stopcrazypp
Well-Known Member
Tesla had never done things like adjusting weights on the fly in NNs via the cloud. It was always done as a software update. I don't see anywhere where they said they would change this.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
FSD v12.x (end to end AI)
- Thread starter Buckminster
- Start date
-
- Tags
- #fixphantombraking autopilot chill is not chill elon is an ass fix fsd extreme late deceleration fsd does not commit to go fsd does not maintain proper speed fsd hard and late braking fsd has decision wobble fsd jerky driving fsd late deceleration fsd sometimes may not detect train crossings fsd stops too early too quickly fsd stops too far behind line fsd tailgating in city streets
Has it not been assumed that Tesla has been updating weights on the fly for quite some time on FSD Beta? I was under the impression that was the cause of the perceived improvement during the course of one build.
If its end to end nets then the utility of this goes up and the utility of software updates goes down.
Last edited:
stopcrazypp
Well-Known Member
Reference to where this was "assumed"? It has been known for some time the maps are updated separately (it shows in the menu), but no evidence NNs are updated via a separate data channel.
I can’t dig up a specific reference but that has been the assumption I’ve been under for probably a year now based on comments from prominent tests on X discussing changes in performance of a build and disparities between experiences in one build between cars.
EVNow
Well-Known Member
Interesting.
Could mean V12 is closer than 2nd half of next year I was thinking.
Ofcourse difficult to say. Anyone know what’s the longest between employee release and rollout ?
As of today, FSD is only partially a machine learning driven system. Roughly speaking, the perception layer is neural nets and all the driving decisions are hand coded traditional software.
There's a couple probable explanations for the perceived improvement during one build that don't involve any adjustment at all to the ML parts:
- Tesla has server side tuning of some limited aspects of the driving decision code. Something like tuning how assertive it is, how sensitive it is to obstacles / people / other cars, how much risk it's willing to take, minimum allowable approach distance to cars / people / obstacles, etc. I've noticed sometimes FSD feels particularly hesitant and needs to be pushed through intersections or otherwise goes too slow. Could easily be explained by the system being tuned to be extra conservative!
- Map updates (potentially collected from other cars, so the fleet is collectively reloading it's shared map cache after an update)
- People just get used to the release and perceive it as performing better
My guess is on #1
stopcrazypp
Well-Known Member
The same thing had been claimed about normal AP and people assuming the car is doing learning on a per vehicle basis, when it has been fairly well established there is no per vehicle learning. There are disparities in how the the car operates between different users just due to huge variance in the environments they operate. Conditions can also change just due to the time of day it is being tested in even if the location is the same.
As for differences in a single car, note that the nature of NNs is there is always a probability of doing something different given similar input, due to the fact it's not deterministic like hard coded functions.
Fair.
Although I’ve seen a few comments that tesla hasn’t told us that they are updating weights on the fly ipso facto they are not. But Tesla doesn’t actually tell us much detail of how the program is run. If they were doing it I don’t expect they would tell us?
In respect to evidence I guess I feel like greentheonly may have - just going off memory - implied it was happening and while I have disagreed with his long term outlook he did seem pretty technically knowledgeable and had torn down plenty of totaled computers
"Humans are basically a big data stream of photons in that produce a tiny data stream of motor commands out" - Elon
Does this disappoint anyone?
Last edited:
Don't hold your breath. This version is unlikely to go to any customers.
Um. I can't speak for other media platforms, but the several times I've made guess-and-by-golly extrapolations around here that Tesla might be doing things like varying weights on a per-user or per-group basis have been shot down, repetitively. Mainly by people who are quoting Twitter/X posts by developers who have claimed that (more or less) it's all in FLASH, checksummed, and can't be changed. (Hope I got that right; if somebody wants to correct that, feel free.)
It's been reported that extensive uploads from FSD-b cars to the Mothership have occurred after driving around for a day. So, there's some guesses that Tesla has been tracking what's been going on with the fleet.
IF (and that's a pretty big "if") Tesla is downloading different weights to, say, different FSD-b users, or groups of users, then playing statistics to see how the weights/algorithms are going, well, sure, that might be possible. And there's some speculation that map data might be getting changed on a per-drive basis.
But notice all the hedge words I've been using: If, might, guesses, speculation, etcetera. And that's perhaps the point: None of us hanging out here on the teslamotorclub forums are actually coding for Tesla. Those that are posting on X/Twitter/what-have-you might be reporting what they're up to, but it's hardly official. So, what you find here is mostly guesswork, guided in part by the occasional video or official statement from Tesla.
And even when Tesla does say or report something vaguely official (like, say, that V12 drive that Elon and Ashook (sp?) took a couple months ago), you'll note that the forum will be up to its keister with posts of, "It's all a fake!", or, "It'll be on Elon time in ten years." or some other fatuity.
As a result, don't trust too much of what you read around here. Certainly without making some kind of judgement call as to how, "true" a post is, for a given value of "true". It's fun speculating, of course, and sometimes those speculations come true. Or don't. Which is part of the fun.
What I mainly use the active FSD threads (as in, the 11.x thread where people are reporting actual experiences) is a guide on What To Look For. While FSD-b isn't all that dangerous if one stays alert as one is supposed to, Not Being Surprised is a good way to keeping life and limb together.
Last edited:
Agreed. We are like blind men trying to figure out an elephant.
Actually, I am pretty sure that green has said that weights are only updated via a software update. The only thing that is "updated" over the air, is that when you use navigation it downloads map hints as part of the routing, which can include things like lanes, stop sign, traffic lights, etc. But that only happens when using navigation, and nothing is updated/stored in the car. (So, if you drive the same exact route with and without navigation you might get completely different results.)
jabloomf1230
Minister of Silly Walks
I think we are like elephants trying to figure out a blind man.
diplomat33
Average guy who loves autonomous vehicles
Humans have other senses too. We also take in sensory data from sound, touch and smell, not just photons. Also, humans output more than just motor commands. We also output thoughts, feelings and emotions. Elon ignores that. To reduce humans to just a vision end-to-end system that only outputs motor commands is a very odd way of looking at the world because it describes humans like we are just emotionless robots. But humans are not emotionless robots. It ignores so much of what makes humans special.
If we continue to discuss this topic we will get to endless arguments about philosophy, religion, abortion, paradise, election...
sleepydoc
Well-Known Member
I think I know which end of the elephant I’m at…
I’m wondering what the team even does if they roll this out to the select group and find issues, just pump in more video clips? Do they pump in a certain set of video clips?
Still lost at what all this end-to-end stuff means in practical terms, so far the process seems pretty much identical.
