Waymo’s co-ceo presents on AI driving

Waymo’s co-ceo gives a long presentation, with lots of good information, including discussion on multi-sensor versus single sensor (camera) and vanilla versus structured end-to-end models (and the exciting Bitter Lesson).

Waymo talk

Highlights:

  1. Multi-sensor may be more likely to achieve higher safety than single sensor.
  2. Single sensor may never achieve super-human safety (or take a very long time to achieve it).
  3. Naive, pure black-box neural network models have weaknesses relative to models with structure informed by human knowledge of subjects like physics, traffic rules and human behavior (even when considering the bitter lesson).

Excerpts

the information from all of those sensors gets fused into a single view of the world around us that is much more precise and generally vastly superior to what you get with any one sensor.

Redundancy, of course, does not necessarily mean multiple sensing modalities, but if you need redundancy anyway, you might as well benefit from the complementary physics of the different sensing modalities

the Waymo Foundation model is a multimodal world action language model. It’s kind of a mouthful, so let me unpack the ingredients. It’s a multimodal model because it is able to process these multimodal sensor inputs: cameras, LiDARs, and radar. It’s a world model because it inherently understands how the world works—the physics, the dynamics, as well as the social and semantic aspect of it. It’s an action model because we are not just passively observing how the world evolves; we’re an active participant.

It’s an end-to-end model, which has a couple of nice properties. It allows us to effectively backpropagate the gradient from the task that we actually care about all the way to the early layers of the model. It allows the encoder to learn the right rich representations for what the generative part needs to solve the task.

a well-known lesson in the AI community that general methods that leverage massive compute and massive data will always beat methods that rely on handcrafted, engineered human knowledge. That’s the so-called bitter lesson that Richard Sutton published and formulated in 2019

one nuanced area where this lesson shows up is the use of structure in your models. Depending on how you use your structure, you can end up on either side of the bitter lesson. Essentially, structure that fights scale will always lose, and structure that channels scale always wins.

But if you need to reach superhuman levels of performance in a fully autonomous agent in a safety-critical environment, just doing that basic vanilla end-to-end is not enough.

But in the physical world, there does exist structure. You have laws of physics, you have rules of the road, you have objects that behave in reasonably predictable ways. You can use that structure in addition to the learned representations to boost your performance, simplify validation, and at the end of the day, just get better scaling laws.

If all you have is a black box end-to-end system, you are forced to do all of your evaluation and all of your training in the end-to-end setup, all the way from sensors to decisions to actuation. Having that intermediate structured representation allows you to mix and match. You can do some training at larger scale and some evaluation in the space of those compact structured representations, and some in the full space of end-to-end from sensors to decisions. Finally, we get strong, verifiable feedback signals for both evaluation and for training, training recipes to support things like reinforcement learning.

To build and safely deploy an agent in the physical world, it is absolutely critical to have a good large-scale, realistic, high-fidelity simulator.

Some related points made in various discussions:

3 Likes

Thanks for the nice overview. So it sounds like in addition to sensor deficiency, Tesla is also using an inferior E2E AI model?

Plus, Tesla also seems to be deficient in focused, simulated training miles. The last I read, Waymo had 15 billion simulated training miles. According to a Google search, Tesla has 7 billion combined real-world and simulated miles with both FSD and robotaxi.

It sure seems like Waymo’s AI driving is superior in every way to Tesla’s.

1 Like

I found the below slide, reproduced in Electrek’s article on the talk, to be very instructive on how Waymo views the debate over sensor technology.

Essentially, they believe that using just cameras can let you develop a really good ADAS system much more quickly than if you try to build using lots of different sensors (like high-def radar, ultrasonic, LIDAR). It takes a lot longer to get the multi-modal system built properly. However, even though the camera-only system can get built faster, it has a lower ceiling. You may not be able to get it far enough down the “march of 9’s” where it can make the shift from being a really good ADAS to being a Level 5 or very broad-range Level 4 system. Because to be able to have an AI driver that can be deployed broadly, it’s going to have to be materially better than a human driver.

No guarantee they’re correct, of course. A lot of very smart people at Tesla are trying to prove them wrong. But it is interesting to see that perspective.

Waymo CEO explains why Tesla’s camera-only self-driving falls short | Electrek

3 Likes

In Spanish we say, “Cada quesero alaba su queso.” Every cheesemaker praises his own cheese.

Tesla used LIDAR early on and decided that it was not worth the trouble for production products, that cameras by themselves were good enough. Time will tell. Unlike many others, Elon Musk focuses on cash flow which is most likely the reason Tesla scrapped LIDAR. I like the focus on cash flow for my investment dollars.

Did Waymo’s co-ceo talk about the impact of LIDAR on Waymo’s cash flow?

The Captain

2 Likes

Yes, but Tesla probably got hoist on their own hubris there. Early on, LIDAR rigs cost around $70K. If you’re looking at things in the short term, then that’s not worth the trouble for products - it’s too expensive. And if you believe that self-driving is something that can be solved in the next year or two (and certainly by 2019), that decision makes sense.

But here we are, a decade later, and self-driving is still unsolved. Meanwhile, the cost of LIDAR has fallen 99%. Those $70K rigs can now be had for $200, they’re standard equipment with no upcharge in a host of Chinese models, and the impact on cash flow is going to be minimal compared to the impact of operational ability of the AI model. Waymo uses LIDAR, and that certainly adds several hundreds of dollars to the production costs of their vehicle - but they have a model that allowed them to run 10x the cars without a safety driver that Tesla can. If you’re looking at cash flow, that’s probably a point in Waymo’s favor, at least so far…

6 Likes

The decision not to go with LIDAR was more involved by Tesla. Musk and Team believed it would complicate decision making longer term. When Lidar and the cameras conflict in terms of decision making, which system should take precedence? The goal was in part simplification and the decision to go with one vs both avoids that conflict. Will it work out longer term? I don’t know but either approach has advantages and some logic to it.

The other thing that is missed is that the camera system allowed Tesla to cost effectively introduce supervised FSD and begin collecting driving data which it uses in development of its physical AI modeling, which provides exponentially more data than what WAYMO is getting, even though it has more purely autonomous vehicles deployed in the wild. It’s literally billions of miles data from the consumer fleet and globally vs tens of millions. Again, it really doesn’t matter until somebody totally solves the problem and then at a cost that makes it financially worthwhile.

2 Likes

Yep. Waymo’s position is that if you go with vision-only, you don’t have to put in the work and effort to solve the decision-making inputs for multiple inputs. That lets you go faster, and you get to a really good ADAS more quickly. But because you’ve taken the shortcut, you can’t get down the March of Nines to an actual AI driver that can scale.

Yeah, nothing would have prevented them from doing that and implementing LIDAR. You could have collected all that data using the cameras on the Tesla for the ADAS, and also developed your AI driver to work with both that visual data and LIDAR inputs.

Now, Tesla is kind of trapped in their own incumbent technology. The massive innovations have come in LIDAR and high-def radar. The former has declined 99% in cost from when Tesla started out; the latter wasn’t commercially available at all back when Tesla started out. Now they’re both cheap and ubiquitous. But Tesla’s just as married to “vision-only” as Kodak was married to film. Perhaps even moreso, since Tesla has kind of legally bound itself to making a vision-only system work in order to avoid problems with how it marketed their cars’ hardware capabilities.

2 Likes

This is, by far, the worst argument I have heard on this issue.

Airliners have multiple sensors, redundant systems, and programming to decide “what’s best.” Same with nuclear power plants. Also other industrial processes in chemical plants, oil refineries, and health care systems.

I’m most familiar with aerospace systems which use voting arrays to “decide” which sensor makes the most sense - or to vote 2 out of 3 - or to turn control back over to humans. It’s not hard, somebody somewhere decides what’s important and programs it into the logic. Let me ask a question: Tesla has cameras with overlapping views. What does it do when the overlap part has dissimilar information coming in? I suspect it defaults to one or the other - or says “I can’t do this” and stops.

I bet if/when the “personal robot” becomes common it will have to have multiple, simultaneous sensors all working in concert: balance, touch, vision, all deciding “what’s most important now?” and proceeding according to whatever it’s programming has been set up to prioritize.

4 Likes

Airliners and Nuclear Power plants ultimately have human oversite and cost millions if not hundreds of millions to build and acquire. We’re talking about an endgame to completely remove human involvement and at a sub $25k price tag per unit.

.

The “Which One Wins?” Problem: In his public statements on X, CEO Elon Musk explicitly noted that sensor fusion can decrease safety because when data feeds disagree, hardcoding logic to dictate which sensor “wins” introduces severe architectural risk.

According to former Tesla AI Director Andrej Karpathy, parsing entirely different data structures (like cameras’ 2D pixel grids alongside LiDAR’s 3D point clouds) exponentially complicates the machine learning training pipelines and requires massive computing overhead.

2 Likes

Yeah. It’s hard. That was Waymo’s point. The easy way - the shortcut - is to just go with a single sensor. That way you don’t have to solve these problems. You collect a single input with a single data structure, and have a simpler AI model (relatively speaking).

But that limits you, in their view. You get a good ADAS quicker. But then the simplicity that you’ve locked yourself into reaches its limits. You can’t keep improving. You hit a ceiling - or asymptotically approach a ceiling - that’s too low to work for a robust self-driver. Your first 10 billion miles of vision data gets you a good ADAS, but no matter how many vision miles you add after that, your progress is too slow to get to a real self-driving AI.

3 Likes

Tesla’s argument is that it substantially complicates the process, creates additional risks, and dramatically increases the compute power to make it all work.

I’m not an expert but until proven otherwise, I’m going to listen to the people and company that have developed for consumer use the closest thing to full autonomy. But, I recognize the story is not fully written.

By that logic, the harder approach, not the easy way out, is in fact reaching full autonomy by vision only AI development. I’m kind of a simple guy and am a long time owner of both Google and Tesla. I have watched Google throw away cash outside of its core businesses. In fact, one could argue that outside of search, it is the youtube acquisition that represents the only other successful financial offshoot. Google has a long history of vanity projects that were never really mindful of the long term ROI.

Musk has built not 1 but several successful companies with a very disciplined approach to process and manufacturing. The fact that Tesla even survived as a going concern is a remarkable feat in and of itself. I see Waymo as looking at autonomy without consideration to commercial profitability. I see Tesla looking at both. Supervised FSD is way better than good and far better than any other commercially available system. But, the last mile may prove the hardest.

2 Likes

Which would be Waymo. Right? I mean, Waymo’s the one that’s actually offering millions of actual full autonomous trips to consumers. Tesla is the company that a year and a half after launching in Austin still has to keep that service severely limited to a small number of vehicles, with only twenty running unsupervised (about a fifth of Waymo’s).

Again, that’s Waymo’s point. If you design your system so you can quickly get to a really good driver assist system, you can get pretty far much faster. But then when you reach that “last mile,” you can’t get there. It’s beyond the capabilities of your single sensor system. You got to the best commercially available assist system faster than your competitors, but you are now locked into tech that can’t far enough down the march of nines.

Your confidence in Musk is warranted by past performance but completely misplaced in the specifics. Musk hasn’t succeeded in his business ventures by being right in every endeavor. He’s succeeded by placing enough different “bets” so that even though some don’t achieve much (or fail outright), the successes outweigh them. His record supports believing that Musk will succeed, but not that any specific effort or initiative by one of his companies will be successful.

About three years ago, we had similar conversations about whether Tesla would continue to grow at the 50% CAGR that Musk was projecting. Skeptics on this board pointed out that was incredibly unlikely, given Tesla’s place in the market. Supporters claimed that it was going to happen. It obviously didn’t happen and won’t happen, of course - Tesla’s sales have actually shrunk since we had those conversations in 2023, and Tesla’s stopped talking about that. Note that this doesn’t mean that Tesla’s stock didn’t do well - their average share price in 2023 was about $217. It just means that the specific aspect of Tesla’s business we were discussing, their projections to grow unit sales by a 50% CAGR, ended up failing. The same is true of the Cybertruck - the fact that Musk built the first new successful major automaker in the US in decades didn’t translate into that specific thing being successful.

The same may very well be true of FSD. Just because Musk was successful in making the Model 3 and the Model Y happen didn’t mean he could make 50% CAGR happen or the Cybertruck successful. It may turn out that he simply chose the wrong path for FSD.

1 Like

I think this is where the logic of your argument breaks down and you are transition from an opinion to then claiming facts and certainty. We do not know for certain that Tesla cannot solve full autonomy with a vision only system. It may be improbable or even unlikely but the future is not yet known. I would also counter your argument that pivoting at some point is not a possibility. As you said, LIDAR cost is dropping dramatically. Who is to say some modification or accommodation may not be possible at some point down the road.

Further, you continue to make assumptions that the drop in lidar costs and manufacturing will eventually make Waymo’s approach profitable. Again, until the economics of it become reality, nobody knows for sure what the margins look like. I would argue if you solve autonomy but can’t make money at it, I find no utility in it as an investors.

I think this is a complete non sequitur. Not only is failure part and parcel of Musk’s approach, he believes it is critical to Tesla’s long term success and to break through with new technological innovations. Look, I’m not sure what you are arguing about. You may end up being entirely correct. My only point is I understand the logic of why Tesla is approaching FSD in this manner. It is not a guarantee that the approach will be successful. I have put substantial capital, by my definition, in play as a result. But, if Tesla is ultimately doomed to failure, you are shorting it, right? If you’re not, I imagine that at least your wallet is allowing for the possibility that you may be wrong.

I think the whole debate about autonomy and the long term winner is speculation until both companies are closer to the finish line. Aside from that, I like rooting for American innovation and will be pleased if either or both come up with profitable enterprises as a result.

I should have asked this at the beginning. Are you based in the US and have you driven the latest versions of FSD? I find that those in Europe that have experience only with the ADAS there are getting a very different experience. It is by all accounts a far cry from supervised FSD in the US.

1 Like

Sure, I just used those examples because they are well known even among non-technologists. Here are a few which operate on multiple sensors, redundant inputs, and which “decide” based on algorithms programmed ahead of time.

Accelerometers, like those used in game controllers, fitness trackers, airbag deployment, or stress indicators frequently installed on bridges and tall structures. Heck, your phone has some to indicate to itself whether it’s in portrait or landscape mode, although they’re probably not redundant.

Gyroscopes, used in satellites, spacecraft, marine vehicles, and even industrial machines where product guidance is required. You’ll find this in VR headsets, gaming handsets and other consumer devices, as well as in sophisticated areas like drones and rocketry as well as GPS satellites. These can use both inertial components (weights) as well as electrocapacitance to maintain spatial awareness, and some use both systems simultaneously as a check/counter-check on each other.

You may have heard the term piezoresistors (at least I have) which test for orientation, but may also be backstopped by multiple paired conductive plates for the same function, and which check each other for surety. You find them in tire pressure monitors on your car, heart or bloodpressure monitoring systems and elsewhere. Very common.

These types of sensor fusion or sensor backstopping is not unusual, and while it sometimes requires extra cost, often it doesn’t, as the function is taken care of on the same chip - yet somehow they manage to decide “how to decide”.

As I say, the “Oh this makes it too complicated” is a shibboleth. Nonsense. It’s harder, sure. It’s also WAY better if reliability is important to the function. For automated driving, I’d have to say that would be near, if not at the top of the list.

1 Like

Yes, they cover cost.

And Waymo is iterating on their platform, their 6th generation platform is entering commercial service now.

The video is long.

But there is a link to a transcript and there are links to various timestamps of the video corresponding to different topics in the presentation.

With every generation, the hardware not only delivered amazing capability, but we were able to drastically simplify and radically reduce the cost of the hardware as well. So betting your company, betting your approach on today’s hardware prices, is just betting your company on a number that has a fairly short shelf life and is going to expire. Hardware will change. Many components will get commoditized and drop in price. So design for that future and be ready to upgrade.

This question reflects a fundamental lack of understanding of how machine learning (and statistical methods generally) combine multivariate data to arrive at the best inference (highest precision, most accurate).

We have covered it so many times, but people are dead set in their prior beliefs.

Ironic how people can believe in all kinds of black-box AI magic but get stuck on this.

Sad that this flawed idea lives on.

The earth is not flat.

I give up.

Probably if you have crude architecture, but not for teams that are at the cutting edge.

Tesla has no data advantage.

They have the data part, just not the advantage part.

At least not for vehicle autonomy.

Tesla data advantage is another myth.

I’m not sure if its relevant, the redundancy in other non-related sectors. The people developing it and putting the capital to work disagree with your assessment. It’s not my opinion but theirs. Until time reveals if vision only doesn’t work or that Waymo’s approach works but can’t be done profitability, we’re speculating on how this plays out. As the Tax Accountant is famous for saying, my crystal ball is broken. The truth is there are few things in investing that have certainty. I think this debate is one such example.

It’s not my opinion, it’s the statements of those building a vision only system and the rationale for it. It’s amazing to me the number of armchair AI and ML experts here that have not found their way into the employment of anyone building autonomous vehicles.

If vision only and the rationale for it is flawed, what is the motivation for Musk to pursue something that will not work or unnecessarily takes on greater risk of failure? Jensen Huang has been quoted as saying Tesla holds a massive structural advantage in terms of physical AI because of its autonomous data. Is he being duped as well?

Absolutely - and I was really just conveying Waymo’s opinion as expressed at this conference (consistent with the subject of this thread). They asserted that an AI system that’s based on a single sensory input system cannot get far enough down the march of nines to capable of robustly driving itself. They might be wrong. But they’re the ones that are actually doing autonomous driving at something close to scale.

True, but practically there’s a massive hurdle. Tesla sold too many cars in the U.S. while simultaneously claiming they had all the hardware necessary to drive themselves one day. If enough time has passed and things get really dire on this point, they might just have to bite the bullet and take the hit - I agree that if Lidar and high-def AI drivers start to get deployed by multiple companies at scale and Tesla still hasn’t solved vision-only, they’ll have no choice but to change. But the consumer liability for their past statements is a really big hurdle.

As for the economics, I also agree that’s the key factor. It’s entirely possible that the economics of autonomy make infeasible for anyone. But if Tesla can’t solve the AI problem with vision-only, then they lose the distinguishing economic difference between themselves and Waymo.

Right - that’s the point. You can’t look at his past successes and say, “this gives me confidence that this initiative by Tesla will succeed.” Because that’s not his MO - he’s perfectly willing to try things and have them fail. So Musk’s gifts and past successes don’t mean that he’s more likely to be right about autonomy than Waymo is.

In the US, but have not driven the latest versions of FSD. But FSD is very good - too good for anyone to form an accurate assessment of its abilities by driving it.

1 Like