Comment by fnordpiglet | Hacker News Reader

Comment by fnordpiglet | Hacker News Reader

I’m been on the fsd beta for a long time now and I have to say it’s remarkably improved over the last year. If it continues to improve at this rate for another year it’s going to be pretty useful. I use it to drive around seattle (which is a difficult driving place with randomly followed infrastructure standards and randomly driving drivers and randomly parked cars) and it used to be a death defying gripping of the steering wheel and grabbing it out of bone headed moves while everyone looks at me askance to a relatively chill experience. Yesterday I needed to take it in for some fender work (my wife isn’t the best driver and I got fsd as a donation to self driving research in the off chance it’ll let her someday delegate driving to the car, but since that hasn’t happened yet I have periodic body work that needs to be done as she pachinkos through the streets). The Tesla service center is in an industrial area in an insane interchange part of I5 with crazy off ramps and strange turns. I always get lost. I let fsd drive door to door and it did it flawlessly. That surprised me, but after each update it’s been getting considerably better.

It still has flaws - it widens it’s place in the lane on every on-ramp which is annoying and weird road conditions can confuse it about what to do next. It sometimes doesn’t anticipate well the turn signals of someone changing lanes leading to people flipping me off for not giving them space. But all in all, given the state of adaptive cruise control 10 years ago, this is a remarkable achievement to put in a consumer grade vehicle.

That you're surprised that it managed drive door to door is telling. And not in the positive way you want to make it seem.

It's very different from the approach taken by Waymo and Cruise. Tesla's FSD uses vision only, and it's available everywhere in the US and Canada. The fact that it can complete most drives successfully (~90% of the time) without intervention in any city in the US or Canada is pretty impressive IMO.

The parent poster is right, I've been watching the progress and it's gotten probably 20 to 50 times more reliable in the last year.

This is what people aren't paying attention to. The rate of progress over only 1 year is astounding.

Driving with FSD supervised by a driver already makes driving far safer and far less stressful. The car has a situational awareness that humans can't match. It's only a matter of time until most drives have 0 disengagements due to human discomfort/awkwardness.

> Driving with FSD supervised by a driver already makes driving far safer

Does it though? Waymo said that their research indicated drivers tend to get complacent and assume the tech will bail them, thus they pay less attention and make more mistakes.

> The car has a situational awareness that humans can't match.

Weird that there are so many videos of Teslas randomly breaking or turning haphazardly then.

Also, the self driving is in the dangerous zone, where it's good enough so the driver will zone out, but bad enough that it's very dangerous to do so.

> ~90% of the time

This is the thing that always baffles me about self driving - while it may be very cool and impressive from a technology perspective, and while of course no technology can start out as completely bug free, when it comes to driving, anything less than 99.9999% success is a worse driving experience.

That is, until I can take a nap in my car, or read a book, or whatever, what's really the point? If I have to keep my hands on the wheel (or yoke) at all times, and have to pay attention 100% of the time, I'd rather be in a mode where I actually really do need to pay attention 100% of the time, instead of one where I really only need to pay attention 5% of the time (oh yeah, and if I miss that 5% I'm in an accident or dead).

In stop an go traffic in a city just having to be aware of whether it’s about to murder you is a blessing compared to screaming and shaking your fist at everyone on the road. There’s a lot of advantage to being able to zone out with highway driving attention vs active driving. I don’t need it to let me sleep, I just need it to do most of the thinking and I just monitor it’s performance - the cognitive load of that is very low vs active driving.

Other solutions are geofenced, not because they could not ostensibly function in other environments, but because it is criminally reckless to do so. Allowing a safety-critical system under test to operate in unvalidated situations is stupid. Allowing a safety-critical system in development that is currently unsafe to operate to be operated by untrained customers in unvalidated environments is criminal.

Waymo had a fully driverless test vehicle in 2015, before Tesla Autopilot was even released. That is a superior system in a specific operational domain. They continued improving it for years with multiple order of magnitude improvements before allowing customers to use it in the same, highly constrained operational domain. Only Tesla is reckless enough to look at a product as good as what Waymo had 7 years ago, decide to ship it, and expand the operational domain to untested situations while they are at it.

OP is talking about being surprised by steady progress to a major human achievement, one that could save a million lives for starters, but also one that in its mature state (where robots can visually interpret and navigate 3d space) unlocks even more labor saving automation than cars, busses, and trains put together.

It has the dark side of military uses, sure, but given how developed our tools for blind destruction already are, robots that can see the world will make a much bigger dent on the positive side.

Progress on this is genuinely exciting and people should let themselves feel that even if they don't like Elon Musk.

Thank you for your thoughtful comment. It is quite sad to see how eager people are to throw stones. It’s always the darkest side of HN that comes out in these threads.

The fact you’re not surprised that a car can drive itself is telling. And not in the positive way you want to make it seem.

It’s magic man. You’re jaded, and have set the goal posts too far and likely increasingly far.

WRT delivery timing, who amongst us had delivered a complex R&D heavy software project on time? I’m not an Elon fan boy but I respect he laid out a vision and worked to execute on it. That he didn’t hit the mark with 100% fidelity doesn’t seem like an issue to me.

> You’re jaded

No, the question is whether Musk committed fraud by selling _in 2016_ self-driving car software which he is still unable to deliver in 2022, and the answer is YES whether or not you are "jaded".

You are a lawyer? Or is that just your opinion based on what you personally believe fraud should mean (or on how much or little you like Musk)? Thanks for clarifying.

Why would you be surprised? Google did that with no safety drivers over 7 years ago [1], before Tesla even released Autopilot, after 6 years of development. The gap between that and a proper autonomous vehicle with no geofencing that is safer than even the bottom 10% of human drivers is immense and has not been achieved yet. That Tesla has achieved less in 6 years than what Waymo did in 6 years when they were trailblazing the field over a decade ago is disappointing. That Tesla executives believe their Full Self Driving product is acceptable to sell and ship to untrained customers, when systems that are nearly a decade more advanced and tens to hundreds of time better are unacceptable, is criminal.

[1] https://medium.com/waymo/scenes-from-the-street-5bb77046d7ce

> You’re jaded, and have set the goal posts too far and likely increasingly far.

Not the original poster. Tesla has set the goal posts too far, not commenters here. And yet, they are misleading consumers about how far the goal posts are.

People aren't saying Tesla's current features aren't marvelous, but that its not the magic that was promised.

> WRT delivery timing, who amongst us had delivered a complex R&D heavy software project on time.

Probably none of us. However, most of us haven't overpromised on the delivery time when we know that the project will take longer than what we have conveyed.

I'm sorry, but your wife should lose her license if she is truly that bad, and she should be required to take further training to get it back.

I wonder what her insurance rate is, and why it’s not so astronomically high that it causes her to reconsider her driving habits.

Right, like driving isn't a single player experience. You are also a danger to everyone else...

Well, she’s never actually hit anyone. Just walls when she’s backing out of a spot.

Seattle's road infrastructure is bafflingly bad and confusing for drivers. But it always seems wrong to me to assume that we shouldn't improve the infrastructure and make it more safe vs putting half baked self driving out in the real world to overcome it. I understand they want more data to feed into their ML datasets and the idea is that that will improve things over time, but we are sacrificing some number of lives for this data without knowing if a convergence point will actually be reached. And what you say at the end has lead me to actively avoid Teslas when I can, they make the road more unsafe for me while they are learning, not just their own drivers.

I don’t think they’re mutually exclusive.

I was in one a couple months ago. My friend likes it but was also clear that you have to pay attention to what it's doing because it makes regular mistakes.

Yep. After the last few updates, I went from a non-believer (you need LIDAR, vision only autonomy is at least a decade away and tesla fucked up going this route) to actually believing in 2 or 3 years we'll have a very nice and useful level 3 from Tesla, and that vision only is the right approach (because I want this to be affordable, not some tech only robotaxis have).

And I think by 2027 we'll have level 4 vision only.

It's unlikely that the rate of improvement, however you estimate it, will continue at the same rate. This is a Pareto effect, with marginal gains ever more difficult to obtain.

And the edge cases FSD can handle will be impaired by the sensors they do (not) install. While they make a good point that sensor fusion can impair comprehension, the removal of radar makes it hard to avoid the edge cases most calamitous.

I actually expect a nice boost of progress as the occupancy network is improved and used more by other parts of the software stack.

And I expect a fairly significant improvement when they release the HW4 hardware upgrade. Better cameras, and more processing are such "free" (as in effort, not money of course lol) improvements to their error rates. I'm surprised they got this far with their current hardware. If you've ever done real time computer vision, you know how hard it is to maintain decent performance (fps and latency) even on fairly beefy GPUs.

>I got fsd as a donation to self driving research in the off chance .

That's likely the reason majority bought FSD

Reply on news.ycombinator.com