FSD v12.x (end to end AI)

[Sharps97]

Dojo was a waste of money. Building chips in house is both difficult and expensive. And rarely necessary since NVidia, Intel, and AMD do custom IP designs for various clientele which would allow the use of decades of architectural experience that Tesla did not have at the time (and does not have now). And if Elon mandated any part of that design based on his intuition, not only is it useless for allowing 3rd parties to use (since it would lack a robust software infrastructure even approaching CUDA with questionable architectural trade-offs that don't make sense), it may not even be fit for purpose (at least on a dollar per watt / performance) relative to either H100, MI300, or even Intel's latest offering. And at least all three companies (NVidia, AMD, and Intel) have experience with designing this type of hardware.

The only 'benefit' to Dojo now is that unless you are a preferred customer, you aren't getting H100 any time soon and MI300 is just starting to ramp up (but has a much weaker software ecosystem). Best bet would have been to prototype Dojo, keep it small scale, and stop investing in it once you start to acquire the real deal, as it were.

The only thing left for Tesla to do now is to re-think the entire FSD approach into something that might work any time soon and off to the races =)

 

[lzolman]

Feel 12.4 will be rhe second profound step after 12.3
Seems obvious, but we might be talking about a logarithmic step
Wow

Good news, now we know 12.4 will at least be interesting.
I thought the idea of using idle Teslas' processors to help do training is beautiful. Tesla better pay me for the energy drain, though! lol

 

[aronth5]

With 1 billion miles driven on FSD and growing quickly do we know if there have been any fatalities?

 

[AlanSubie4Life]

With 1 billion miles driven on FSD and growing quickly do we know if there have been any fatalities?

To be clear, about 3/4 highway miles.

Probably something like 300 million City Streets but we’ll never know exactly.

Presumably soon will be growing less quickly unless they move AP/NoA to FSD v11/v12.

 

[Larry Hutchinson]

Had a bit of a scary situation the other day. Was driving to the airport using FSDS and had just merged onto the freeway when a motorcycle moved into our lane and started slowing dramatically. The car paid absolutely no attention and I had to break hard to avoid a collision. Still came about 6 feet from him before he pulled off the road and stopped. Left a message and saw the screen say “message sent.” The car is still at the airport, so it’ll be a few days before it will be connected to WiFi.

A little later the motorcycle and his buddy came blowing by us lane splitting while we were doing 80. The car moved a bit to the left and I thought, well that was good (although it might have been in response to a semi in the adjacent lane.) But then a second later as the cycles were receding in the distance, the car suddenly slowed way down and I had to goose it. My wife remarked that the car did nothing when about to run over the motorcycle but then freaked out when they were long gone.



Yes, this was V11 at work but I wonder if 12.3 would be any better. Do they really have training videos for something like this?

 

[kabin]

Feel 12.4 will be rhe second profound step after 12.3
Seems obvious, but we might be talking about a logarithmic step
Wow

If it's gonna happen at all it should be most dramatic with early version releases.

 

[lzolman]

Had a bit of a scary situation the other day. Was driving to the airport using FSDS and had just merged onto the freeway when a motorcycle moved into our lane and started slowing dramatically. The car paid absolutely no attention and I had to break hard to avoid a collision. Still came about 6 feet from him before he pulled off the road and stopped. Left a message and saw the screen say “message sent.” The car is still at the airport, so it’ll be a few days before it will be connected to WiFi.

A little later the motorcycle and his buddy came blowing by us lane splitting while we were doing 80. The car moved a bit to the left and I thought, well that was good (although it might have been in response to a semi in the adjacent lane.) But then a second later as the cycles were receding in the distance, the car suddenly slowed way down and I had to goose it. My wife remarked that the car did nothing when about to run over the motorcycle but then freaked out when they were long gone.



Yes, this was V11 at work but I wonder if 12.3 would be any better. Do they really have training videos for something like this?

As a rider, this is good info to read about. But I never split lanes (when cars are moving at least). I think it's illegal in MA but I'm not doing it with an RT anyway.

 

[zoomer0056]

As a rider, this is good info to read about. But I never split lanes (when cars are moving at least). I think it's illegal in MA but I'm not doing it with an RT anyway.

Absolutely necessary, and legal, in SoCal when traffic is just crawling.

 

[EVNow]

To be clear, about 3/4 highway miles.

Probably something like 300 million City Streets but we’ll never know exactly.

Presumably soon will be growing less quickly unless they move AP/NoA to FSD v11/v12.

Without providing more data, that billion miles is pointless in terms of city driving. Give us the intervention rates - categorized. Thats the only way to judge FSD progress.

 

[lzolman]

Absolutely necessary, and legal, in SoCal when traffic is just crawling.

I'm as afraid of stop-and-go traffic when riding as I am of taking FSD through my funky intersection down the street. The former I avoid like the plague, the latter I bite the bullet and try every once in a while.

 

[Ben W]

Good news, now we know 12.4 will at least be interesting.
I thought the idea of using idle Teslas' processors to help do training is beautiful. Tesla better pay me for the energy drain, though! lol

I think the distributed supercomputer idea is quite a good one. Tesla's FSD computers are optimized for computational performance per watt (unlike say my old PlayStation 3, which was an absolute power hog when I once tried to use it for SETI-at-Home), and with increasing time-dependent renewable energy sources (see current situation in California with "too much solar" during long sunny days), having a use for all that excess power will become quite important. It's unlikely the Tesla computers could be efficiently repurposed for cryptocurrency, but for machine-learning inference and perhaps training, they're ideal. Probably the "older" generations of FSD hardware won't be economically competitive relative to the newer ones, but with continued growth of the fleet, the majority of computers will stay fairly new.

This should be paired with a "smart grid" approach that adjusts consumer prices based on realtime supply and demand, rather than on pre-set schedules. It's absurd that wholesale electricity prices currently go negative in the mid-afternoon, while my EV rate schedule is charging me higher prices at those times. The Tesla fleet (and EV fleet in general) should easily be able to absorb all that excess production and use it overnight for compute. (Or for driving in general.)

In 2022, California "wasted" about 2400 GWh of excess solar power. All the EV batteries in California add up to somewhere around 100 GWh. If each car were to do the equivalent of 24 full charges per year using this excess power, that would solve the entire problem. (The average Tesla goes through a full charge perhaps once a week, so this would represent roughly half of all EV charging.) Both solar and EV's have the potential to scale up at a similar pace, so the problem should remain solvable. Add a little extra power for compute here and there, and it could help balance the grid even more effectively.

 

[lzolman]

Good vision. If you know about municipal power companies and solar, I've got that combination and it's 16kWp of battery nightmare. The only saving grace is I'm being good to the planet, if not all that much to my bank balance. We'll never break even, and I'll probably have to replace the SolarEdge crap and crap LG batteries (my second set, still flaky, or something is, I don't know for sure yet what lol), meaning everything except the LG PV panels, which seem fine, with a couple of PowerWalls, costing me yet another $20K. But the one thing that does keep working is that the energy I produce is being used, one way or the other, and it's a bunch (my app says 68.6MWh after roughly 4.5 years), and a lot of it is going straight into our Teslas, so I can feel good about it.

[EDIT: Plus this is in Boston, so under the best of hardware reliability conditions, the productivity of the system is gonna suck.]

 

[pg3ttt]

I'm on FSD 12.3.10. Is anyone else, also? My wife and I live in rural NW Wisconsin and use FSD 90% of the time. Although only statistics would hold the answer, FSD 'feels' safer to me than my own driving when we drive into Minneapolis in heavy traffic. I just focus on not hitting anything and let the car follow the navigation. Occasionally, it takes the wrong exit or turn, but I let it figure it out. It mis-identified where we should turn a few days ago and drove down a dead end street. At the end, there was a business with parking surrounding the building. Very slowly, the car circled the building and rerouted. I found that amusing.

 

[drtimhill]

The wobble is the anti-robotaxi - solution isn't clear:

Interesting if you watch the screen carefully just as the car plows forward the "snake" is shifting back and forth between left and right forks in the road. For some reason the car wasn't able to make its mind up .. I wonder if it would have recovered if he hands intercepted. Though I found it amusing the just before that intervention he was talking about how he likes to take curves fast and was hoping the car would. Guess the NNs were listening to him!

 

[AlanSubie4Life]

The wobble is the anti-robotaxi - solution isn't clear:

What garbage. I just kept watching after the initial botching and it just got worse. Had to turn it off after 45 seconds or so.

Of course, 12.3.5 is the essentially the same as 12.3 so this should not surprise anyone.

We’ll see.

 

[swedge]

In 2022, California "wasted" about 2400 GWh of excess solar power.

In 2022, PG&E's Diablo Canyon nuclear power plant produced around 18,000 GWh of power. The problem with nucs is that they are not able to throttle down when they are not needed, such as during peak solar production. Trying to throttle down is what blew up Chernobyl, and also what melted down three reactors at Fukushima. Anyway, that produced some of the power which was in "excess" of what was needed. Also some large, remote solar farms, which also load the transmission lines.

With our solar, we are finding it difficult to charge our EV charging during peak solar production, because the car is usually out and about mid-day. Sufficient home battery capacity to buffer EV charging would roughly double the needed home battery size, making the capital investment for batteries even less attractive than it is now. Same with ubiquitous plug-ins for street and lot parking. So I'm not sure EV's will solve the "too much solar" problem. Shutting down nucs, however, will help. As will grid scale storage, which utilities will be happy to "charge" consumers for, even if the generation is fuel-free.

But I'm not sure why this is a V12 FSD issue....

 

[JulienW]

I'm on FSD 12.3.10. Is anyone else, also?.......

What you smoking or are you from the future? If from the future why didn't we get 12.4.x yet and how were 12.3.6, 12.3.7, 12.3.8 and 12.3.9?

 

[JulienW]

Feel 12.4 will be rhe second profound step after 12.3
Seems obvious, but we might be talking about a logarithmic step
Wow

Then 12.4 shouldn't just "arguably be called v13" it should be called v13. Of course if Elon's Kool-Aid inducing hype on each upcoming version was near accurate we would have been at L5 yeeeeears ago. I have sure learned by years of experience to temper my expectations from what Elon over pumps.

 

[lzolman]

Then 12.4 shouldn't just "arguably be called v13" it should be called v13. Of course if Elon's Kool-Aid inducing hype on each upcoming version was near accurate we would have been at L5 yeeeeears ago. I have sure learned by years of experience to temper my expectations from what Elon over pumps.

OTOH, we know he's timeline-challenged, but then delivers eventually...which in the case of L4/5, if Telsa finally pulls it off, will have been worth the wait. "The difficult we do immediately, the impossible takes a little longer." Yeah, that's my story and I'm sticking to it...

 

[aronth5]

To be clear, about 3/4 highway miles.

Probably something like 300 million City Streets but we’ll never know exactly.

Presumably soon will be growing less quickly unless they move AP/NoA to FSD v11/v12.

Without providing more data, that billion miles is pointless in terms of city driving. Give us the intervention rates - categorized. Thats the only way to judge FSD progress.

Any reason why you didn't attempt to answer the actual question? Do we know of any fatalities with 1 billion FSD miles driven? Or for that matter serious injuries?

The run rate is now another billion miles driven every 68 days.
Tesla FSD users pass 1.3 billion cumulative miles

National average is 1.24 fatalities per 100 million miles driven. (2023 -Q2)
NHTSA Estimates Traffic Fatalities Continued to Decline in the First Half of 2023 | NHTSA