Tesla FSD confirmed to fail at stop sign as attempts to discredit tests backfire | Hacker News Reader

Tesla FSD confirmed to fail at stop sign as attempts to discredit tests backfire

twitter.com

46 pointsVeserv23 comments

As someone who has been driving FSDbeta. for the last 2 years or so, I can tell you there are a lot of spots that the car can't navigate without intervention.

I can also tell you that number is decreasing over time. In fact just yesterday for the first time in two years the car successfully made the left turn at the 6th and 20th Queensborough bridge onramp in Vancouver. High fives all around.

FSD has recently overcome a local maximum in performance and is enjoying rapid progress for the last few months after almost a year of minimal gains. It seems like the devs have already seen the next crest in development because version 12 appears to be another top to bottom rewrite of the NN architecture according to this recent talk. https://youtu.be/OKDRsVXv49A

I'm not surprised in the least that it blew the stop sign, but I would be very surprised if this error persisted over time.

It's interesting how much backlash something that was already well understood generated.

* FSD Beta is a level 2 system. This means it requires constant supervision. This crystal clear to every FSD Beta driver, the car will both monitor your face, and nag you to touch the wheel every so often to make sure you're paying attention. Too many "strikes" and it will lock you out of FSD Beta for a few weeks. It's basically expected that the software will fuck up every now and then, otherwise it would be level 3/4/5 (i.e require no supervision)

* If you watch the channels that review FSD Beta (Chuck Cook, Dirty Tesla, etc), the vast majority of drives require at least one disengagement. Only very recently with recent versions you would sometimes get the mythical 0 disengagement drive. So FSD beta requiring frequent disengagements is something already well known and understood.

It’s interesting how people are upset about a system marketed, sold, and operated on public roads as Full Self Driving is in fact nowhere close to fully self driving.

Very curious indeed!

I just can't imagine why people are upset that Tesla is gambling with their lives to juice their stock price by turning a half-baked self-driving implementation loose on public roads. Truly baffling!

It's the only company that advertise it as "auto pilot". Meanwhile all the other major players have the same tech with none of the backlash

Interesting

I think if Tesla called it “Enhanced Autopilot” people would be less mad but at the moment their “Full self driving” software does not fully self drive.

Releasing a beta product to operate a product that weighs a ton, and quickly accelerates to 60mph in seconds, while not strictly evaluating the end user before granting access, and then marketing it as "Full self driving", while it even fails to do basic, "slow down at curves", is something which makes them deserved for the hate.

For background:

https://twitter.com/RealDanODowd/status/1672325190112641024

This event occurred during a short 1 hour drive with Dan O’Dowd, FSD critic, and Ross Gerber, FSD proponent. Ross Gerber driving his Tesla on FSD 11.4.4, the most recent released version.

The Tesla community had a immediate backlash claiming that Ross, a major Tesla investor, deliberately sabotaged FSD during the intersection to make Tesla look bad. Alternatively, and simultaneously, they claimed that the car would have stopped if only Ross Gerber would have waited and his intervention was unnecessary and detrimental to Tesla FSD marketing.

TeslaBoomerMama, a FSD proponent, who wishes Tesla legal would do things in these sorts of situations to prevent the spread of disinformation [1], organized a effort on Sunday to test the intersection to prove that Dan was engaging in disinformation by claiming that the car was engaging in a dangerous behavior and would have run the stop sign and potentially crashed into the oncoming car.

The testing conclusion is that it is “slow to stop when crossing straight” and “this intersection has a problem… the system does not detect the stop sign in a timely enough manner and you are entering at about 60 kph or about 35 mph, so the problem is real”. In other words, it runs the stop sign as originally claimed by Dan. However, they go to great lengths to try to state that as obliquely as possible while still claiming that Dan is dishonest and mischaracterizing the nature of the event.

[1] https://twitter.com/TeslaBoomerMama/status/16730354224050995...

Hold on: the title says that FSD is ignoring the stop sign and your quote says that FSD doesn’t detect the stop sign in a timely enough manner.

In my mind, those are two different things. Both bad (but for different reasons).

I was echoing the naming on a earlier thread talking about the first event I linked: https://news.ycombinator.com/item?id=36453633

You are technically correct. I seem to be just in the edit window so I have updated the title to reflect that.

Let the record know that the title originally said: “ignore stop sign” instead of “fail at stop sign”.

I watched the video just now.

That stop sign does sneak up on you, which is why they put the very visible sign saying that there is a stop sign ahead!

FSD obviously needs to recognize the “stop sign ahead” sign and start to slow or do some scanning or whatever to look for a hard-to-see sign.

To be honest, this seems like it would be somewhat of a challenge for ML

Even if the stop sign does sneak up on you, the fact that the brakes didn't come on as soon it came into view is concerning. I feel like that would have ended very badly if the driver hadn't manually applied the brakes.

It should work if they train it on non-FSD driving at this intersection, most humans would slow down. I wonder why it doesn't, perhaps they don't have the right set of input and output features.

> The Tesla community had a immediate backlash claiming that Ross, a major Tesla investor, deliberately sabotaged FSD during the intersection to make Tesla look bad. Alternatively, and simultaneously, they claimed that the car would have stopped if only Ross Gerber would have waited and his intervention was unnecessary and detrimental to Tesla FSD marketing.

Wait.... so the driver was being attentive and saw a situation where the FSD may not react properly, and took manual of the control of the vehicle.

And now a said driver, a Tesla investor and and proponent, is getting backlash for doing the the very thing that the Tesla community has been insisting is necessary when driving beta FSD, which is to pay attention and be ready to take manual control of the vehicle...? And they're saying it's not a glitch but sabotage?

.......... Seriously?

Yes, the conclusions are:

- _intersection_ has a problem

- speed limit is too high, its City fault!!1

- You're Holding It Wrong! car works "flawlessly" if you manually limit speed before crossing this one intersection.

Aka conclusion is cult good, critics bad.

Only really naive and credulous people fall for Tesla's tricks these days. Rest of us knew from a very long time that all their claims are shady marketing tricks and there are better cars on the market.

"The guy is not clever enough to have found this intersection"

What kind of non-sense talk is this

It sounds like someone took for granted that all problems with a system would be non-reproducible and thus deniable, and is shocked that isn't the case.

As Tyler Durden said "You have to consider the possibility that God does not like you".

FSD doesnt work, its basically partially smart assistant, you still need to be in charge. Even tesla says FSD isnt perfect and makes you keep your hands on the steering wheel.

I dont get all the hate or love for tesla and FSD, its not a perfect or real FSD and its not a scam either. Its just software thats a work in progress.

I think the problem here is that Tesla sells it like it’s a prod version.

Yeah when you enable it you get all the warnings etc. but still you pay an important sum of money for something not properly working.

I personally admire Tesla sales & marketing team for financing the development through this “trick”, but they’re working on it for ages now and still we experience this kind of “regressions”.

Well the problem is Dan O’Dowd has history of faking issues with FSD but when it happened for real ofcourse people are ready to discredit. It is a real issue hope Tesla addresses this soon.

There is no such history. The only claim with any backing was Fred Lambert of electrek reporting on the first ad claiming that FSD was not enabled during the testing as FSD was not enabled on some footage in the ad.

https://electrek.co/2022/08/10/tesla-self-driving-smear-camp...

However, Fred made this claim before reviewing the raw footage which can be seen here:

https://dawnproject.com/the-dawn-projects-new-advertising-ca...

After seeing the raw footage he added a teeny update at the bottom of the first link acknowledging that FSD was actually enabled during the safety tests in the report as originally claimed. Some footage in the ad most likely just came from dry runs where FSD was not enabled.

Since then, the same group of people including individuals such as WholeMarsBlog and TeslaBoomerMama have continuously made false accusations of impropriety citing as evidence that the first ad was falsified even though it was actually the claims of falsification that were discredited. They are deliberately promulgating disinformation to support their stock portfolios.

Since then the proponents have claimed that the accelerator was pressed, so the critics increased the camera angle to include the accelerator. Then the proponents claimed the steering wheel was turned to force collision, so the critics hovered their hands over the wheel. Then the proponents claimed a small device was added behind the accelerator to depress it secretly, so the critics added a flashlight to the wheel well. Then the proponent claimed the software was hacked to hide warnings and activate the accelerator remotely, so the critics planned the livestream with Ross Gerber who got to use his own, unmodified Tesla. TeslaBoomerMama tried to get ahead of the test by claiming that hidden lights would be placed on the route to blind the Tesla and mannequins would be specially painted with magic paint that looks normal to humans, but would make them invisible to Tesla cameras. Then when the test happened the the proponents claimed Ross sabotaged the tests, and that is where we are now. Does anybody else see a pattern here?

At literally every step of the way the claims of the proponents were met with increased transparency and proven to be unjustified. Yet the proponents use every increase of transparency as tachyonic evidence of wrongdoing. If they show the accelerator now, then they must have been hiding it on purpose in the past thus proving that they were pressing the accelerator in the past. They only show it now because they figured out a new diabolical trick that allows them to show more while still getting the same results. It shows how clever they are at falsifying which is even more evidence they were lying in the past as they are so skillful now that they can still hide with all of these transparency measures.

Here we are now, no way to claim falsification (though they tried to blame Ross), so they had no choice but attempt to duplicate the tests. And lo and behold, the tests done were in fact true, credible, and trustworthy. This shatters the entire thesis that everything is faked as, when the chips are down, the tests are legit.

The actual people who should be discredited are the people continuing to spread known false information that has been retracted by the original author and the people using that as a basis for their attempts to discredit.

Reply on news.ycombinator.com