail-safe mistakes.

FSD is a completely separate system with different tolerances. One does not have any relation to the other

The transition from low quality to high quality driving will be sublinear, not linear or greater than linear. This has been the case at every other self driving company. You get big gains from working on low hanging fruit early on in the process, then there's a long, arduous march to make improvements once those fruit are gone and avoid regressions. That's why Waymo, Cruise, Zoox, etc. Have been able to do impressive demos for the last decade, but it took Waymo a full decade to actually launch a public ride-sharing service.

Before you bring up Tesla's "Fleet Learning", it's not used at all outside of gathering perception training data for labelers. greentheonly got access to a dev build of the FSD beta - ML is just used to generate a representation of the current scene - all of the driving and planning logic is rules-based and hardcoded by engineers.

Seems unknowable still since Waymo and Cruise are now both deploying fully autonomous vehicles ahead of most peoples' expectations. If graphed, I'd imagine this implies a non-linear curve given these expectations were grounded in linearity imo. So it's under-determined basically.

They're deploying AVs way (1-2 years, at least) behind their own schedules, by their own admission. Ask anyone inside of the industry (including Musk) and they'll tell you that the problem is the long march of nines I mentioned earlier.

This is also why you saw a bunch of AV startups with < 50 employees that got big initial funding rounds but folded after a couple of years - it's pretty easy to grab a few dozen grad students, slap some LIDAR on a car, encode basic road rules, and get a demo working that's compelling to investors. Once you get past that initial pitch though, you get into the weeds of trying to progress without causing regressions elsewhere.

Also, what makes you think those expectations were grounded in linearity? Waymo went from an initial project at Google in early 2009 to driving 10 complex 100 mile routes without intervention in mid 2010 - on the surface, it would have seemed they were 95% of the way there. Improvement has been sublinear for a very long time and everyone in the AV industry is aware of it.

The linear assumption I'm referring to is forwards-looking not backwards-looking. Up until say last year I think the consensus was that a production-deployed FSD vehicle was ~5 years away. It shipped this year for both Waymo and Cruise. Unless you're arguing that the pessimists and optimists have just been arguing over the slope of the line, not the curvature (which may be fair.) If that surprise is any indication it leads me to think a similar consensus-breaking event may happen with Tesla, but time will tell when that is, since as you state they are behind in several ways.

Cruise and Waymo Gamble Lives with Unsupervised Self-Driving Cars

https://jperla.medium.com/tesla-saves-lives-waymo-gambles-pa...

Also a Tesla owner, on latest AP hardware: I would be amazed if the FSD update manages to solve all the problems, I have been through a few updates where things noticeably improved but we now have to pay Tesla money because while using autopark it just completely reversed itself (at a pretty high speed) into an object. It was insane. We’ve had weird AP issues too and now we don’t feel confident using any of those features.

I agree Tesla's shipped software that doesn't inspire confidence in the current version. I only use it on highways and remain fully engaged. (If I had to guess, in that scenario it does provide a marginal improvement in safety over fully manual driving.)

However I don't think lack of confidence in the current software, particularly the production autopilot software, is really giving me much signal in helping predict if Tesla will have FSD anytime soon or not. Now that FSD (at least in Phoenix conditions) is known to be possible the prior that it will never be achieved by Tesla to me is near zero, beyond the chance that LIDAR is strictly necessary and they have a doomed hardware stack. I think there is a pretty big range of possible timelines though - I could see it happening in as soon as 18 months or as late as 10 years depending on how poorly things sift out with their hardware stack.

Can you fill me in on where the levels of Tesla software are? The version of the OP who you are calling out as "apples" (the old, legacy one), what SAE autonomous level is that one? And the "oranges" one that is new and beta, is that a different level? If so, which level?

* Level 0 (No Driving Automation)

* Level 1 (Driver Assistance)

* Level 2 (Partial Driving Automation)

* Level 3 (Conditional Driving Automation)

* Level 4 (High Driving Automation)

* Level 5 (Full Driving Automation)

Different features of Tesla's software have different levels. As a whole, it's level 2.

If you restrict it to certain parts of driving, eg TACC, that might be level 4 (doesn't conform to the chart though). It can handle everything in it's domain, albeit in a rudimentary way, so a human doesn't always have to shadow it, but they do need to be ready to take over in some cases.

The apples to oranges bit is using the cruise control software to make guesses about FSD. I'm not certain but it would be very surprising if the software that provides basic cruise control shares much of the underlying algorithms and models as FSD, and if it does I'd be surprised if it's using models that aren't quite a bit older.

> It's not clear that the transition from low quality to high quality full self driving will be linear. Most claims around anecdotes come with an unspoken assumption we ought to see steady improvement over time vs a step-function like behavior. The latter seems somewhat possible given the nature of how improvements in AI systems seem to happen.

That is true, but it's actually an argument _against_ the assumption of full self-driving capability being just around the corner. Because that assumption itself is based solely on observed recent progress, which - as you correcly state - can be very misleading when used as a means to extrapolate the future due to the tendency of the AI field to have long stretches of very little to nearly no progress at all, divided by huge leaps followed by stretches of very fast progress in specific sub-areas of the field. We clearly have seen such a stretch recently, triggered by the deep learning idea and enabled by an explosion in parallel computing capability over the course of the last decade.

I'm not claiming it will be linear or will not be linear. I'm saying it's a bad idea to assume either when trying to predict the future.

The problem is: you have to assume "something", because otherwise you simply can't make any predictions. Either assume progress is linear (which seems to be the instinctive default most people use if they don't know much about a particular field) or assume it's non-linear. In case of the latter, you can choose between assuming exponential progress to apply or a step-wise, non-continuous kind of progress. It's a bad idea to not assume one of these to be the case if you want to make a prediction, because it's not possible to predict anything without that assumption.

I mean, you can of course refrain from making any predictions. But I have never met a single person who could resist the instinctive urge to extrapolate current developments into the future. It is as much a core part of being human to do this as it is to breathe air.

It's the assumption of linearity in combination with anecdotes that make the prediction weak, not the assumption itself. If you hold that assumption and use a variety of data to justify it and extrapolate from that's different than what I was criticizing.

"You weren't using the actual FSD software, but the live production autopilot, so it's Apples to Oranges."

Do you think Tesla has completely separate groups of engineers working on this? That doesn't seem like Tesla's style at all.

It's the same people but FSD was a rewrite of the system.

The rewrite was to do image recognition and inference not from individual still images (pre-FSD) but from video feed from all cameras. This was enabled by the FSD chip Tesla started putting in cars about a year ago.

Plus the improvements that come simply from gathering more data from cars and using it to improve the training of neural network.

It's documented in Musk's tweets and Karpathy's talks.

> Beyond all this, without understanding the underlying cause of the behavior you mention, it's hard to say from a risk perspective if the car made the right choice.

What argument could there be that suddenly slowing down on the freeway with no visible obstruction is the right choice?

It does not avoid any danger, but it creates a danger of getting rear-ended by cars behind you.

The argument would be: the car saw something you didn't.

I'm not saying that's what happened, just that it could have happened and therefore would be an explanation.

Your position here is, essentially, the same as a Q believer. Just trust in Elon, wait for the big reveal. Despite a complete lack of evidence in favor of the idea, and also despite an abundance of evidence to the contrary, FSD will emerge fully formed in the near future, without any intermediate signs of progress.

You have a reading comprehension problem. Saying "what you just said does not confer evidence to your claim" (which is what I said) is not the same thing as saying "I am 100% certain of the counterclaim." Responses like this one are why discourse on the Internet has fallen so far. The analogy to Q lands you squarely in the "lol fag" tier of responses.