-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Navigate on Autopilot is Useless (2018.42.3)
- Thread starter wk057
- Start date
boonedocks
MS LR Blk/Blk 19”
So the 10,000 earlier beta testers were not "normal"? And now 100,000 testers, what are they? FSD is far from perfect (that's why its beta), but the "doom and gloom" predictions have all turned out to be wrong. First we were told the car would not be able to see traffic signals (wrong), then it would crash within a few days of the start of beta (wrong), then it would not be able to see far enough to make safe turns (wrong), then it would run over every child it could find (wrong).
Sure, there are people out there who gush over FSD as if it were perfect, but for every one of those there are "True DISbelievers" who are equally irrational in their criticism. Neither contributes anything even remotely useful imho.
"True Disbelievers" hope to spread the fear of god about using it, Ultimate Fans say you can sit with no hands on the wheel and play with your phone.
You decide which approach appeals to you. Hint, Tesla recommends the first one: "FSD Beta can do the wrong thing at the wrong time".
No, overwhelmingly they were employees.
How many disabled it? I'm betting a lot of people don't use it except after a release to have a good laugh, and I bet a decent proportion have disabled it completely and gone back to standard firmware. Obviously the only numbers we have is Tesla's picked cherries, sort of like their "autopilot is safer" numbers that are easily picked apart.
Have they? My prediction since day one has been absolutely rock solid. Right down to their behavior around "smart" summon and FSD "beta" testing. Meanwhile, the predictions of the CEO have been absolute fantasy since 2015. I'd suggest that people that don't understand ML are the worst at prediction, which is likely what we're both seeing.
I didn't see anybody saying that the car couldn't see a red light, since the filters on the camera favor red. What we do see is the car running red lights, stop lights, and other traffic control devices. Meaning if you expect it to stop every time, you'll eventually be sent sailing into an intersection. Which is unsafe no matter how you try to slice it.
It absolutely did that. Then the fans said "no major crashes" which it promptly did, then they said no pedestrian crashes, which it promptly did. So I don't really know what you're talking about here, because it did all of those things. You can even see them in NHTSA reports.
That's an absolute fact. Crossing traffic moving at 50 MPH is a common thing in the US, and those cameras do not see nearly far enough. Even Chuck Cook's videos demonstrate that.
Like all those fan videos trying to prove dan wrong, and then their car hit the simulated child? You've used four examples that the car actually does. That's not a great position to take.
One huge difference between Tesla’s cameras and human vision is humans have the ability to focus our vision and brainpower in the highest threat direction. I wonder if Tesla will begin trying to focus its processing power in different areas depending on driving conditions.
I think it's more correct to say humans can only focus in one direction. Sure, we are good ar figuring things out in that direction, but only that direction. A car wont be as good as a human in that regard, but it can look in all directions at once. Look at how many accidents are caused by missing an entire car in a blind spot.
Well, several of the "True Disbelievers" here pretty clearly want FSD to fail, presumably because that would please their egos "I was righ!!t!". Tesla, and others in here (including me) urge everyone to be realistic and use caution while driving the car, not expect miracles, and ignore Elons predications about dates. Like most things in life, the extremes are to be avoided.
Yes, compared to an average human I would say Teslas have much better peripheral vision but worse focal vision.
Well you should probably buy a car running Dan's software then, I'm sure it will be wonderful.
I mean, just look at the first post date of this thread, and compare it to today's date. Consider that the system still can't handle detecting high beams or rain drops properly. We aren't talking about other companies, we're talking about Tesla, and the desire for Tesla to do better.
And your point? That Elon is a joke when it comes to dates? Old news, we all know that, I've said so many times. Nothing to see here.
That FSD is late? And what could "late" mean, aside from Elon's made-up dates? Later than expected? By whom? Based on what? No-one has ever built something like FSD before. EVER. So how can anyone know how long or how complex such a project could be? Probably the whole FSD team was naive about the dates too. So what? To an extent, you HAVE to be a bit naive to take on such a project.
You keep pointing out that FSD cannot do X or Y, as it it was some titanic discovery. Yes, we all already know that. So do Tesla. That's why its in beta. What you dont seem to notice (or perhaps dont want to) is that with each release it moves closer to achieving their goal. The car is VASTLY better at driving than a year ago, and continues to improve. So yes, I HAVE looked at the post date, AND todays' date. And I see huge progress. Does that make me an optimist? Perhaps, but you seem to think anyone seeing any advance at all is an optimist. You keep looking for flaws, as if to prove a point. But to my mind, the only point you are proving is that you seem to want FSD to fail.
Then why does the CEO continue to call it "superhuman" as recently as this week?
That's not what Beta means. Beta means major development is complete, Alpha testing has completed, and now a small base of public users is putting the product through its paces. That's not at all what FSD is doing. It's still under heavy development, making it a proof of concept rather than any kind of release quality.
It hasn't moved an inch in improving features or behavior. How is it that Tesla needed to spend three months on a single user's turn experience when he's been pressing the feedback button for two years straight? And when the declare mission accomplished, it still fails frequently. That's not closer to achieving a goal, that's sporadic changes at best.
Not for me. Still random braking with oncoming traffic, still not using blinkers to enter turning lanes, still not getting into turning lanes properly, still attempting to make illegal passes, still random braking as I pass through intersections, I could go on. But all of these behaviors were present last year, they're still present today. Auto wipers still don't work reliably, auto high beams are still not working. I don't doubt that Tesla is spending a lot of electricity adding more annotated training data to their set, but we're so deep into diminishing returns territory that it's not making a difference.
Nope. IMO it already has, but I want it to improve because there's hundreds of thousands of customers that have paid for it and they deserve something for their money.
I've already addressed the issue of Elons optimism. Nothing more needs to be said.
Sure, 30+ years ago that's what it meant, but its meaning has drifted significantly over the years. Gmail was beta for many years, Apple have iOS and macOS betas that are FAR from feature complete and are available to a wide audience. And anyway, so what? You're just quibbling over a name (seriously?). The reality, of course, is that in the non-deterministic world of NNs, sustained and widespread training and testing with real world data is the only way to build a product, which makes the whole concept of beta somewhat outdated.
Nonsense. it's improved vastly in the last 15+ months I've been testing it, and pretty much everyone else who has been doing so for long enough has said the same. The overall driving experience is smoother, with the car able to negotiate situations it would simply have failed at utterly in (say) 10.2 or 10.3. You are free to deny this, of course, but you should remember that denial is not the same as considered criticism.
Anyway, as I for one remember Churchill's definition of a fanatic, I think I've said all that I wish or need to say on this subject.
Last edited:
I keep seeing certain people mention this. Did your life depend on gmail? no.
...an audience of developers. You are not an FSD developer, nor do you have access to the telemetry necessary to verify its performance.
Well, lives are at risk with the general public, as we've seen countless times. That deserves consideration for most people. I suppose you believe different though.
No, you brought up "beta". I'm telling you what Beta means, you're the one mentioning it as some sort of defense for extremely low performance software being on public roads with members of the public that didn't consent to being test subjects.
Nonsense.
Certainly the Tesla Marketing Team. But basically nobody else.
A meaningless metric we keep hearing from certain people. Now it "smoothly" jerks the wheel toward oncoming traffic, and "smoothly" accelerates toward pedestrians, and "smoothly" steers rapidly when in intersections. Meanwhile, functionally it still doesn't path plan properly, it still doesn't handle countless common scenarios, and it still doesn't interpret or obey countless control signs like "road closed". But at least it'll be smooth when it crashes into them.
...like jerking the wheel toward oncoming traffic? No, wait, it did that back then too. You'd assume that if the "prime directive" is "don't crash" that it wouldn't attempt to head-on with another vehicle. But I guess that's just "beta" for you, right? That's what you'd tell the victims? Relax, sure, your legs are putty, but it was a beta.
Guys, you're never going to resolve this argument. @drtimhill and @DrDabbles have different experiences with their cars. It would be like Tim having a problem with this local McDonalds (poor service, bad food, unclean restrooms, etc) and Dabbles having a great experience with this local McDonalds. You're not going to convince each other that McDonalds is a great or poor place to eat. One of you can show videos of great McDonalds, and the other can show videos of crappy McDonalds restaurants.
This may just be an "agree to disagree" moment.
This may just be an "agree to disagree" moment.
True...true.The difference is these cars are supposed to behave the same across the fleet since they are running the same software.Unclean restrooms at McDonalds is understandable though. lol
I would still love to check a theory about different cars. I would love to take two models built at the same time and test them on the same road (following each other) and see the performance. They will obviously not be identical as the NN's will have some variation in their performance, but I'd expect it to be close enough that it would be hard to tell.
There are so many variables with different cars built at different times, cleanliness of the sensors, calibration of the cameras, wear and tear of the tires (there is someone here who replaced the stock tires with different diameter tires, but the system doesn't know about it, so the speed is not accurately displayed, and that may affect ADAS functions), etc. However, there should not be wildly different outcomes based on these things, more subtle differences. Though, a corrupt calibration would cause significant problems with ADAS.
Similar threads
