Search results

I think the question is can a model of this sort assist with tasks required by AVs and the answers from this paper make it clear: absolutely they can. GPT4V wasn't even trained to do this thing. You can imagine a smaller model built for much more specific purpose and trained on only the most...

Full paper here: https://arxiv.org/pdf/2311.05332.pdf

@Bladerskb how’s Xpeng doing, two years on from this thread?

OP, FSD or FSDbeta? They are different...

Stop sign behavior is variable. Sometimes it behaves relatively humanlike and other times it's too cautious.

Perhaps I'm not being clear. Sometimes, you see a sign and it tells the car what it can do. Sometimes you see a sign and it tells the car what it can't do. Sometimes you don't see see a sign and, depending on where you are in the world, you are allowed to do something or it is illegal to do...

Right, and that's my point: most of the time, a sign will tell you what you can't do. But sometimes it's the absence of a sign that tells you what you can't do. How does the model figure out which is which?

Opposite reaction: extremely f'ing believable.

Thought experiment for current end-to-end AI systems: Imagine a world in which in 99.999% of the world, you may turn right on red UNLESS there is a sign that says "No right turn on red." In %0.001 of the world, it is illegal to turn right on red unless there is a sign that explicity allows you...

It's interesting right because DeepMind created AlphaGo, giving it knowledge of many human games, and AlphaGo ended up able to defeat the world champion. Since then AlphaZero throws out the human games altogether and simply learns from itself.

OP, any thoughts? Seems pretty good to me for alpha:

Yeah if 10x better than 11.4 isn’t “better” to you then yes, this won’t be “better.” I realize we do need 1000000x better than 11.4 to get to robotaxies but for my commute, 10x better than 11.4 is amazing (as I already use 11.4 for every single commute). We then turn to, what is the rate they...

Based on the recap right above this that I posted, I suspect this will age poorly. Car is more humanlike than ever in that video.

Can someone transcribe the Qs and As?

My commute features an intersection with a flashing red for me every day. In previous versions, FSD didn’t support, but 11.4.4 handles it ok some of the time.

Thank gord

Just pointing out it isn’t the worst. Perhaps the worst for you. But we’ve seen much worse.

You clearly never drove it on the very early FSDb versions...

New behaviors on 11.4.4 on my morning commute today...one big lane selection issue that was super consistently wrong in weeks' past is now working properly.

Not sure but I think the paper you cite was from 2022 and predates GPT-4. (At least, there are no mentions of in the paper you cite and the paper says "Submitted 10/22".)

Ripping out 300k lines of C++ with a Dojo-trained neural net is definitely a "rewrite" but the reason why it's not a total rewrite is because presumably all of the other neural nets they've already deployed into production aren't being changed significantly.

That’s never seemed to matter for me.

Relevant to the earlier discussion about “a mind”: Generative AI Space and the Mental Imagery of Alien Minds

I've had FSDb "improve" while on the same version. I've also had it regress. I attribute it to map data but tbh no one knows.

Sure, what I mean is the single stack, which comprises both highway and city street driving, is radically different from the original AP stack. So in some sense, "AP" has very much changed. My FSDb car on the highway is radically better than it used to be.

"AP" on FSDb is radically different. Would be great if they can push it out to most of the fleet.

This is why I pre-emptively slow down the FSD when I'm about to encounter traffic like this...I greatly prefer it to be 8-10 car lengths away from the car in front, just to ensure stuff like this can't happen.

Here is MD it handles them ok, not great, not horrible. I almost always let it drive thru the booths (I use EZPass).

D Good we actually super agree with each other. Once those crappy human-written 300k lines of code are ripped out and replaced with a neural net that has taught itself to drive, then we can start to compare FSD to a fly.

Exactly. We are using the words in totally different ways. Pretty sure Elon is using it in the same way I am.

First of all, if my Roomba had some neural nets to control its planning and to understand the world around it, I would actually 100% say it has a mind. My understanding is that my Roomba has some C++ code that makes it funciton in a brittle, non-thinking manner, and that's the main difference...

My cat has a mind...she knows how to navigate around the house, find the liter box, go to where the food and water is, where her favorite cat tree is. She knows when I say the word "food" it means it's time to go to the food bowl. She has some sort of memory and when she bites my fingers in...

Pretty pointless to argue about what Elon meant by his tweet (too vague to really mean anything) but "literally everyone" is 100% incorrect. Just one example contradicting this: Oversight of A.I.: Principles for Regulation | United States Senate Committee on the Judiciary From the three...

Has a mind probably means has a basic understanding of the world. You can see sparks of this in LLMs. Seems reasonable for this to be in the neutral nets of FSD too.

Well that and theoretically there can be "infinite" virtual drivers...and also theoretically the costs could quite easily come down the moment a competitor appears.

First off, we don't know what v12 alpha is like so it's hard for us to judge. But I have long believed that for us to actually get to "real" FSD, we'd need AGI, since driving itself requires a bunch of different intelligences (vision, planning, reaction to enviornment). I've also thought that...

If you care more about FSDb than...idk...some minor upgrade to Spotify or whatever, then it seem prudent to stay on the betabuild. I personally feel that the best aspect of the car is the FSD software stack, so I would want to stay "enrolled" in beta for as long as possible.

OP, the good news is that EAP is great even without the disabled features. Autolane change and exit taking on highways is fantastic.

Sorry to be a blind dumdum but ...the message on the screen, what is it saying? It's too blurry for me to read.

Well that's definitely one way to look at it...probably not the best way, but a way. My CRT monitor back in 1990 had some nice features too that 2023 monitors simply don't have (example: you can use it as an amazing doorstop!).

In general, completely agree, we don't really know where the wall is. Recent remarks from Elon seem to indicate that they have realized the problem is simplier than they realized...unclear what that means, but perhaps suggests that end-to-end neural nets are the breakthrough we need here, or...

As someone who has used 10.69 and 11.4.x, I can 100% say there has been some improvement. Maybe not as fast as one would hope, but improvements for sure (ex: roundabouts on my commute are much better in recent builds). They're in the process of yanking out old logic-based code and replacing it...

This is what we know for sure: at the moment, we are SOFTWARE LIMITED, meaning the thing holding us back from better FSD is the software, not the hardware. How do we know this? Because the software keeps getting better under HW3. You might not like the rate of improvement, but unquestionably...

It's not arbitrary nor is it hidden, you just lack historical context. First, don't call them alpha testers, they never get "alpha" releases. (We aren't even sure who gets those but definitely not the youtubers.) They get betas weeks before normies like you and me. Back in the day, 2020...

IDK, someone said "are you talking about FSD Beta or FSD since those are different" and then another person said "There is NO SUCH thing as just FSD" and all I'm saying is there is definitely such a thing as "just FSD." I agree that Tesla has done themselves no favors here by their naming and...

Please post video. Hard to diagnose problems from textual descriptions alone.

There IS such a thing as FSB non-beta software...it's what you have when you don't opt-in to the beta builds. The software is pretty trash (ex: it requires you to confirm passing through green lights on city streets) but it exists nevertheless.

Once you get your VIN enrolled in the Beta branch, it seems like a bad idea to try to convince Tesla to unenroll you. Yes, right now, you're behind the main branch in terms of features, but you can imagine a future after the Beta and Main branches are merged into a single stack, and then...