If Tesla releases 12.x for MCU1 vehicles (with HW2.0 computer and cameras upgraded to HW3), that's a pretty good signal of end-to-end being more compute efficient than before. I believe reports of MCU1 FSD Beta 10.x is that the visualizations were slower, but maybe that's less of an issue with 12.x given that it seems to be mostly slower 11.x visualizations anyway. So far older 2016 vehicles that upgraded to MCU2 and HW3 have not gotten 12.x according to TeslaFi, so hopefully these 8-year-old vehicles also get 12.x when it releases wider.
-
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 jerky driving fsd late deceleration fsd stops too early too quickly fsd stops too far behind line fsd tailgating in city streets
A proper answer to this question can’t exist until FSD is finished, because until it’s finished nobody knows what the necessary hardware is. An additional complication is that there’s no actual definition for what ‘finished’ means.
It’s pretty unlikely that the minimum footprint for FSD to be viable is defined by the difference between HW3 and HW4 across all its aspects (that is, that a HW3 vehicle can only be made FSD capable by upgrading every single part to the HW4 spec). It’s somewhat more plausible that it turns out to need more compute power than the HW3 computer has, or that it can only manage a lesser capability than the newer cars because the cameras can’t discern detail at the same distances.
Of course upgrades are possible. Engineers manage to keep spacecraft working decades after their expected lifetime, I’m pretty sure it’s feasible to get more compute out of the same physical space and power that an HW3 computer has. What Tesla have communicated via their denials so far is that they don’t want to upgrade it because it’d be hugely expensive. I think it’s a pretty safe bet that if it became a critical issue preventing them selling FSD to the ~5 million HW 3 vehicles that are out there they’d work it out. Knowing how the customer service side of Tesla operates, it’ll be a paid upgrade, obviously.
One notable thing is that the HW3->4 upgrade is not that revolutionary. No new camera positions, no return of radar(as yet) and no other sweeping changes. Either Tesla are so hostage to the HW3 base that they refuse to acknowledge they need something different, or it’s a relatively mundane refresh because they’re happy that they’re on the right track.
And yes, Tesla haters, we all know you’re certain it’s the former, save your internet-breath. The fact is that nobody knows until Tesla or someone else gets it working.
Version 12 has never been released. It has just been hyped by the carnival huckster, Tesla, employees, and Internet, shills.
I purchased the product over five years ago, you would think I would know if it was released. Maybe Tesla will come up with another test to see if I am a good candidate for the product they sold me. One day they might even let me use my car like I paid for it and own it, and stop putting me in time out if I don’t pay enough attention for dojo.
All that has happened is internal product, testing and hype. They have come up with some new buzz words like end to end AI, single stack.
They’ve been talking about edge cases for years. It is just marketing to keep pumping the stock. Sandy is running out of videos to make on what a game changer everything is a Tesla. he might have to go back to taking apart cars if Elon cuts off the money supply.
Scientific studies prove my car does not hallucinate any more than old hippie, it just needs better coping skills.
powertoold
Active Member
12.2.1 day 3 observations:
1) my fav setting so far is aggressive with my own 14% speed offset. Auto max speed is too unpredictable with aggressive, sometimes too slow in parking lot, and sometimes too quick accel
2) lane changes are usually smooth as butter and flow in chains for multilane changes, very natural and easy to supervise
3) still unnatural creeps and freezes at unprotected turns at stop signs
4) has different driving behaviors in light vs heavy traffic. Still does some mind-blowing decisions from a software pov, however, I'm not sure if decision making is reliable, it may do one thing one day and make a different one another day with slightly different environmental conditions
1) my fav setting so far is aggressive with my own 14% speed offset. Auto max speed is too unpredictable with aggressive, sometimes too slow in parking lot, and sometimes too quick accel
2) lane changes are usually smooth as butter and flow in chains for multilane changes, very natural and easy to supervise
3) still unnatural creeps and freezes at unprotected turns at stop signs
4) has different driving behaviors in light vs heavy traffic. Still does some mind-blowing decisions from a software pov, however, I'm not sure if decision making is reliable, it may do one thing one day and make a different one another day with slightly different environmental conditions
I thought a salient version 12, would flash its lights, to tell them to go ahead. Your car is not very courteous, and now the other people probably think you aren’t , bad dojo.
I don't mean to quibble, as I agree that your basic point is good: stopping for no reason is unwise. But I am trying to understand why TSD did stop. To see what I noticed, look again, please. Viewing on X (link) may have better resolution, but the red car ahead of ego is following a white car which slows as it approaches the tall truck, or whatever that is. The lane ahead was obstructed, making it possible that ego would not be able to clear the intersection. The white car stops, and the read car passes it, as does yet another car from the right lane. Some of this is more clear (though intermittent) in the visualization.
To your point, the tentacle does show a planed stop very early on, so perhaps FSD was being too defensive. Also the approaching car does appear to pull forward in response to ego's stop, so he too thought ego's stop was a courteous move. But once ego did stop, staying stopped was probably wise, as the oncoming car had already started his left turn in front of us.
Most of my very few V12 disengagements are defensive, avoiding situations where I might not have time to intervene if FSD really was doing something unsafe. For example, making a left through opposite direction traffic, FSD started to go when I was uncertain there was enough of a gap, so I hit the brake. Had I let it proceed, I suspect it would not have crashed, but my job is to make certain that it does not crash. I count this as an FSD failure to make me comfortable, not proven to be a safety failure.
Tesla can test FSD using simulations based on these sorts of disengagements, so while these are valuable data, Tesla does not need us to be actual crash test dummies where the safety score is based on severity of injuries.
As one of the lucky few who did get V12, I really look forward to the wider release so that the comments in this thread can be more informative and less speculative.
gottagofast
2022 Model 3 Performance
Since ver 12 is Ai based isn’t it supposed to be constantly improving on its own with not much input from Tesla?
No. There is no learning done in the individual cars.
Artful Dodger
"Neko no me"
You call that a "severe flaw"? That's the way people drive, and it wasn't unsafe, it was a judgement call at best. I'm continuously entertained by detractors trying to use inaccurate language to paint a picture of a fantasy reality. It's the RATE of progress which matters, not the STATE of progress...
People, if the car isn't driving the way you would, then take over. THAT'S how Tesla collects training data: it's the DISENGAGEMENTS that matter (that's the real User feedback), and that comes out by analysing the STATISTICS, not in breathless rhetoric.
Just a dissatisfied customer, that believed the hype 6 years ago, now I enjoy watching the timeline continue to extend, and the shills continue support of utter BS.
The circle flirt is strong with Eon.
Typically the response I get is if you’re not satisfied with your car, you should sell it for something else.
I’m still waiting for promises to be fulfilled, or someone to admit that it was a scam all along
I got the hardware upgrade for free cause I paid for full self driving, what about the fools that paid the $1500 to upgrade hardware that only spots traffic cones, and now that goes away in 12. No computer hardware or camera upgrades for me.
I took away my radar and my ultrasonic sensors, which I paid good money for. And all I got was two
No early adopter button
It can’t even work a playlist better than Winamp amp n 1999.
Took me three years to get Way points
Can we maybe get a recall on the control arms?
They should last more than 20,000 miles
The circle flirt is strong with Eon.
Typically the response I get is if you’re not satisfied with your car, you should sell it for something else.
I’m still waiting for promises to be fulfilled, or someone to admit that it was a scam all along
I got the hardware upgrade for free cause I paid for full self driving, what about the fools that paid the $1500 to upgrade hardware that only spots traffic cones, and now that goes away in 12. No computer hardware or camera upgrades for me.
I took away my radar and my ultrasonic sensors, which I paid good money for. And all I got was two
No early adopter button
It can’t even work a playlist better than Winamp amp n 1999.
Took me three years to get Way points
Can we maybe get a recall on the control arms?
They should last more than 20,000 miles
Last edited:
zoomer0056
Active Member
There is nothing anyone can do for you at this point, aside from just letting you vent. The "buy a new car" option is the most viable in the USA - we vote with our dollars. That's how we affect the free market. You sell your Tesla and get another MFG. That takes money away from Tesla (you are no longer a repeat buyer), and gives money to the new MFG, which hopefully gives you want you're looking for. That tells the new MFG that what they're doing is good, and teaches Tesla that what they're doing is bad (if enough people share your view).Just a dissatisfied customer, that believed the hype 6 years ago, now I enjoy watching the timeline continue to extend, and the shills continue support of utter BS.
The circle flirt is strong with Eon.
Typically the response I get is if you’re not satisfied with your car, you should sell it for something else.
I’m still waiting for promises to be fulfilled, or someone to admit that it was a scam all along
I got the hardware upgrade for free cause I paid for full self driving, what about the fools that paid the $1500 to upgrade hardware that only spots traffic cones, and now that goes away in 12. No computer hardware or camera upgrades for me.
I took away my radar and my ultrasonic sensors, which I paid good money for. And all I got was two
No early adopter button
It can’t even work a playlist better than Winamp amp n 1999.
Took me three years to get Way points
Pursing your lips, crossing your arms, holding your breath and stomping up and down isn't going to really do anything in the grand scheme. Unfortunately, your outrage won't make much, if any, difference. It's just making you miserable standing on your hill, trying to stay warm with your righteous indignation.
Don't get me wrong, I sympathize with you. I'm just telling you what your realistic path forward looks like, hopefully to bring you to a state of happiness. Another option available to you is legal, but that's going to take time/effort/money from you to pursue it. There have been many class actions against Tesla, and only 1 (many years ago) actually made it. The others have failed, or been forced out of class action and into arbitration. There are many small-claims cases (see the thread about them), with various levels of success, but you have to be very careful how you lay out your argument to avoid getting your case dismissed.
gottagofast
2022 Model 3 Performance
For the most part, AI doesn't really run locally. When you use ChatGPT or Microsoft CoPilot, the AI is being done on massive server farms somewhere else and the end result is piped to your computer. In the case of Tesla, the training for FSD neural nets is being done on Telsa's big server farms. The end result of that training is to create a new neural net (many refer to it as a "black box" where input goes in, and controls come out), which is then downloaded to our cars via software updates. The car cannot learn on its own.
I tried 2 v12.2.1 U turn tests today.
The first one made me feel nervous. When FSD was waiting for left and u-turn light behind other cars at about 15 yards from the intersection, it turned the wheel 90 degrees to the left. I was not sure what it was doing and ready to take control if it ran over the street dividing island. When the turn green light was on, FSD returned the wheel to the straight position (phew) then moved to the intersection and successfully made the U turn.
The second U turn did not have the above behavior.
The first one made me feel nervous. When FSD was waiting for left and u-turn light behind other cars at about 15 yards from the intersection, it turned the wheel 90 degrees to the left. I was not sure what it was doing and ready to take control if it ran over the street dividing island. When the turn green light was on, FSD returned the wheel to the straight position (phew) then moved to the intersection and successfully made the U turn.
The second U turn did not have the above behavior.
A bit more complicated answer than @brkaus.
OK: We all own a working Neural Network: i.e., our brains. And, no kidding, our brains can learn and all that jazz.
But just because something is called a neural network doesn't mean that it works exactly like, or even similarly to, our brains.
For one thing, our actual living brains will actually grow neurons under the right stimulus, a capability that I'm not aware of any physical man-made hardware having.
Now, in the Generalized Block Diagram of how any old neural network can be put together (this includes our brains and Other Things), one can have feedback from the output of the neural network back into the input of a neural network. Neural networks work, kinda, on the idea of "weights", where the multitudinous inputs on each cell of a NN have individual "weights"; when enough signals get through these weighted inputs, then the cell in the NN generates its own output, which then goes to further cell's input nets and so on, ad infinitum, or until one gets to the output(s) of the NN.
With feedback, the outputs can be sent back to modify the weights. If done correctly, one might term this, "learning". So, in principle, again, one could train a NN to do something different. And, in fact, that's definitely what Tesla is doing in their development process.
The question of whether that's done on the hardware/software released on the Teslas driving around the landscape has, at least in this forum, ruled out. Apparently, some Tesla developers have made statements on Twitter/X that the actual weights and programs are checksummed and delivered as a fixed package. Of course, Twitter/X isn't the world's most stable platform, so that leaves wiggle room.
On the other hand, it's possible to dream up a scenario where some semi-fixed scratch pad area in the flash RAM gets modified over time.. if Tesla's doing that, then nobody's talking.
So, the proper answer to your question is: Nobody Knows For Sure, but the preponderance of evidence, such as it is, is that Tesla's Not Allowing per-car learning.
$DIETY help me, "AI", or "Artificial Intelligence", is a moving target/marketing term.
Back in the late 80's, AI was used by programmers to describe gigantic programs with weights and multiple if/then/else statements. As an example, people with this software would follow expert diagnosticians in hospitals (think, "House", that medical series on TV), asking said practitioners why they asked certain questions, what answers they were looking for, what tests they were running, and generally how they came to some conclusions about what disease/condition/what-have-you that a patient was suffering from.
When they finished, the software wasn't as good as the expert, but might be better than a poor diagnostician. The package was marketed to hospitals and diagnosticians as a double-check on what might ail a patient, but not the primary check.
I had a neighbor who, of all things, got sucked into the development of such a package, so his expert knowledge of the parts catalog of Mercedes-Benz cars (I am not making this up!) could be foisted on call center people without the neighbor's decades of work in the field.
What people are calling "AI" these days ranges from GPT-whatever large language models which, more or less, are working up the probability of the next word or phrase based upon the words preceding that word or phrase. Just like one can complete sentences for one's loved ones, a GPT can run on amusingly long and actually sound coherent.. but there's no brain in there.
People sometimes say that, what with a Neural Net in a Tesla, that the Teslas embody some kind of AI. Um. No. In this scenario, one is simply saying that there's a powerful computer in there, and then throwing in the words, "Artificial Intelligence" because it sounds cool.
Reminds me of the definition of a "real time" computer. No digital computer I'm aware of is "real time"; that is, able to do its compute with zero delay. (Analog computers can get close to that..). "Real time" is a marketing term; it's usage typically means that "Fast enough for the problem at hand", and not much more than that.
No grok in V12. Grok is a product of xAI, not Tesla.What do you mean by "a concrete example"? Do you mean an analogy to they way humans drive? Okay.
In a rally race team, the driver (pilot) is responsible for the immediate control of the car. If it starts to slide, the pilot countersteers. If a deer runs in front of the car on a blind corner, he dodges it without being told. It's all the 'twitch' parts of driving.
Meanwhile, the navigator (co-pilot) says that an intersection is coming in 200 meters, make a right turn. Slow down for the blind corner ahead (because the co-pilot knows its a high accident location).
The real point of all this is that text prompts can be sent in a just-in-time fashion to guide the pilot to the destination while dealing w. the unexpected. The pilot remains fully in control and carries out the instructions from the co-pilot, which is implemented as a large-language model (LLM) in the Navigation stack.
HTHs.
Last edited:
Similar threads
- Article
