FSD v12.x (end to end AI)

[mtndrew1]

I don't follow your logic. It was peanuts to upgrade FSD purchasers from HW2.5 to HW3, but it will be prohibitively expensive to upgrade the compute on roughly the same group of vehicles? Unless you're saying tons more people are buying FSD at it's current $15k price.

I’ve assumed for some time now that the reason the price of FSD kept rising was to leave a buffer for pricey future retrofits.

If they’re selling this thing for $15k and each car needs $3k worth of work they’re covered.

I’ve been in the FSD beta group for almost two years. While the system does some things much better than it used to, it’s no closer to me being able to hand the car off to an inexperienced friend without fear.

Given the glacial rate of progress towards actual usefulness I expect that the fleet will need new computers and high-res radar retrofits to even begin to approach the lofty promises.

 

[clydeiii]

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

 

[Knightshade]

By that logic, they didn't need to upgrade HW2.5 buyers to HW3 either, but they did at no additional cost for those that already bought FSD.

They're selling the HW3 retrofit for $1,000 retail right now, including labor. So how much can it cost them to swap a board? $200 in parts and $200 in staff time, max?

If V12 is required for the levels of safety Elon said was achievable with HW3, then I think they could make a retrofit happen.

It's VASTLY harder to retrofit HW4.

HW3 was designed to be drop-in compatible with HW2.x cars- and ONLY required swapping the computer.

HW4 has an entirely different set of cameras, and the computer is physically different, the wiring and connectors are different, it also requires upgrading the MCU-- and on top of that both the MCU and driving computer require the new 16v power system the newer lithium-low-volt cars come with.


It's apples and brain surgery in comparison.


I suppose Tesla COULD develop something that was physically compatible with HW3s unit, and also make it compatible with the older 12v system, and also design a Ryzen-based MCU that runs on that wiring.... All of that would be a lot of cost in the first place... then the added cost of future development having to handle BOTH HW4 systems with old 720p cameras vs HW4 with newer 4k cameras.

So none of that makes much sense financially for Tesla.

Especially when the vast majority of FSD buyers bought their cars after ~March 2019, meaning Tesla never promised them more than L2 anyway.


They will need to do SOMETHING for the rough-ballpark 40-50k FSD buyers prior to March 2019, but it's unlikely to be a retrofit of existing vehicles.

 

[gsmith123]

With V12 adding end-to-end AI on the horizon, what do you all think this means for development in the short term?

End to end AI was (presumably) motivated by challenges to cleanly implement fixes and features in V11 without end to end AI. Even improvements that are possible in V11 could still be wasted time if they need to basically implement the same fixes in V11 as well.

Over the next 6 months will releases slow down? It's been almost two months since 11.4.4 (the version most testers are on) came out.

On the other hand perhaps releases continue if either a) the migration to end to end AI can be done incrementally or b) they continue to iterate on hand-coded driving policy in parallel to provide the guardrails for the AI driving policy.

 

[goRt]

V11 - a GUI update (as with most Vxx)
V12 - a backend update - does that mean the RoW will just wait, that the RoW will get all the fancy stuff, or Musky's made yet another unthought-through boast???

 

[mongo]

With V12 adding end-to-end AI on the horizon, what do you all think this means for development in the short term?

End to end AI was (presumably) motivated by challenges to cleanly implement fixes and features in V11 without end to end AI. Even improvements that are possible in V11 could still be wasted time if they need to basically implement the same fixes in V11 as well.

Over the next 6 months will releases slow down? It's been almost two months since 11.4.4 (the version most testers are on) came out.

On the other hand perhaps releases continue if either a) the migration to end to end AI can be done incrementally or b) they continue to iterate on hand-coded driving policy in parallel to provide the guardrails for the AI driving ppolicy.

11.4.7 releases soon
V12 is in early testing with Tesla internal fleet

https://twitter.com/x/status/1689948253532659712

 

[JB47394]

They will need to do SOMETHING for the rough-ballpark 40-50k FSD buyers prior to March 2019

I assume that the arbitration clause was part of their agreement, and that it limits the opportunity for a lawyer-fueled class action lawsuit. How many people on arbitration will choose to go through that process as a result of their dissatisfaction? They were early adopters, buying very expensive cars, and FSD wasn't as expensive as it is today. Just on a cost basis, it seems like doing nothing is the way to go. It's certainly not good marketing for Tesla, but I wonder how much harm even that would cause.

Personally, I'd like to be able to run V12 on my 2022 Model 3 whenever the software is ready.

V12 is in early testing with Tesla internal fleet

I think Elon is getting better at his public statements. He included "imo".

 

[Knightshade]

I assume that the arbitration clause was part of their agreement, and that it limits the opportunity for a lawyer-fueled class action lawsuit. How many people on arbitration will choose to go through that process as a result of their dissatisfaction? They were early adopters, buying very expensive cars, and FSD wasn't as expensive as it is today. Just on a cost basis, it seems like doing nothing is the way to go. It's certainly not good marketing for Tesla, but I wonder how much harm even that would cause.

All arb costs are borne by Tesla past the filing fee.... so they'd rather do one class action than 50,000 arb cases they pay for- on top of which customers could, and at least some did, opt out of that clause-- on top of THAT the clause also says you can simply file in small claims instead- which is generally VERY cheap, and a number of owners have already gotten FSD refunds this way without waiting on a class action.

On top of THAT-agreeing to individual arbitration does not, inherently, waive your right to be part of a class action. Individual and class arbitration are different things legally-- see Lamps Plus Inc v. Varela where SCOTUS made clear agreeing to individual arb does not waive your right to be part of a class action.

 

[Buckminster]

https://twitter.com/x/status/1689948253532659712

 

[JB47394]

All arb costs are borne by Tesla past the filing fee.... so they'd rather do one class action than 50,000 arb cases they pay for- on top of which customers could, and at least some did, opt out of that clause-- on top of THAT the clause also says you can simply file in small claims instead- which is generally VERY cheap, and a number of owners have already gotten FSD refunds this way without waiting on a class action.

But they won't face 50,000 cases because it would require 50,000 people motivated enough to go through the arbitration process. A class action suit only requires motivated lawyers, and they have significant motivation because it's their chosen profession and, if successful, there's a big payout. The arbitration group wasn't even motivated enough to opt out of the arbitration agreement (given that they were aware of it).

agreeing to individual arbitration does not, inherently, waive your right to be part of a class action

My right. Heh. Funny.

 

[Knightshade]

But they won't face 50,000 cases because it would require 50,000 people motivated enough to go through the arbitration process.

Even if you only had a 10-20% action rate that's a LOT of arb bills to pay.

More than it'd cost them to just make them a better offer- and with a ton less bad press. On top of that in most (though not all) states arbitration can include punitive damages-- which I'd certainly be arguing for if I was showing a company had sold me something 8+ years ago, took 100% of the $ up front, and never delivered nor even today seems capable of delivering...

That said- anyone who does want to take individual action can also as I mention go small claims route anyway- simpler, quicker, cheaper, and I've yet to see anyone go that route and have Tesla even bother to defend themselves significantly.... though no punitive there and you're likely looking at at best a refund-with-interest there rather than say a car buyback someone mentioned as that'd likely exceed what you can ask for in small claims.

The arbitration group wasn't even motivated enough to opt out of the arbitration agreement (given that they were aware of it).

My right. Heh. Funny.

Again, the arb group isn't necessarily excluded from being part of the class action- I mean- it's a legally different thing is the point.

Individual arb waivers do not prevent a class action from being filed per the supreme court in the case I cited.

 

[DrChaos]

https://twitter.com/x/status/1689948253532659712

If they're doing an almost total rewrite now, it means they're very far from production L4. Something like that takes years of tuning on a reasonably stable codebase. Going from the per frame networks to the 3-d stitching and occupancy networks was already a total rewrite. And oops, they did it again?

It's possible it's not really a total rewrite but the internal managers have figured out what Elon thinks is Kewl and prevents them from getting fired in a snit.

Or it is, because at long last some of the smarter people have finally convinced him that the path they were on was not going to make it to the destination

Or it's internal programmer churn as too many people leave and the next batch comes in and they can only progress in their career by throwing out what predecessors did and saying it was junk (when it wasn't) and how they can do so much better (and they find they can't as it was a really hard problem and there's a conservation of difficulty issue) or there was huge technical debt and confusion because Elon pushed for crazy deadlines, and then fired too many people or drove them away. That results in a *sugar* and confusing codebase so they have to rewrite because everyone who knew the old junk is gone.

Or Elon being distracted at Twitter let them finally fix things (best outcome)

 

[Buckminster]

If they're doing an almost total rewrite now, it means they're very far from production L4. Something like that takes years of tuning on a reasonably stable codebase. Going from the per frame networks to the 3-d stitching and occupancy networks was already a total rewrite. And oops, they did it again?

It's possible it's not really a total rewrite but the internal managers have figured out what Elon thinks is Kewl and prevents them from getting fired in a snit.

Or it is, because at long last some of the smarter people have finally convinced him that the path they were on was not going to make it to the destination

Or it's internal programmer churn as too many people leave and the next batch comes in and they can only progress in their career by throwing out what predecessors did and saying it was junk (when it wasn't) and how they can do so much better (and they find they can't as it was a really hard problem and there's a conservation of difficulty issue) or there was huge technical debt and confusion because Elon pushed for crazy deadlines, and then fired too many people or drove them away. That results in a *sugar* and confusing codebase so they have to rewrite because everyone who knew the old junk is gone.

Or Elon being distracted at Twitter let them finally fix things (best outcome)

It shouldn’t be compared to a rewrite as it barely needs any new code. It will need several training iterations and is limited by compute. Dojo will speed things up. I expect beta to go to first customers in October.

 

[uscbucsfan]

It shouldn’t be compared to a rewrite as it barely needs any new code. It will need several training iterations and is limited by compute. Dojo will speed things up. I expect beta to go to first customers in October.

Even though Elon is calling it a re-write.

 

[clydeiii]

Even though Elon is calling it a re-write.

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.

 

[diplomat33]

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.

I think Elon mentioned that the lines of code being replaced was for planning and control. So it would only be a rewrite of the planning/control component, not the entire stack. However, Elon also specified that V12 would be end-to-end, ie vision in, control out, which would imply a total rewrite. Although it is possible that Elon was simply referring to the individual components (perception, planning and control) being NN based and not actual end-to-end where there is only one NN for the entire stack. I suspect the latter. I think Tesla is keeping separate perception, planning and control components, just each component will be all NN now. I don't think Tesla will actually do true end-to-end where it is just one NN from vision to control. I feel like that would require too big of a rewrite.

 

[Buckminster]

https://twitter.com/x/status/1689948253532659712

100-99=1
1% is written code. Not a rewrite. New NNs. Maybe reconfiguration?

 

[mongo]

V11 NN : layers followed by output that feed C++, so a big array of all things the code needs
V12 NN : layers plus more layers (probably) followed by speed and direction output.
Highly different format (at least near the end) and that end change propagates into training.

Each shift from discreet function NN with quantized outputs to combined NN without intermediates increases training load even if total weight count remains the same.

 

[Knightshade]

Not a rewrite.

You've made this claim twice now, yet you directly quote Elon saying it's an "almost total rewrite"

Why do you keep insisting Elon has no idea what he's talking about?

 

[Buckminster]

You've made this claim twice now, yet you directly quote Elon saying it's an "almost total rewrite"

Why do you keep insisting Elon has no idea what he's talking about?

After a neural network has been created, it must be configured. The configuration step consists of examining input and target data, setting the network's input and output sizes to match the data, and choosing settings for processing inputs and outputs that will enable best network performance.

Just making sure folk realise how little human written code is in this build. Literally not written as traditional lines of code. No debugging etc. Elon is just using traditional wording.