Waymo’s co-ceo presents on AI driving

Stocks A to Z
,
61

Again, I never made an investment as a shareholder relying on this thesis or did it enter my calculus as a consumer. I can only speak to my own reasoning.

I would agree that there has been a shift from Tesla as a dominant participant in personal EV sales to the cybercab as an important focus, one of many.

I think if autonomy is solved, Tesla goes from pitiful small levels to very high rates of adoption, potentially.

I have no idea of what the economics of robotaxi look like for the company. I can tell you that I have little regard for where the P/E stands for certain investments. Particularly for large winners in my portfolio where the potential addressable markets of the business are very large and important, I tend to simply hold through, absent some fundamental change in the underlying thesis, which could happen with Tesla.

I do have other investments where valuation is an important part of when to buy or add. But in almost all cases, I am far more interested in the the importance of the market in terms of potential size, the moat of the business or its perceived advantages, the quality of management and if they are a stakeholder. I have lost far more money exiting early than holding until the outcome is more certain.

Would I be cautious about starting a new position in Tesla right now at this price? I probably would But, Tesla is substantial winner in my portfolio I follow Peter Lynch and David Gardner’s advice with respect to those positions. Valuation alone is rarely a good reason to sell and the default position should be to water the flowers or as David would say, “let them run big”. The truth is with almost every great long term business. You could have bought them at interim all time highs, where they seemed wildly overvalued, and still outperformed the market over the long haul.

62

I mean - maybe? Tesla’s spent a lot of money to develop FSD. They’ve gotten whatever they’ve received from sales/subscriptions. Some manufacturers have developed more advanced ADAS systems (God’s Eye, Super Cruise) - many have not. If the value proposition isn’t there for a very costly to develop ADAS that isn’t self-driving, you won’t see too many manufacturers doing it.

They were definitely saying that you can get to an actual driver much faster and with less effort if you go multi-sensor rather than single-sensor, and that you had a lower ceiling with the single sensor. IOW, the end state of the multi-sensor driver will be much much safer and effective than the single-sensor driver. You can get to very good ADAS more quickly with vision-only, but then trying to make the last jump from ADAS to AI driver takes a lot longer if you didn’t integrate multiple sensors on your way there.

63

The most successful investment of my entire career has been AOL. Should I have let that run? Or should I have sold it, suffered a massive tax bill, and moved on?

Valuation was part of it, the market’s “toppiness” was part of it, the entrance of cable internet was part of it (I also owned @Home, which was wonderful while it was wonderful, and terrible when it became terrible).

Thing is, things changed. I believe things have changed in the EV market. I also believe things have changed in the Robotaxi market. I also believe things have changed in the self-driving market, though not as dramatically as the first two. I am sure things have changed in the evaluation of Tesla as a car company (just as they changed for AOL in the “on line” market), but perhaps Musk can keep the plates spinning for a while as he promises AI or whatever next comes along. Steve Case and Bob Pittman didn’t, and people thought they were pretty sharp too, at least at their height. (Pittman arguably more, as he went from success in radio to AOL to inventing MTV to real estate mogul back to broadcast. Still didn’t stop the collapse of his earlier successes.)

64

My approach is not a guard against giving back huge gains. We each have to approach investing in a way that best serves our mental game. For every example you can give me, I can give you just as many others where the valuations were declared outrageous and questions raised about the business, just as the stocks went on to make even more enormous runs. I suspect that your timing on AOL that worked out well, but probably includes a number of investments where you pulled the trigger too soon, either underestimating the long term runway or the macro environment.

As Mark Twain said, history doesn’t always repeat but it does Rhyme. It’s entirely possible that Tesla suffers the fate of AOL. There is no doubt the underlying thesis is shifting in terms of the future source of revenues. But, I believe in being slow to act and wait for more certainty, even if it means selling at lower a price or even a loss.

In the end, if you don’t let a 5 bagger become a 10 bagger, it never gets to become a 100 bagger and so on and so on. The fact that some or even many of them will give back those gains, is an acceptable risk for me. The investments that have done well have far outpaced the losers. I wish I had the ability to see the thesis deteriorating in real time to separate intermediate winners from the truly long term ones. In a few limited circumstances, I have but for the most part I simply let it play out. But, we each have to play our own game.

65

This may be an argument in FAVOR of the way Tesla does it. All these examples are extremely complex, take years to deploy, and come with massive amounts of regulation, both up-front in the form of testing, and ongoing to ensure that all the complex systems continue to work properly with each other. And, of course, as a result, they are extremely expensive.

These conversations about theory and about the future are interesting, but when it comes down to the current reality, all I care about is “What is available out there today for me to buy?” I have no “brand loyalty” anymore, and I will test out anything, and if the new company does a better job at the features that I want, I will buy it. So right now, I am willing to test any car that has the feature of entering an address (or tapping on a saved location, etc) and pressing a button called “GO” and have the car bring me there. Parking alone is optional because I usually like to pick my own spot, if the device can learn where I like to park, that would be a big plus. I want a car that has cooled seats because it becomes brutally hot here in the summer (6+ months a year). And I want a car that doesn’t have keys or fobs or anything like that. This has become a critical feature because we constantly share cars in our family, at any time, any one of me, or my wife, or my 5 kids, or my 2 sons-in-law may be driving the cars. I want the ability for all 9 of these people to simply walk up to the cars and drive them when necessary. This also enables another nice feature - I can easily park one of the cars at the airport and then when someone arrives, they can simply walk to the car, load their bags, and drive away. No muss, no fuss. That is something I want. And one of my cars, entirely on it’s own, saved my wife and sons life (or at least prevented severe injury) yesterday. They were driving on a local road, stopped at a red light, and then after the light turned green, they began to cross the intersection. Suddenly a car moving quickly perpendicularly went straight through the intersection on red. The car slammed on the brakes just in time so the car that was about to hit them passed within inches of their car. I also want that feature. I have the whole incident on video, that’s another good feature, that the car stores video from all (most?) of the cameras when necessary.

So whenever I see publicity/ads about cars with the features I want, especially those that include all the “must have” features, I will definitely test drive them in preparation for the next vehicle purchase. Any suggestions what to test drive next?

(I have a whole list of “must have” requirements, both about the vehicle and about the company. For example, I will only buy from a company that has a very high likelihood of remaining solvent over say the next 5 years.)

1 Like
66

The airliner stuff is algorithmic, pre AI, making the comparison with Tsla FSD moot

GoogleAI:

You are entirely correct that commercial aviation autopilot systems rely on deterministic, rule-based algorithms, whereas Tesla Full Self-Driving (FSD) utilizes a probabilistic, deep neural network trained on video data. [1]

Comparing them as equivalent technologies is a flawed analogy because they operate on fundamentally different engineering principles, face entirely separate operational environments, and solve different problems.[1]

Comparing Autopilot vs. Tesla FSD

Feature Commercial Airliner Autopilot Tesla Full Self-Driving (FSD)
Core Technology Deterministic, rule-based algorithms Neural networks and deep learning
Primary Inputs Instruments, radar, radio beacons, GPS High-resolution cameras (vision-only)
Decision Making Strict if/then logic and control loops Probabilistic pattern matching
Environment Highly structured, regulated airspace Unstructured, unpredictable city streets
Edge Cases Rare weather events or system failures Constant erratic pedestrians, construction, debris

Why the Comparison Fails

1. Deterministic vs. Probabilistic Logic

  • Aviation Autopilot: Operates on exact mathematical models (like PID control loops). If the aircraft drifts 2 degrees off course, the system applies a mathematically calculated correction. The code is entirely auditable, predictable, and verifiable by regulators like the FAA.
  • Tesla FSD: Relies on an “end-to-end” neural network. It processes visual pixels and outputs steering and acceleration commands. It does not use traditional “if/then” code lines. This makes it a “black box” where developers cannot easily trace exactly why a specific network weight caused a specific maneuver. [1, 2, 3, 4]

2. Environmental Complexity

  • The Sky: Contains no traffic lights, pedestrians, hidden driveways, or unmapped construction zones. Air traffic control strictly sequences distances between aircraft.
  • The Road: Requires real-time negotiation of chaotic human behavior, poorly painted lines, non-standard signs, and sudden physical obstructions.

3. Redundancy Architecture

  • Airliners: Achieve safety through hardware redundancy. They use multiple identical algorithmic computers that “vote” on actions. If one computer disagrees due to a hardware fault, it is overridden.
  • Tesla: Achieves safety through continuous software iteration. It relies on the human driver as the primary fallback mechanism (Level 2 automation). [1, 2]

The Economic and Financial Context

From an investment perspective, conflating these two systems distorts valuation models for autonomous driving technology:

  • SaaS vs. Hardware Margins: Traditional aviation automation is baked into the capital expense of an aircraft. FSD is structured as a high-margin software-as-a-service (SaaS) subscription model designed to scale across millions of consumer vehicles.
  • Regulatory Pathways: Aviation software follows strict DO-178C compliance certification, which takes years to approve changes. Tesla utilizes over-the-air (OTA) updates to continuously deploy code changes directly to consumer fleets, shifting the regulatory paradigm. [1, 2, 3]

The Captain

67

Do your requirements include

  • unsupervised autonomy
  • the manufacturer assumes liability for driving
    ?

or are you happy to supervise and assume liability?

68

As I have explained before, I used those examples only because they are well known, and in response to the argument that “You can’t have multiple sensors because it’s too complicated to have them decide which is right.”

Apparently I should have said “Your Wii game controller has multiple sensors, some redundant, which do the same thing.”

Your iPhone has multiple thermistor temperature sensors distributed across logic boards, CPUs, the battery, the charging port, the camera and elsewhere which provide thermal readings which the phone reads simultaneously and then uses to make macro decisions, including whether to “shut down” if it’s too hot. But it can also change the charging speed, or make other internal changes depending.

If it’s not “too complicated” for the phone in your pocket which costs $1,000, then it’s not “too complicated” for a self-driving automobile costing many multiples of that.

(For the record, you’ll find layered systems doing this same sort of “smart input balancing” in GPS units, gyroscopes, fingerprint scanners, facial recognition GPUs, and lots of other places. The entire argument is absurd.

Musk did it because 1) history; LiDAR used to be vastly expensive and he did not foresee that the price would crater. 2) Ego; it’s hard now, 10 years later, to say “oops”, and 3) legal peril; a promise was made to buyers that they would not need to upgrade, and quite obviously they do if “vision only” can’t be made to work.

As for those who say he’s not afraid to say “oops,” well, I can only point out that the Boring Company still exists, and is (literally) going nowhere.

1 Like
69

I think all three points you have made are on target.

However, I find this #2 point about “ego” is one that I started wondering about a few years ago if this was part of the problem with FSD and now the struggling we see with Tesla Robotaxi. I have never posted this concern previously but only because it’s near impossible to substantiate. But is it a very real problem right now?

Could it be one of the bigger problems? Hard to say - we don’t know what’s really going on in Elon’s head but it just seems incomprehensible about how stubborn he is. He just seems to be willing to keep wasting lots of shareholder $$$ rather than use common sense as to why FSD has failed to achieve expected safety levels.

Another one I don’t understand is his hope that end to end AI will succeed. Another complete unknown but also questionable premise. Can billions and billions of miles of data EVER offset the dangers of undiscovered edge cases? Considering the probabilistic nature of AI, the long tail distribution of edge cases just seems too long for AI at this time. And unfortunately increasing AI computing power doesn’t seem to be an answer as the silicon brains just can’t seem to come close to what the human brain can do when it comes to interpreting and reasoning through such unknowns.

2 Likes
70

I suspect not, because there are always a bazillion more possibilities to be included, and as some point the “edge case file” will be larger than the “how to drive” file. (Making a point, not saying that’s how the memory is actually stored.)

In a parallel thread I linked to an article about Waymo’s increasing frustration with edge cases. Here: Another one on “Edge Cases”. Waymo, this time

And one of the examples given is a couple in the back seat where the car suddenly takes off into a construction lane, marked with orange cones, except they’re on the wrong side of the cones (weaving back and forth thru them, actually) accelerating to full speed in the closed lane, and with the passengers terrified and unable to stop the car from doing whatever it wanted. No kill switch, no control whatsoever.

If that happened to you, would you ever get in anybody’s Robotaxi again? The slogan “safer than a human driver” would be of little comfort knowing that at any moment the chips and electrons could conspire to do whatever they wanted, no other input of any sort allowed.

It took a while for people to get comfortable with elevators at all, and when “automatic elevators” came about in the 1920’s it took many years before people became comfortable without having a human operator for those, as well. (And those ride a predictable, never changing path!)

I’m sure acceptance is going to be different based on age group and so on, but I see lots of hurdles ahead. (And I’ll be that couple in the back seat of the errant Waymo was glad that it had Lidar before it ran into a construction truck.)

1 Like
71

I’m not even sure it’s edge cases that are the problem, it may be just normal stuff like how to handle a left turn controlled by a traffic light. I’ve watched a lot of robotaxi videos on Youtube and robotaxis consistently have problems with left turns controlled by a traffic light.

Several months ago in Austin there was a problem with robotaxis running left turns against a red left arrow. Just recently, I watched a robotaxi video in Florida and the robotaxi didn’t make a left turn when it had a left green arrow. The robotaxi just sat there until the green left arrow shifted to red. It did finally go once the green left arrow came back again.

So, I think the training of even basic cases like red and green left arrows is difficult for an E2E black box AI system.

2 Likes
72

No but the argument is moot. Perfection will never be achieved, the limit is where you accept the consequences.

GoogleAI:

Accepting the residual risk of undiscovered edge cases is the actual operational threshold for deploying complex automated systems, rather than waiting for an impossible state of zero errors.

The Reality of Edge Cases

  • The Long Tail: Rare, unpredictable events (the “long tail”) cannot be fully cataloged by any amount of driving data or simulation.
  • Comparative Safety: Systems are judged against human error baselines (such as traffic fatality rates) rather than absolute perfection.
  • Regulatory Limits: Legal frameworks and liability standards define the exact point where society agrees that a system is “safe enough” to operate in public spaces. [1, 2]

The Captain

73

I agree absolute perfection has never been the goal; however, that doesn’t mean Tesla has an acceptable excuse for their status quo of acting like their main issue is more billions of miles of training data.

At some point will Elon capitulate on his sensor and AI decisions that continue to limit scaling?

Another interesting dilemma is how the Cybercab is compromised. My understanding is that the Robotaxi driver software can’t deal with the smaller chassis size??? So the goal is “drive anywhere, any time” but apparently that is only if you have the right chassis size?

1 Like
74

I think that overstates what’s been discussed publicly. Tesla has said that the software has to be calibrated to the different chassis size, not that it can’t handle the smaller chassis size. So we don’t know yet whether it will be simple or complicated - much less whether it won’t be possible - for Tesla to make that shift.

1 Like
75

As an admittedly failed engineering student, I have to say this is weird. Calibration to a different chassis size should be fairly trivial; it’s not like it’s an airliner with differing weight and balance characteristics flying through the air. And if it is not trivial, what does that say for porting the idea to various manufacturers, all with different body length, height, width, weight, wheel base, and wheel track (thank you Miss Vito), not to mention positraction and independent rear suspension? And there are several firms trying to do just that.

As I say maybe that’s a real thing, it just seems that it shouldn’t take very long at all given the engineering prowess the company has already displayed.

76

I believe Waymo will field test its new 6th generation hardware, Ojai, about 6 - 9 months before it enters paid commercial service.

Of course a lot of development and testing would go into building a new platform, before field testing, for hardware such as Ojai, with a reduced sensor suite and different form factor.

77

I can think of a few. Chrome, which wounded Internet Explorer to the point where Microsoft Edge is built on the Chromium operating system, Gmail, Android, Maps, Drive, Photos, and Workspace. Each of those has over a billion users. Not to mention Google Cloud. And of course Google invented Transformer Architecture.

And yes, they have made lots of speculative bets, big ones even, on product lines that didn’t work out. Google+ arrived right when social networking started dying, for example. But then some of those bets turn into rocket ships. Alphabet is the most profitable company in the world. And its profits are increasing at a double digit pace AND its margins are increasing as well.

Alphabet is really, really good at making money.

We’ve talked a lot about the economics of TaaS, so no need to revisit it in detail, but by whatever metric you care to use, miles, trips, vehicles in operation, cities, announced planned expansion, operational capabilities, Waymo is orders of magnitude ahead of Tesla robotaxi.

And of course, Waymo intends to licence their tech. That’s high margin, recurring revenue. That’s where the real money is going to be.

2 Likes
78

Who knows? Tesla hasn’t made a lot of statements about this. I was just pointing out that they haven’t said that they’re unable to port the driver software into the new chassis, just that they have to do some calibration. We don’t know if it is difficult or trivial yet.

If you’re cynical about it, one nice thing about having to “calibrate” your system is that it’s a useful explanation for why cybercabs haven’t been deployed anywhere yet, despite having entered continuous production back in April. It’s four months later, and Tesla has produced more Cybercabs than it has autonomous vehicles in its fleet. Probably several times over that amount. It’s pretty reasonable to ask, “why aren’t they being used anywhere yet?” Calibration is a good answer. Well, it’s an answer - given the massive capital costs of the production line and the not-massive-but-still-large capital costs of having several hundreds of cars just sitting around unused, one could reasonably ask why the “calibration” wasn’t completed earlier in the ramp up.

79

They are fantastic at making money. They also spend a lot of capital in areas that have been wholly unproductive in terms of ROIC. My point is that Waymo may not be approaching autonomous driving with the endgame of profitability in the same fashion that Tesla may be doing so in terms of decision making. Tesla may also not be in a position to spend capital in the way that Google/Waymo to rollout the service at great expense to the company.

I don’t know what the economics business will look like for Waymo long term or for Tesla. But, if Tesla solves autonomy, with a simpler and more cost effective approach and with manufacturing at scale already in place, the pace at which it then goes national, IMO, may close quickly whatever gap there may be between the two companies.

But, I fully acknowledge these are all guesses at this point and if we want to use the current metrics, Waymo is ahead. But so was blackberry at one point and then all of the sudden, it wasn’t. Compaq was the leading provider of personal computers until Dell took the spot. It will be fascinating to see this latest heavy weight battle play out.
It’s also possible that Tesla never fully solves autonomy but the edge cases are so rare, it becomes the default standard for supervised FSD and also licenses that version for use by other manufacturers.

80

Keep in mind that there’s a lot of important detail in what one means by “solving autonomy.”

It’s very likely that we can ‘solve autonomy’ in the sense that you can have fully autonomous vehicles travelling the public roads if they are supported by an established and robust network of local employees. A network of employees who can attend to the vehicles IRL, do some remote guidance if they get into trouble, and interface with local LEO and emergency responders. That type of “solving autonomy” is kind of what we have now in a few places. And if that type of autonomy ends up being relatively cost-competitive with taxis, you could get fairly thorough market penetration with just that. It wouldn’t necessarily scale up nationally fast, though. Just at a Waymo pace, because the bottleneck is the robust local network, not the cars. And that type of autonomy might end up only with TaaS cars that are more expensive than privately owned cars, and so no self-owned autonomous vehicles even as they are ubiquitous in the TaaS space.

A second possibility is that you can end up with autonomy that doesn’t require much of a robust local network, and thus is theoretically capable of being self-owned, but is significantly geo-locked - either operationally or through regulation. That’s a bit of a market question, but it might not be viable to sell an autonomous vehicle that can only be used in some places but not others within a market. Certainly some people would buy it, but a lot of folks would be reluctant to buy a car that can only go to X% of the addresses within an area.

And the third possibility is full-on Level 5 autonomy, which is what I think you mean by “solving autonomy.”

1 Like