Latest comedic fantasy from Tesla Q2 2025 earnings call

There’s nothing in Waymo’s blog on its architecture to support what you’re saying. The modules aren’t redundant to each other, they feed output to input. According to Waymo, there is but one module taking sensor inputs and outputting the world view:

The prize isn’t in delivering rooms, but in completing the house. Waymo has, to date, provided more rides in more cities than Tesla. But, if we’re looking say, 5 years down the road, who will have the lion’s share of the autonomy dollars?

As one of the articles linked above says:

In a way, Elon’s strategy is a bet that the same scaling laws that created LLMs will lead to self-driving cars working too. As Richard Sutton argues in his piece The Bitter Lesson, general-purpose methods that exploit ever-growing computation and data usually outperform systems packed with hand-engineered, domain-specific knowledge.

Have you read The Bitter Lesson. It created quite a stir back in 2019, but led to big advances in AI:

We have to learn the bitter lesson that building in how we think we think does not work in the long run. The bitter lesson is based on the historical observations that 1) AI researchers have often tried to build knowledge into their agents, 2) this always helps in the short term, and is personally satisfying to the researcher, but 3) in the long run it plateaus and even inhibits further progress, and 4) breakthrough progress eventually arrives by an opposing approach based on scaling computation by search and learning. The eventual success is tinged with bitterness, and often incompletely digested, because it is success over a favored, human-centric approach.

While not attacking modularization per se, it does attack the creation of AI software that attempts to mimic how humans understand and process data, and modularization is most often set up in such fashions.

The Bitter Lesson was astounding and controversial in 2019. Today, pretty well accepted, as the rise of giant Nvidia data centers attests.

There isn’t one prize - or even a prize at all. A finished room has economic value regardless of whether the whole house is finished.

If we’re looking five years down the road, there’s a decent chance that Waymo has the lion’s share of the autonomy dollars - depending on when during that five years Tesla manages to “solve” for robotaxis. Most of the “house” in this metaphor is pretty much irrelevant. If you can cover about 75-100 metro areas, you’ve covered at least 3/4 of the robotaxi market. Sure, there’s hundreds and hundreds of other cities - but who cares?

There’s an advantage to getting into those markets first - being the first mover, getting your brand established, getting the better free media coverage, etc. If Tesla rolls into town a few years later, and Waymo’s already in place and covered the market, they run the risk of being the “InDriver” to the Uber/Lyft duopoly. Especially since Tesla doesn’t have much institutional capability for driving branding or advertising, and has a lot of consumer brand issues right now. So even if we wake up in three years and Tesla has figured out a way to run robot axis everywhere, that probably won’t matter much. They’ll have the Cedar Rapids, IA market all to themselves - but they’ll have ceded Waymo years of head start in the big cities.

As for The Bitter Lesson, it’s the same type of question - what does it mean for one approach to be “better” than the other approach? The compute-centered approach leads to massive jumps in some aspects of performance, hence the rise of those data centers. But you end up with a product that (to continue the metaphor) has painted 90% of every room in the house. It can do vastly more than human-centric approaches do, but it is never 100% reliable because of the way it was developed.

Here, I freely admit I might be colored by my own specific experiences - which I know should not be generalized, but its where I’ve had the most interaction with AI. I’m a lawyer at a BigLaw firm, and have had to sit through a lot of training sessions on the state of AI (both of our own use and for advising clients). AI is great at replacing junior lawyers - it is a good substitute for a middling entry-level performer. But you always have to check its work, because it is prone to confidently making mistakes that no junior lawyer ever would (they’d make different mistakes).

So that’s a leaps and bounds improvement for replacing whoever’s doing your first drafts in your “working with words or symbols” industry (law, communications, coding). That gives you a lot of economic utility. But it’s not useful in applications where it has to be correct 99.99999% of the time the first time.

Breadth vs. accuracy.

The hope, I guess, is that with enough compute and enough data you can also solve the accuracy issue. But not a lot of progress on that front as of yet. AI is improving a lot faster on the “breadth of application” dimension than on the “eliminating hallucinations” dimension:

https://www.nytimes.com/2025/05/05/technology/ai-hallucinations-chatgpt-google.html

3 Likes

So many “ifs” in the above narrative. It become meaningless.

Here is reality.

Here is Elon’s response. This is what progress looks like.

That’s part of reality.

It’s exactly consistent with what we were talking about. Breadth vs. consistency. AI in general is improving by leaps and bounds - but it is improving in a certain way. When it gets it right, it’s getting it right a lot better. It’s more impressive, more fluid, faster, better written, etc. BUT - and this is the big thing - it hasn’t shown any real improvement in cutting down on the rate of mistakes. If it worked 99% of the time before, it’s even more impressive the 99% of the time it works - but it still fails 1% of the time.

This is why humans end up getting into a lot of trouble with AI. Humans don’t work this way. We would expect a junior lawyer (to use my field) to both make some basic mistakes and have less polished work product on the stuff he did right. As they learn and develop, they quality of their work product gets much better - and they make fewer basic mistakes. That’s how human improvement in skills usually works. So if we observe someone doing the thing they do amazingly well, our heuristic is to assume that they have the same level of consistency of someone who is capable of that kind of work.

But AI isn’t like that. If you have an AI driver that’s both clunky when it drives and making lots of mistakes (or disengaging), and then compare it to an AI driver that “feels a lot more like a person was driving it,” the natural thing to do is to infer that the second AI driver is going to make fewer mistakes (or have a lower rate of disengagement). It’s a better driver! Better drivers make fewer mistakes! But AI isn’t like that. It’s not only possible, but it seems the dominant tendency right now is that AI gets better and better at doing what it does while still making the same amount of mistakes.

That’s why people are getting into massive trouble using AI improperly. It’s why I keep having to attend mandatory sessions on the limits of AI. When AI was kind of clunky, people naturally intuited that you had to check its work (but even then there were exceptions). Now that its output is really really polished, people are getting fooled into thinking it’s reliable. And it’s not. AI breaks the connection between quality of output and reliability of output - it produces output that looks like it was done by a seasoned pro, while still making the same types of mistakes as an inattentive rookie.

IMHO, for a robotaxi service or for autonomous driving more generally, it’s the mistakes that matter. It will be irrelevant whether FSD drives more smoothly than Waymo, if it makes mistakes/disengages at too high a rate to remove the employee from the car. Looking at how the car operates when it’s driving well is just looking at the wrong thing - what matters is whether it’s going to hurt or kill people when you pull the employee out. Services that hurt or kill people end up getting destroyed in the market (RIP Cruise, at least until their upcoming comeback). It doesn’t matter whether the FSD feels alive if it disengages once every few thousand miles or less.

This is factually wrong. You are just making stuff up.

1 Like

I’m not making stuff up - it’s literally what the NYT analysis I posted upthread pointed out. AI is getting better, faster, more polished, improving its output - and it’s hallucinating even more than it was before. This is what our AI consultants have also been telling us - AI is getting vastly more professional at generating legal work product, but it is still just as prone to hallucination and other malpractice-level errors as it was a year or two ago. We’ve been told in no uncertain terms to never rely on AI-generated work product as anything but a working draft, no matter how good it might look.

These types of end-to-end trained neural net AI systems don’t work the same way human brains do. Humans typically don’t vastly improve their work product without also improving their reliability. AI’s can, and do. So if all you do is look at how incredible AI is at doing a particular task when it’s getting it right, you can form a really incorrect impression of how reliable it is.

I don’t disagree that there’s “a decent chance” Waymo is the winner, or at least a winner.

I do, however, think that winning means more than taxis in just 100 cities. Even Uber runs in over 10,000 locations today. With lower costs and rider preferences, the robotaxi market is even bigger IMO. And even bigger than that is autonomy for personally owned vehicles. Who buys a new car with a stick shift today? Even Indy cars went first to paddle shifters, not needing a clutch, and today use EAS (Electronic Activator Shift) that doesn’t even need pneumatics. In short order (say a decade or so) buying a car that doesn’t drive itself will be like buying a car with a stick shifter today.

The white paper makes it clear what better means and how it’s revolutionized AI capabilities.

Human-centric approaches aren’t 100%, either. You can try to deny the advances made, but everyone serious in the AI world has learned the lesson.

You sound like people in the 1920’s saying to get a horse because cars don’t reliably start, especially in cold weather, don’t do well in the snow, and heck, there aren’t even enough gas stations to drive everywhere in the US without carrying extra fuel in cans.

Yeah, but I bet Uber gets a very large proportion of its earnings from a very small subset of those 10,000 locations. And since Uber has almost non-existent capital or personnel costs to service a new location, they have no reason not to cover any location where a driver’s willing to download the app. I’m pretty certain that unlike Uber, driverless cars will need company personnel physically to be in the same service area as the cabs, to do the things that Uber delegates (independently contracts) down to the drivers today.

With 100 cities, you’ve covered 3/4 of the population, and I bet almost every metro where there’s enough business to warrant having an employee there. I think you’re getting nearly all the market with the first 100 cities. And if it’s 150 instead of 100, that still doesn’t go beyond the size capabilities of a Waymo model.

I’m not denying that advances have been made. What I’m denying is that there’s been much progress towards very high levels of reliability. AI is vastly better at generating the first draft of a complicated legal document than it was a few years ago. But it’s no better at being reliable enough to generate anything that can go out without being double checked by a human.

Not at all. All new products have pros and cons, and simply because a proposed product has some “cons” doesn’t mean it shouldn’t be bought and can’t be successful.

But that’s not what we’re talking about here. If Tesla doesn’t get the error/disengagement rate down far enough, then the car can’t operate without an employee in the vehicle. And obviously it doesn’t present any real advantage over “ordinary” Uber if that’s the case. We’re not talking about “the car doesn’t reliably start” versus a horse - we’re talking about “the car can kill you and other people if it doesn’t have a driver, either behind the wheel or in the passenger seat.” You can sell lots of cars in the 1920’s even if they’re not 100% reliable to start in cold weather; but you’re not going to get many takers on a robotaxi service in 2025 if it’s materially more likely to kill the occupants or other people once you pull the driver out.

Not if they lean on the subscription model. There’s a whole world out there that is price sensitive (and another which charges everything and pays the minimum, but I digress.)

I’m sure you’re right, but only to a degree, and it will depend on what the additional costs are: whether it’s a “one time upfront” or “continuing payment” model, which will impact acceptance. (Intercst let the FSD free trial lapse; now that he needs it, he’s back on it.) That’s why I’m skeptical that we will have cars talking to each other any time soon, you need almost 100% compliance for it to do much good; the random airplane who is not squawking his location is a very big threat around airports, to use an analogy.

[Which brings me to one of my favorite examples of government costs being public and complaint ridden, while offer private benefits that few contemplate. “Visible costs, invisible benefits”. Readers are dismissed, I’m going sideways:

The humble stop sign was enough until it wasn’t, and was replaced by a traffic light on a timer. That cost more, of course, and taxes had to pay for the new signal. Eventually they figured out that they could trigger a stop light with pressure pads underneath at high traffic intersections that are little used in one direction , so they replaced the old traffic lights at taxpayer expense. That cost money, but it also sped up your wait coming out of an intersection - sometimes. Nobody says “Hey, I got through the light 30 seconds faster today” but you did save that bit of time and frustration.

Later on they figured out to use electric loops in the road to measure impedance so smaller cars (which didn’t trigger the push plates) also sent the signal, and eventually they replaced some of those with motion sensing apparatus which helped traffic move even better. Each of those changes cost visible money, the benefits were invisible and uncontemplated, but there they were.

If we get cars “talking to everybody (including traffic signals)” there could come a day when your car says to the traffic light “Hey, I’m approaching. If there are no cars coming the other way, how about giving me a green?” And it could happen. It will cost tax dollars to do so of course, even if you paid extra for your “talking car” since the signal apparatus will have to be upgraded (again), but your ride will be smoother and faster, roads will handle more cars, and everybody will be happier. Well except people who think all taxes are theft but they will drive and enjoy the benefits as well.

Trivial example, I know, but now extend it to upgrades at airports, interstates, railroad crossings, financial rules, FDIC, car exhausts, sewage treatment, NASA inspired inventions, basic research at universities, and so much more and maybe it’s not so trivial at all. And woe unto whoever suggests mandating “communicating cars” for the general good. That’s the sort of thing we cannot abide.

OK, class dismissed. I have work to do. You too, probably.]

1 Like

I agree with that, but that’s not relevant to autonomy. I don’t think anybody is still seriously expecting autonomy to be based on any V2V technology.

Interestingly, it used to be that automatic transmissions were an upcharge. Then in the 1990s it switched and automatic transmissions were standard, with an upcharge for manual transmissions. Today it’s hard to find a car with a manual transmission even available for sale.

Once the safety benefits are factored in (eg insurance), it’ll be cheaper to have an autonomous vehicle, and like I said earlier, there will be areas prohibiting manually driven vehicles as well.

What this has to do with V2V is beyond me.

I couldn’t find any history supporting that, but did find this:

In 1915, Cleveland received an electric traffic signal. Detroit, the center of the automobile industry, is credited with installing the first proper stop sign that same year.

So both have always co-existed I guess.

What you’re denying is that there will be progress towards that.

That’s a big unlikely “if.”

But, if you think that’s probable, you should probably short the stock.

BTW, automobile deaths were a big problem in the 1920s, but that didn’t seem to stop Model Ts from selling like hotcakes.

Not that there ever will be. Just that I don’t expect there to be a ton of progress in the very short term, and probably not using Tesla’s current approach.

After all, Waymo has built a system that can operate with sufficient levels of reliability that it can be run without an employee in the car. It’s clearly possible. But Waymo has a different software structure and different inputs (incl. Lidar and mapping), and apparently has built a set of safety protocols to kick in rather than ever have to cede control to a driver. Given that Tesla hasn’t yet done the same, I think it raises the likelihood that Waymo’s ability to do it stems from something particular to their system: modular systems give a lot more visibility into what’s going on than an end-to-end system, and more opportunities to “tweak” the system in edge cases, for example.

I don’t think it’s unlikely at all. Tesla has made massive improvements to how well the car drives when it’s driving. It’s been a minute since I last rode in a Tesla with FSD, but it’s pretty amazing how well it works.

But there’s been no indicia that there’s been a corresponding increase in reducing the error rate. Absence of evidence is not evidence of absence, but you would expect to see either Tesla trumpeting a reduced error rate (like they herald other aspects of improved performance), or for crowdsourced monitoring of error rates to show material improvement, or for Tesla to start submitting intervention/disengagement data to regulators in order to move towards permits…..something.

And it’s what we see in nearly all of the other big, rapidly advancing areas of AI. Most of these systems are seeing exponential increases in performance without similar-sized improvements in reliability. They can do more things and do them better, but they still hallucinate at about the same rate.

No way. Far too many people like ubn and dividends20 out there who will support the stock price based on their personal interactions with FSD, convinced that if the driver is that good when they use it, that means Tesla must be close to full autonomy. Even if I knew for certain that Tesla was more than 5 years away from “solving” self-driving, I wouldn’t know for certain that shorting the stock would be the right move.

Because people wanted to make the trade-off. They were willing to trade off higher car deaths in exchange for the massive improvement in transportation that a car brought.

I don’t think customers or regulators are going to be willing to do that for Tesla - or AV’s in general. If they’re not safer than humans, they won’t see widespread adoption in the U.S. regardless of how convenient it would be to not have to drive yourself.

Which should be a fairly rare event with all the training data.

How rare?

Rare enough to result in about (or less than) 1 accident per million unsupervised miles?

1 Like

Your written explanation is confusing to me.

You write “attack the creation of AI software that attempts to mimic how humans understand and process data.”

This is confusing because a core approach of AI is neural networks, which are modeled on human neurons networked together like in the brain, which are certainly a part of how humans understand and process data.

In fact, neural networks themselves are fundamentally modular, with individual neurons (represented with mathematical formulas, ie, functions) combined into groupings in all kinds of ways, with all kinds of special, technical names.

Sutton laid it out pretty well, I thought. Which part of what I said doesn’t match Sutton’s blog?

Understand that modular includes having separate handlers for different types of sensors leading potentially to different views of reality like the Cruise case with the pedestrian.

Me! If I could find one!

It can include that.

It can be different from that.

It can be any degree of modularity.

There is no 1 unique modular architecture.

And for this:

different types of sensors leading potentially to different views of reality like the Cruise case with the pedestrian.

I think I have addressed this at lease once, maybe more, maybe even in this thread.

goofy has addressed it I think > 1 time

Let me emphasize, again:

Machine learning can directly evaluate models with different modular architectures and identify the best architecture to perform the driving task.

This choice of model architecture is exactly the kind of problem machine learning is designed to solve.