As is true for every other human on the face of the earth.
-
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 12.4.x is coming autopilot chill is not chill elon is an ass elon rocks fix fsd extreme late deceleration fsd does not commit to go fsd does not maintain proper speed fsd doesn't read signs properly 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 stops when there is zero visibility fsd tailgating in city streets fsd very slow at stops fsd will never enable l3 with hw4
AlanSubie4Life
Efficiency Obsessed Member
There’s a concept called credibility.
Of course there are no guarantees. But that is not the point!
As a world-class (best ever?) sales/marketing guy who places less importance on technical correctness, Elon definitely uses creative license from time to time!
Last edited:
That's still no different from everyone else selling products.
AlanSubie4Life
Efficiency Obsessed Member
Um. I've been reading, listening to some fairly interesting podcasts, and most of this sourced, one way or the other, by Real Researchers. Which of course, I'm not.
But a year or three ago I was listening to a real researcher describing how the human brain, when processing images, separates everything into 5-sided polygons.. and that's how he started, he promptly lost me on the second turn.
I had the distinct understanding after a half hour of all this that he wasn't just putting up random gonzo stuff on what he thought, but that the base level (those polygons, for example) were settled fact.
There are actual structures in the brain, visible with the right equipment, that handle image and color processing.
The point is, here, is that I'm just scratching the surface. Forty years ago, your statement, "No one knows how are brains actually work, yet." may have been absolutely correct. These days.. I don't think that's the case.
At one time people envisaged the operation of human brains as being like Lovelace-style computers, complete with gears. Later popular ideas centered around computers and flip flops.
But Neural Networks aren't founded upon us playing around with silicon: It's actual operation of the axions and such of the nerves and connections in our brains. I've read reports which talk about the layers of the brain and the directions that these signals travel.
A whole, working model? Nope. Getting closer, with active research? Sure. Feedback mechanisms and math? You betcha. Silicon chip models that mimic the real thing? It's in our cars.. How close is what the car does to what our brains are doing? By no means the same.. but there are hints. Mother Nature by dint of bloody tooth and claw weeds out the stuff that doesn't work as well; if something like flip flops were more efficient the NNs, we likely would have developed that, instead. And the fact that HW3 apparently works faster under 12.X than 11.x, well..
The parallel with the brain is that both “see” what their sensors see and learn what the input means. Even as adults we have all had the moment when we looked at what was in front of us but didn’t “get” it for a second or two—or even longer if the data was bad (dark, snowing, etc). That’s enough to use parallel language.
Driving rules might be useful as a boundary check before FSD executes a move, such as the EU g-force limitation on autonomous vehicles taking curves on a highway ramp. They would have to be non-absolute for safe evasive maneuvers, for example. Once all vehicles are moved by computers they can come up with their own rules.
Driving rules might be useful as a boundary check before FSD executes a move, such as the EU g-force limitation on autonomous vehicles taking curves on a highway ramp. They would have to be non-absolute for safe evasive maneuvers, for example. Once all vehicles are moved by computers they can come up with their own rules.
As the CEO of the only tech company in the world with an ADAS working on city streets in an entire country, Elon’s credibility is very high.
That’s not marketing. City Streets ADAS is SHIPPING FUNCTIONALITY WAY ahead of everyone else.
diplomat33
Average guy who loves autonomous vehicles
In light of the V12 livestream, I have some questions about Tesla's end-to-end video training approach:
- Since the approach is dependent on video training, will overfitting be an issue? Could we see V12 work better in some areas where Tesla has more video data and work less well in other areas with less video data?
- Will there be diminishing returns as Tesla gets to more of the long tail? I feel like in the beginning, as Tesla is training on very common cases that are easier to find, they won't need as much data to get the same results. Progress will be faster. But as they get to rarer edge cases, Tesla will need more and more data to get the same results and progress will slow down. Is that a real concern?
- In the US, different areas have different road infrastructure, different driving behaviors, different traffic rules. Won't be hard to truly generalize one NN to handle all of those differences? Won't Tesla need to collect training data from basically everywhere to train the NN on all the differences? Will Tesla need to backtrack on the "all nets" and use some special rules for certain areas?
Last edited:
Elon said on the stream & I quote "We got FSD 12 test drivers around the world"
Patience! This is a hugely positive news that the “ADAS test operators” that Tesla has been recruiting all over the world are testing FSD V12!
(Damnit, I should have applied for the job
)Sure, or many other cases.
Because our perception of the world "driving off a cliff" doesn't necessarily match the internal perceptions and intuitions of a big bag of grey neural goo whose representations we don't understand. Like it's never seen cliffs before or know what a concept of a cliff is.
We don't understand human internal representations either, but humans brains are the product of millions of years of vertebrate evolution which had sharp reinforcement feedback on doing stupid things. And even then, teenage drivers are really bad with poor judgment.
That's a giant unsolved problem in machine learning (and humans too): the confidently incorrect. And suppose it alerts like a frightened mother-in-law screaming from the other seat 400 milliseconds away from a disaster?
Sure, but declaring a level doesn't solve the problem technologically.
diplomat33
Average guy who loves autonomous vehicles
No surprise, the head of Tesla's FSD says FSD will be the bestest thing ever. lol.
Technologically that can be solved by having some nets be segmented (whether fully or partially) or fine-tuned by region and other parts remain the same in an end to end training. Or, like the LLM fine-tuning, a low-rank overlay delta that's region specific on top of a generic base model.
The big problem will of course be the data curation and testing, as now certain adherence to rules can't be tested with normal rule-based but only inferred from a huge set of test cases (and there can be overfitting indirectly there too).
And some time back the previous versions were a "kludgy mess" and V10 or V11 was going to be the awesome sauce
Andrej Karpathy providing some insights into keeping neural network modules around such as vector space outputs not only to bootstrap end-to-end training but also provide insights into what's affecting end-to-end video -> control.
goRt
Active Member
Yours is GA, he refers to a couple of test mulesElon said on the stream & I quote "We got FSD 12 test drivers around the world"
Patience! This is a hugely positive news that the “ADAS test operators” that Tesla has been recruiting all over the world are testing FSD V12!
(Damnit, I should have applied for the job
Our stack is 7 years old
AlanSubie4Life
Efficiency Obsessed Member
I would not argue with this at all (with the caveat that I have no experience with other systems with exception of primitive Toyota).
However that is irrelevant to Elon’s credibility. I was speaking to his specific statements on what is going on under the hood with V12. Not his promises of what will be delivered (setting aside when it will be delivered - which refreshingly he did not talk about in the video…as I recall).
Hallucinations have been used to describe when large language models put together words that sound like facts but are false. More generally and more relatable to what we've experienced with FSD Beta is false positives, e.g., believing there's a pedestrian at a mailbox when there's nobody there. With traditional control code, it tells the vehicle to go around pedestrians, but potentially with control network, it could determine that the mailbox doesn't require swerving around based on other signals over time.
But more broadly, now with end-to-end networks, any incorrect control behavior could somewhat be considered a hallucination if you want to call it that.
In fact, he refers to a few countries.. And these countries have test mules in all the major cities.
Tesla Hired a Lot of ADAS Test Operators in Europe, and We All Know What It Means
Tesla has opened many ADAS test operator positions throughout Europe, hinting at the Full Self-Driving program beginning testing in full force
www.autoevolution.com
BTW, NFI what GA means.
AlanSubie4Life
Efficiency Obsessed Member
Yeah that is what I am referring to. I think exactly how this sort of thing will manifest in driving is unclear (to me at least). But I also think that occasional unexplained (or easily explained by a human) errors are likely with an unconstrained approach (as advertised in the video - this is what I specifically wonder about with Elon’s claims).
I always took Ashook as a yes man guy who wants to be liked by Elon compared to Andrej.
Andrej would never say some nonsense like this.
It sounds like something a child would say. "My mom is..."
Infact I recall Andrej saying the opposite when asked about timelines.
He would always downplay/deflect and say how hard things they were working on were.
100% cringe
