Tesla's Robotaxi Path Just Got Clearer

Added: 2026-05-12

[00:00:00]Let's go through the list of features that they added in 14.3 and then you guys can pick the one you want to talk about but just get the audience here up to speed.

[00:00:09]>> Yeah, well, I mean a good number of those items can just be rolled up under the RL category.

[00:00:15]Right, that's all they are. I mean essentially what they've done is I mean a lot of the training they did in this release was reinforcement learning use cases. And so uh the the coming pothole avoidance will be exactly that. Like it's exactly that.

[00:00:29]Yeah, I mean avoiding animals is exactly that. Better school zone is exactly that.

[00:00:34]All the other ones that we saw there.

[00:00:37]I I call some attention to you know they talked about improved vision encoder.

[00:00:43]Here again I think I think a couple things are true here. One is that you know they can take in more raw data and of course the vision encoder is a big piece of this architecture we rarely talk about and and in and of itself you know they can optimize that piece of that component of the stack um because it has to set up the data for the neural network to process it and you know without that we don't really have any input data. So so it's an important component. Again, that's in the big bucket of optimization. That's in the bucket of optimization. Let's let's use all the data at the potential at the potential data rate we can use that data. And I think that's a you know it's a big theme here which is that at the end of the day uh whatever the neural network can can compute. In other words, whatever it can decide, whatever it can we also want to optimize the entire data pipeline, right? So we got to get that to the proximal limit.

[00:01:47]You know let's get that to the limit on the hardware that we're dealing with and then having have having that which is what 14.3 effectively has done for the team. And I'm sure there'll be more improvements but that's the big box of work there.

[00:02:03]Now they can send that data through uh in any number of scenarios with this reinforcement learning and get the outcomes that they're looking for in the planner. And I think that is you know in my mind much more important than some categorical I'm going to increase the parameter count by 10x. Because if you increase the parameter count by 10x and then you were left with um less than optimal input data or less than optimal run time then what was the point in the first place?

[00:02:34]Right? And further if you increase the parameter count by 10x and then found that having not done those things it couldn't operate in the hardware environment that you have then what was the point?

[00:02:49]So this is the tee up for them to be able to do that and now everybody out there who doesn't really understand what even a 10x improvement in the model 10x parameter increase in the model would do for them will be happy that they do that at some point.

[00:03:04]>> [laughter] >> Yeah, I found it I sorry I'm being a little cynical there but I found it to be completely insane that I mean look, a lot of this falls on Elon for not to say things he there's things he should he doesn't need to say, right? Uh to advertise an increase in the model of 10x leads everyone to think that that has to happen for something to be true.

[00:03:26]Um and and then he doesn't say what's missing in that. You know and so then people are left with filling in the gaps the way they do. Um So you know I I I laugh a little bit when I read those 10x model improvements because I'm like well what do you mean it's it's not you're not improving the model 10x or simply adding 10x the number of parameters. You have no data yet to support that that builds a better outcome. Right? Uh Tesla needs to get to the place where it can get that on that kind of hardware to decide whether it improves the actual driving outcomes.

[00:04:01]Yeah. Herbert, I think there's also something really important. If you go back to your second last slide, this idea that the car can now recover from its mistakes, I think they're recovering without driver intervention reducing unnecessary disengagements.

[00:04:17]This is also actually I think very important.

[00:04:20]So in the past, you know, you might get you know the red hands you know take over now.

[00:04:25]Um I've noticed already with the the builds that I have I'm not on 14.3 yet that that's happening less and less.

[00:04:32]And I think this is a further indication of that. So the vehicle if it finds itself in a situation where normally it would say you need to take over is recovering from that itself. It's it's more and more capable of dealing with these problems that previously it couldn't.

[00:04:46]And certainly in a unsupervised robotaxi scenario that is absolutely critical. It needs to be able to figure it out for itself at all times. Mhm. All right, there's at least the Right. You don't want to phone home all the time and say hey can you get me out of this situation? I I went down a street and it turns out that there's a tree on this blocking the street. What do I do now? Well the car needs to be able to figure this out.

[00:05:08]So I think this is also very important and I'm I'm curious to see as we get more evidence from 14.3 exactly how much more advanced this part of it is. You know I bring up a sub point there which I think is super important. You're kind of saying it without emphasizing it. It's you know robotaxi has yeah you're right.

[00:05:27]In the robotaxi situation it has to it has to square up all these problems all by itself. There is no it's not like you said it maybe there's a tree in the street and that it was never going to hit the tree in the first place. But it has to have a plan to get out of that.

[00:05:41]And if it doesn't have a plan to get out of that it has to call the operations people, right? And so what they're left with is here they're trying to build a system that works at scale.

[00:05:51]Right? And what they've done here is all these piece parts and they don't want to be left with those little components where the the network itself I mean the entire thing can't scale because it's left with all these silly drop off problems or you know indecision problems where it has to call the operations people. So I think that's another thing what they're looking at is they're saying look if we multiply this by a thousandfold, two thousandfold, five thousandfold in the field, what would our operations center look like?

[00:06:24]And they need to get that down to very minimal number of calls to run a service like this.

[00:06:30]So your point is you know very very correct in terms of of robotaxi specifically because in the other scenario we as users that drive our cars we handle this. We never call anybody. We never do anything. We don't need to do anything. We just write about it on X the next day. But in in the case of somebody that's in a robotaxi that problem needs to be solved and that operations center needs to only deal with absolutely critical problems, not everyday problems. And I think that as we go further you know Tesla wants to make this feel very smooth and I don't know when they get to unsupervised on hardware four vehicles but I predict that they won't get there on hardware three vehicles. And so whatever 14 light or whatever ends up there will be um unsupervised and so they want to make sure that it works as accurately as possible for the best possible experience. Wait. Okay, let me argue the the other side of it. The whole reason why they're trying to put FSD upgraded FSD on hardware three so that you can put the hardware three cars into robotaxi which is needs to be unsupervised. It's not going to happen.

[00:07:42]It's not going to happen. Not going to happen. I mean nobody I don't I'm not saying that I've heard them promise that but I'm trying to understand the words that they use about that they won't let the hardware three guys Yeah, I don't I don't think hardware three vehicles will ever run unsupervised in the robotaxi network or in a customer scenario and that's my Okay. my prediction and I could be wrong and I would be pleased to be wrong but I don't think that's going to happen. All right, so just to clarify everything as far as I heard, okay? We agree the three of us here I believe that 14.3 with the improvements that they have uh is increased safety, therefore they will scale unsupervised robotaxi.

[00:08:25]Now what about the comment about unsupervised coming to regular customers like us?

[00:08:32]If if 14.3 is enough to do unsupervised robotaxi, let's say just in the geography of Austin, then Suren should be off offered the opportunity when he's got his car in that geography to say hey you know you can go ahead and do unsupervised just as if you were taking a regular robotaxi.

[00:08:49]Or do you guys think that you need to wait for version 15 before they can then say all Teslas in whatever geography is unsupervised? No, I don't think they need to wait for that version. I think I mean I think unsuper I mean largely customer unsupervised and robotaxi is the same it's the same problem. I mean there's some differences. Yeah, there's some obvious differences here but you know I think that they are like I've said before I think so much of their focus is on getting robotaxi to go. Um I don't think they're as concerned with customer unsupervised. I think a lot of us are obsessed with it but I don't think they're as concerned is like they know that there is a laundry list of items that are specific to that.

[00:09:31]But I think they're quite busy over there.

[00:09:33]And I think that they're focusing most of their attention on the laundry list of items that make robotaxi scale.

[00:09:39]And I think so I absolutely think robotaxi goes out before they worry about the customer side of unsupervised and then they say okay, let's go and chop off that list. And I don't know what's exactly on that list. I can hypothesize but you know there's sort of non robotaxi things that are on that list in the customer bucket. And I think it's just priority two, if you will. I would agree with that and I would just overlay the the business risk aspect on it. It would make more sense to scale up robo-taxi, start earning a bunch of revenue, making profit, then release the customer unsupervised. You at least have a massive then cash flow to deal with any lawsuits that come your way because of whatever stupid things happen. Elon had a post this week where he said, you know, we'll we'll probably end up avoiding 90% of accidents and we'll save millions of lives.

[00:10:27]But the other 10% you know, they're unavoidable and we're going to get sued for that.

[00:10:33]And so, what's the best, you know, way to deal with that? Well, one is is the best software you can possibly make so that you reduce the chances, but the second thing is just to have a lot of revenue and cash flow to deal with the cost of doing this.

[00:10:46]>> that. Yeah, it's cost of business. Got you.