-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
FSD Beta 10.69
- Thread starter Buckminster
- Start date
powertoold
Active Member
Is your car new? Already drive 100 miles?
Did you try to go to software, and tap on advanced 5 times to check for update?
Knightshade
Well-Known Member
I'm sure they had one for HW3 too.
Didn't stop them from being wrong about their guess on how much HW was needed for FSD to ever go beyond L2/city streets.
No reason to think having a HW4 simulation will either- unless you think that simulation is already safety-enough-for-public-release operating such a system.
I've had it over a year, 14k miles, tapping 5 times doesn't do anything nor does pressing and holding. I had subscribed back in sept before wide release and never got it. Now that it's wide release I was hoping it would update right away, but it just says there's no updates available. Looks like 44.25.5 or 44.30.5 is FSD beta, so I should be able to get the new update.
AlanSubie4Life
Efficiency Obsessed Member
It was not released in 2022, over the holidays, is the major issue here with Elon’s hopes and dreams. However, the very poor L2 system now appears to be in the process of being wide released, though it is not clear exactly whether we are quite there yet. So he was close, but not quite correct.
Deadlines aside… I have to admit: it’s a surprise to me that they released it in its current form. It is a lot better than it used to be, but maybe they are figuring most people will just not use it (they have plenty of data on user propensity at this point). That certainly seems to be the design intent.
I don’t care about intersections. I can take care of those details. It seems way too much to ask of the system to figure out how they work, reliably.
If you wait for it after requesting and acknowledging the requirements, it will probably take between a couple of hours and a couple of days, assuming your car is connected to Wi-Fi. This seems typical based on comments here and the experience of myself and a friend.
Some people suggest you can speed this up by actively checking for updates in the software menu. I never checked and it took about 24 hours for me, about 48 hours for my friend. It sounds like you've tried this already, so I'm saying don't worry too much unless it goes two or three days.
powertoold
Active Member
Totally different story and time though. When they were developing HW3, the current stack wasn't even discovered / envisioned yet (transformers, occupancy network, autoregressive, etc.)
According to Karpathy, when he arrived (around time HW3 was first being designed), Tesla's NN team was training models on their desks.
Knightshade
Well-Known Member
Ok. Not sure how that changes anything.
At the time they thought they had a path forward, that HW3 would be enough for. They were wrong.
Then they changed paths (the big re-write) they ALSO thought HW3 would be enough for. They were wrong.
They may well think the new path today means HW4 will be enough. But there's 0 evidence that's any more of a "we think we are on the right path but don't know until we get there" thing than all the previous dead ends into local maximums turned out to be.
Eventually (we hope) one of these iterations really WILL be the one that works, and then we will know how much HW actual autonomy requires. Until then everyone- including Tesla- is guessing.
I recall the current FSD h/w dual core was promoted as a redundant backup safety measure which presumably met the robotaxis wet dream. Of course that idea went away quick.
Disagree with only knowing after the fact. Important company projects minimize risks by spending the $$ up front to identify the issues, requirements, specifications, and then alpha-beta converge to a reliable solution versus wham-bamming something into always failing revisions/releases/promises for improvements. By using proper design cycles, a company always has the option of killing the project, saving face, and money.
The Cybertruck's FSD hardware was rumored to be ~3x more capable versus current hardware. Hopefully Tesla has a better estimate for today's FSD needs.
Tried FSDb on my 2023 Model 3 for the first time today and was very impressed. Was it perfect driving? Of course not. But it was far more impressive than any other ADAS available for purchase. Went 11 miles round trip and only disengaged once, because it was trying to make a right turn and creeping very close to cars going by at 55 mph, and I wasn’t comfortable enough with the system to let it do that yet.
powertoold
Active Member
A whole lot has changed since HW3 was being designed. We have a technique (bird's eye view, occupancy network), strategy (auto-labeling), a system that currently works doing very complex maneuvers, and a clear understanding of what's left to solve. I think people forget how far and magical Fsdb currently is (because we're mostly focused on all the frustrating issues or dangerous maneuvers).
We don't hear about fsd hardware failures, fsd going haywire, etc. And it's currently on hundreds of thousands of cars. This is very amazing...
If you consider all the things fsd can possibly do but doesn't (all its possible failure modes), it's mind boggling.
Nice to see some positive feedback for a change. You're right about creeping. It can be pretty scary thinking the car will launch you into traffic. After awhile if you're brave and wait, FSD will jerk to a stop. Just have you're foot over the brake. Tesla really needs to change this behavior so it creeps less aggressively like a human driver does.
AlanSubie4Life
Efficiency Obsessed Member
Or it will perceive other cars aren’t moving and incorrectly try to take right of way. (Specifically on ULTs, when left-turning traffic from main road is stopped and has right of way in front of you onto the street you are turning from. Have a good video example I will post soon.)
Definitely good to use that brake pedal early and often (when stopped).
And, just going along these lines:
The local cognoscenti on this forum have Made The Decision That The Hardware Isn't Capable Enough To Do The Job. (Which kind of fits with the glass-is-barely-damp mindset seen around here.)
Um. It's not like Tesla has said that. In fact, one has to dip straight into the what looks like conspiracy theories if internal Tesla people believe that the hardware's not capable.
First off, if they truly thought that HW3 couldn't do the job, why even bother writing software for a platform that can't keep up? Quit doing that and write software for a platform that can. Doing anything else is throwing $$ down the outhouse hole.
Unless.. The intent is to deny, defraud and so on. Right into Tesla-Q territory. There are certainly posters on this forum that think that way.
The problem with that thinking is that it Only Takes One Whistleblower. NDA or not, one whistleblower to the world at large will corral the largest number of shareholder lawsuits the world has seen, action by the SEC, and so on. And it's just amazing what the discovery process will bring out in such a case.
Yeah, yeah: Everything that Elon says is aspirational, future performance isn't based upon past performance, and so on. But Tesla has been talking robotaxies and FSD for some time now like it's a thing that's coming. If they figure that they really can't do it, then some kind of damage control would have to kick in. Refunds. Promises of new hardware. There's been none of that.
And, as far as the Cybertruck and its purportedly more capable CPU: HW3 was designed in 2018 or thereabouts. This is 2023, five years later. Moore's law isn't dead, just slowed down. It would make financial sense to build a cheaper, faster CPU, using less silicon physical area, using more modern technologies. When one does silicon shrinkage, one can keep the same number of transistors and clock speed, the die will shrink, and they'll be much cheaper to produce. Or one can keep the die the same size, put down a heck of a lot more transistors at the same clock speed, and get more capability. Or, finally, do the thing that most CPU producers do: Shrink the die some, use more transistors, up the clock speed a bit, and get something more capable and cheaper than the old hardware. Geez, five years is like six dog lives in Internet Time.
The Cybertruck is coming out, what, late this year sometime? So it would make sense to stick the latest and greatest in that box, since, well, not worth the trouble to switch hardware right after a launch. Shortly after the Cybertruck hits the road, I'd imagine that the other Tesla products would start getting the new hardware. Again: Not necessarily because it's required, but more likely because it'll be cheaper.
EVNow
Well-Known Member
I think you disengage too often. I doubt you will get even 3/10 CULTS if you tried it. I guess his best was 8/10 ?
I’ve been getting around 1 in 10 miles. Have yet to analyze the holiday release.
In general for me to get to the next stage (say 1 in 50 miles), I just need couple of improvements
- handle roundabouts better
- pick the right lane, as in don’t make mistakes FSD knows it should correct in 10 seconds
- recognize school zones and lower speed limits
If they put their minds to it, they can do the above in a quarter. Then it will be similar to highway NOA in usability.
AlanSubie4Life
Efficiency Obsessed Member
I’ll post another video at some point which does include another failed ULT. I have one I will post shortly but I have another. But yes very low success rate, all valid disengagements though. It’s just not very good at them especially without a center line marking. It’s particularly bad at dealing with left-turning traffic during a ULT (tries to cut them off through positioning (see first video) or by actually attempting to take right of way (future video)).
(Here’s the first example: What are your most common intervention scenarios?)
I intervene because in most cases there is no excuse for using friction brakes. It’s mind-boggling that it is still doing this. And also because it can’t accelerate, as covered.
Last edited:
Knightshade
Well-Known Member
A whole lot changed from HW2 to HW3 as well- and that also didn't turn out to be the correct technique.
Why were they wrong repeatedly in the past when they were sure they were not, but you're confident THIS TIME THEY MEAN IT?
One of Teslas towering strengths is their ability to recognize when they went down a wrong path, and take a different one.
Assuming THIS ONE must be right and STICKING WITH THAT is how Waymo has been working on this stuff since 2009 and only works in parts of 2 cities. Tesla does the opposite.
We didn't hear about HW2.x failures either.
That's 1000% irrelevant to it not having enough compute to perform >L2 driving assist though. Same for HW3.
May or may not be true of 4. We do not know
Neither does Tesla- because they haven't achieved a working >L2 public-safe system yet.
Since the system they're aiming for has never been done, how, specifically, could they identify the requirements in advance?
It's impossible to know what they actually are.
They guessed at what they thought they would be, based on their approach at the time-- and we got HW2.
That was wrong.
So we got HW2.5
That was also wrong.
So we got HW3.
That was also wrong.
So now HW4 is coming.
They changed fundamental things about their approach multiple times during this process. They may well do so again.
They told us multiple times the current HW was sufficient for self driving, then admitted it wasn't (over different versions of the HW as they realized they were hitting the limits of them). They may well do so again.
Or you could look at publically available info showing they're already maxing available compute use, and have been for a while now, across both nodes to run a single instance of the overall software.
And since redundancy would be required (and was called out AS a requirement when HW3 was first shown) we know, for a fact, HW3 can't do even L2 city streets with redundancy.
And that's a system fundamentally lacking multiple pieces (by Teslas own admission) that would be needed for >L2 (see the CA DMV docs where they explicitly tell us this)
No conspiracy theories needed- just known facts.
Because you can still test PIECES of what you hope is your eventual solution this way. In the real world, on a huge fleet on public roads.
Again we KNOW FOR A FACT they did this on HW2.x ahead of HW3 so there'd be no reason to doubt they'd do it on HW3 ahead of deploying 4.
None of which you can't do on HW4 since it doesn't exist in such a fleet.
NOW we're into conspiracy theories, but it's not from the folks who know HW3 can't do >L2.
Similar threads
- Article
