Elon in shambles
> Our experience as the only company operating a fully autonomous service at this scale has reinforced a fundamental truth: demonstrably safe AI requires equally resilient inputs. This deep understanding of real-world requirements is why the Waymo Driver utilizes a custom, multi-modal sensing suite where high-resolution cameras, advanced imaging radar, and lidar work as a unified system. Using these diverse inputs, the Waymo Driver can confidently navigate the "long tail" of one-in-a-million events we regularly encounter when driving millions of miles a week, leaving nothing to the imagination of a single lens.
Waymo is absolutely delighting in their luck that Elon is so stubborn that he has kept Tesla from being anywhere close to catching up.
According to Elon, "sensor ambiguity" is a danger to the process [1], and therefore only a single type of sensor is allowed. (Conveniently ignores that there can be ambiguity/disagreement between two instances of the same type of sensor)
The fact that people still trust him on literally anything boggles my mind.
[1] https://x.com/elonmusk/status/1959831831668228450
Sensor fusion allows you to resolve that ambiguity, I wonder if Elon is really as in touch with this as you would expect. No single sensor is perfect, they all have their problematic areas and a good sensor fusion scheme allows you to have your sensors reinforce each other in such a way that each operates as close as possible to their area of strength.
No single sensor can ever give you that kind of resilience. Sure, it is easy in that you never have ambiguity, but that means that when you're wrong there is also nothing to catch you to indicate something might be up.
This goes for any system where you have such a limited set of inputs that you never reach quorum the basic idea is to have enough sensors that you always have quorum, and to treat the absence of quorum as a very high priority failure.
Even if it doesn't allow you to resolve the ambiguity, knowing that there is an ambiguity is extremely valuable. Say the lidar detects a pedestrian but the camera doesn't. Which one do you believe? Well, you propagate the ambiguity and take appropriate action, i.e. slow down, change lanes, etc. Don't drive through an area where there's a decent chance that you're going to kill someone by doing it.
Yes, absolutely. Knowledge about the fact that a conflict between sensors exists is valuable in its own right, it means you are seeing something that needs more work than simple reinforcement.
Fail safe, always. That's what I tried to get at with 'absence of quorum', it means you are in uncharted territory.
Last time I checked I relied entirely on vision to drive autonomously.
That's a very naive way of looking at this.
You have an extremely detailed world model including a mental model of the drivers and other road users around you. You rely on sight, sound, experience and lots of knowledge. You are aware of the social contracts at work when dealing with shared resources and your brain is many orders of magnitude more powerful than any box full of electronics.
What you can do with 'just vision' misses the fact that you are part of the hardware.
You also have hearing, you can move your head and wear sunglasses to avoid glare, etc.
You have a much better GPU than it has.
Sensor ambiguity is straight up useful as it can allow you to extract signals that neither sensor can fully capture. This is like... basic stuff too, absolutely wild how he's the richest person in the world and considered this absolute genius
Agreed, anyone who has worked on engineering a moderately complex system involving sensing has explored the power of multi domain sensing... without sensor fusion we'd be in the stone ages.
I've been trying to fuse my stone knives and bearskins, but I fear I will never craft a tricorder.
More importantly you can detect a failed sensor.
Truly. I don't understand why Tesla fans think camera/lidar fusion is unsolvable but camera/camera fusion is a non-issue.
Unsure if you’re trolling, but you haven’t listened to what Tesla are actually saying.
Having more sensors is complicating the matter, but yes sure you can do that if you want to. But just using vision simplifies training a huge amount. The more you think about it, the stronger this argument is. Synthesising data is a lot easier if you’re dealing with one fairly homogenous input.
But the real point is that cameras are cheap, so you can stick them in many many vehicles and gather vast amounts of data for training. This is why Waymo will lose - either to Tesla or more likely a Chinese car manufacturer.
I do not like Elon because I do not think nazi salutes or racism are cool, but I do think Tesla are correct here. Waymo wins for a while, then it dies.
Because they bought a Tesla with only cameras on it.
Admitting this would be admitting their Tesla will never be self driving.
I bought mine with cameras and a radar, which they then deprecated and left an unused. Even though autopilot was better when it had radar. Then I realized that this thing would never be self-driving and that its CEO was throwing nazi salutes. Cut my losses and got rid of it. Gotta admit defeat sometimes.
Add a tow hitch to Waymos and any car can be autonomous!
Do Tesla fans think that? I've seen plenty of Tesla fans say that lidar is unnecessary (which I tend to agree with), but never that lidar is actively detrimental as Musk says there.
I mean, humans have only their eyes. And most of them intentionally distract themselves while driving by listening to music, podcasts, playing with their phones, or eating.
I get your point about camera vs lidar. Humans do have other senses in play while driving though. We have touch/vibration (feeling the road surface texture), hearing, proprioception / acceleration sense, etc. These are all involved for me when I drive a car.
To be fair, humans are fairly poor drivers and generally can't be trusted to drive millions of miles safely.
Actually humans are fairly good drivers. The average US driver goes almost 2 million miles between causing injury collisions. Take the drunks and drug users out and the numbers for humans look even better.
I don't think averages work that way
Incorrect. Humans are fairly good engineers, so cars are pretty safe nowadays.
If you include minor fender-benders and unreported incidents, estimates drop to around 100,000–200,000 miles between any collision event.
This is cataclysmically bad for a designed system, which is why targets are super-human, not human.
Personally as much as people like to dunk on Musk, he did build several successful companies in extremely challenging domains, and he probably listens to the world-leading domain experts in his employ.
So while he might turn out to be wrong, I don't think his opininon is uninformed.
I fully agree with your first point: Musk has shown tremendous ability to manage companies to become unicorns. He's clearly skilled in this domain.
However, if you think about this for 2 seconds with even a rudimentary understanding of sensor fusion, more hardware is always better (ofc with diminishing marginal value).
But ~10y ago, when Tesla was in a financial pinch, Musk decided to scrap as much hardware as possible to save on operational cost and complexity. The argument about "humans can drive with vision only, so self-driving should be able to as well" served as the excuse to shareholders.
I'm an EE, I have worked with things like sensor fusion professionally. In short sensor fusion depends on what sensors you have and how you combine them, especially if two sensors' outputs tend to disagree - which one is wrong and to what extent, and how a piece of noise gets reflected in each sensors' outputs, to avoid double counting errors and coming up with unjustifyably confident results.
This field is extremely complex, it's often better to pick a sensor and stick with it rather than trying to figure out how to piece together data from very dissimilar sources.
> I'm an EE, I have worked with things like sensor fusion professionally. In short sensor fusion depends on what sensors you have and how you combine them, especially if two sensors' outputs tend to disagree - which one is wrong and to what extent, and how a piece of noise gets reflected in each sensors' outputs, to avoid double counting errors and coming up with unjustifyably confident results.
> This field is extremely complex, it's often better to pick a sensor and stick with it rather than trying to figure out how to piece together data from very dissimilar sources.
Whether sensor fusion makes sense is a highly domain specific question. Guidance like "pick a sensor and stick with it" might have been correct for the projects you've worked on, but there's no reason to think this translates well to other domains.
For what it's worth, sensor fusion is extremely common in SLAM type applications.
> humans can drive with vision only, so self-driving should be able to as well
In May 2016, Tesla Model S driver Joshua Brown died in Williston, Florida, when his vehicle on Autopilot collided with a white tractor-trailer that turned across the highway. The Autopilot system and driver failed to detect the truck's white side against a brightly lit sky, causing the car to pass underneath the trailer.
Our eyes are supported by our brain's AGI which can evaluate the input from our eyes in context. All Tesla had is a camera, and it didn't perform as well as eyes + AGI would have.
When you don't have AGI, additional sensors can provide backup. LiDAR would have saved Joshua Brown's life.
What doesn’t make sense to me is that the cameras are no where as good as human eyes. The dynamic range sucks, it doesn’t put down a visor or where sunglasses to deal with beaming light, resolution is much worse, etc. why not invest in the cameras themselves if this is your claim?
I always see this argument but from experience I don't buy it. FSD and its cameras work fine driving with the sun directly in front of the car. When driving manually I need the visor so far down I can only see the bottom of the car in front of me.
The cameras on Teslas only really lose visibility when dirty. Especially in winter when there's salt everywhere. Only the very latest models (2025+?) have decent self-cleaning for the cameras that get very dirty.
FSD doesn't "work fine" driving directly into the sun. There are loads of YT videos that demonstrate this.
For which car? The older the car (hardware) version the worse it is. I've never had any front camera blinding issues with a 2022 car (HW3).
The thing to remember about cameras is what you see in an image/display is not what the camera sees. Processing the image reduces the dynamic range but FSD could work off of the raw sensor data.
Nobody cares that you think v14.7.22b runs well on HW3.1. Literally nobody.
It doesn't run well on HW3 at all. HW4 has significantly better FSD when running comparable versions (v14). The software has little to do with the front camera getting blinded though.
Especially the part where the cameras do not meet minimum vision requirements [1] in many states where it operates such as California and Texas.
[1] https://news.ycombinator.com/item?id=43605034
And to some extent, I also drive with my ears, not only with 2 eyes. I often can ear a car driving on the blind spot. Not saying that I do need to ear in order to drive, but the extra sensor is welcome when it can helps.
There is an argument for sure, about how many sensors is enough / too much. And maybe 8 cameras around the car is enough to surpass human driving ability.
I guess it depends on how far/secure we want the self-driving to be. If only we had a comprehensive driving test that all (humans and robots) could take and be ranked... each country lawmakers could set the bar based on the test.
The other day I slammed the brakes at a green light, because I could hear sirens approaching -- even though the buildings on the corner prevented any view of the approaching fire trucks or their flashing lights. Do Teslas not have this ability?
I don‘t know whether Tesla‘s self-driving mode could do that.
However, notice that deaf people are allowed to drive, ie. you are not expected to be able to have full hearing to be allowed on the road.
Nuanced point: Even if vision alone were sufficient to drive, adding sensors to the cars today could speed up development. Tesla‘s world model could be improved, speeding up development of the vision only model that is truly autonomous.
> I fully agree with your first point: Musk has shown tremendous ability to manage companies to become unicorns. He's clearly skilled in this domain.
I would firmly disagree with that.
What Musk has done is bring money to develop technologies that were generally considered possible, but were being ignored by industry incumbents because they were long-term development projects that would not be profitable for years. When he brings money to good engineers and lets them do their thing, pretty good things happen. The Tesla Roadster, Model S, Falcon 9, Starlink, etc.
The problem with him is he's convinced that he is also a good engineer, and not only that but he's better than anyone that works for him, and that has definitively been proven wrong. The more he takes charge, the worse it gets. The Model X's stupid doors, all the factory insanity, the outdoor paint tent, etc. Model 3 and Model Y arguably succeeded in spite of his interference, but the Dumpstertruck was his baby and we can all see how that has basically only sold to people who want to associate themselves closely with his politics because it's objectively bad at everything else. The constant claims that Tesla cars will drive themselves, the absolute bullshit that is calling it "Full Self Driving", the hilarious claims of humanoid robots being useful, etc. How are those solar roofs coming? Have you heard of anyone installing a Powerwall recently? Heard anything about Roadster 2.0 since he went off claiming it would be able to fly? A bunch of Canadian truckers have built their own hybrid logging trucks from scratch in the time since Tesla started taking money for their semis and we still haven't seen the Tesla trucks haul more than a bunch of bags of chips.
The more Musk is personally involved with a project the worse it is. The man is useful for two things: Providing capital and blatantly lying to hype investors.
If he had stuck to the first one the world as a whole would be a better place, Tesla would probably be in a much better position right now.
SpaceX was for a long time considered to be further from his influence with Shotwell running the company well and Musk acting more as a spokesperson. Starship is sort of his Model X moment and the plans to merge in the AI business will IMO be the Cybertruck.
You say that you disagree with my point, but then your first paragraph just restates my argument. And your subsequent paragraphs don‘t refer to my comment at all.
I never claimed he‘s a good engineer, nor that he has high EQ, nor that he is honest, nor that he has sole responsibility for the success of his companies.
Home batteries are being installed at insane rates in Australia at the moment. Very few of them are Powerwalls because Tesla have priced themselves out of the market (and also Elon’s reputation is toast).
I think his companies succeeded despite Elon. Tesla should be a $5T company and he fucked it up.
Stongly disagree. I don‘t like the fella but thinking that he founds and successfully manages SpaceX and Tesla to their market value _by chance_ is ridiculous.
[dead]
His autopilot has killed several people, sometimes the owner of the car, sometimes other drivers sharing the road. It is hard to root for this guy.
I certainly don't trust anything he says 100%.
This is - to me - entirely separate from the fact that his companies routinely revolutionize industries.
> The fact that people still trust him on literally anything boggles my mind.
Long-distance amateur psychology question: I wonder if he's convinced himself that he's a smart guy, after all he's got 12 digits in his net worth, "How would that have been possible if I were an idiot?".
Anyway, ego protection is how people still defend things like the Maga regime, or the genocide; it's hard for someone to admit that they've been stupid enough to have been fooled to vote for "Idi Amin in whiteface" (term coined by Literature Nobel Prize winner Wole Soyinka), or that the "nation's right to self-defense" they've been defending was a thin excuse for mass murder of innocents.
I've always wondered how people who are not 1/10th as smart as Elon convince themselves that he is not intelligent after solving robotics, AI, neuralink, and space all simultaneously.
And what fraction Elon-Intelligence is needed to believe he actually invented/solved all that by himself?
Or did I miss the sarcasm?
Well, given that Elon openly lies on investor calls...
One of his latest, on the topic of rain/snow/mist/fog and handling with cameras:
"Well, we have made that a non-issue as we actually do photon counting in the cameras, which solves that problem."
No, Elon, you don't. For two reasons: reason one, part A, the types of cameras that do photon counting don't work well for normal 'vision'/imagery associated with cameras, and part B, are not actually present in your cars at all. And reason two, photon counting requires the camera being in an enclosed space to work, which cars on the road ... aren't.
What Elon has mastered the art of is making statements that sound informed, pass the BS detector of laypeople, and optionally are also plausibly deniable if actually called out by an SME.
If only there was a filter so we could fuse different sensor measurements into a better whole..
I don't thing it's purely stubbornness. Tesla sold the promise of software only updates resulting in FSD to hundreds of thousands of people. Not all of those people are in the cult of Tesla. I would expect admitting defeat at this point would result in a large class action lawsuit at the very least.
It wouldn't keep them from equipping _new_ models with additional sensors, spinning a story around how this helps them train the camera-only AI, or whatever.
I know it's "illegal" and technically sold as FSD (assisted), but just 2 days ago I was in a friend's Model Y and it drove from work to my house (both in San Jose) without any steering wheel or pedal touch, at all. And he told me he went to Palm Springs like that too.
I shit on Tesla and Elon on any opportunity, and it's a shame they basically have the software out there doing things when it probably shouldn't, but I don't think they're that far behind Waymo where it really matters, which is the thing actually working.
I suspect they have a long tail problem with FSD. It might work fine 99% but that's simply not good enough.
Palm Springs from San Jose? Albeit freeway throughout but that's quite impressive.
The terms of service probably require you to sue Tesla in that Texas district with his corrupt judge pal.
Elon cult members still to this day will tell me that because humans only use vision to drive all a Tesla needs is simple cameras. Meanwhile, I've been driven by Waymo and Tesla FSD and Waymo is by far my pick for safety and comfort. I actually trusted the waymo I was in, while the Tesla I rode in we had 2 _very_ scary incidents at high speeds in a 1 hour drive.
> humans only use vision to drive
I love this argument because it is so obviously wrong: how could any self aware person seriously argue that hearing, touch, and the inner ear aren't involved in their driving?
As an adult I can actually afford a reliable car, so I will concede that smell is less relevant than it used to be, at least for me personally :)
> hearing, touch, and the inner ear aren't involved
Not to mention possibly the most complex structure in the known universe, the human brain: 86 billion neurons, 100 trillion connections.
Involved? Yes. Necessary? Pretty sure no.
If it makes you happy, you can read "only vision" as "no lidar or radar." Cars already have microphones and IMUs.
1. in US you can get a driver's license if you're deaf so as a society we think you can drive without hearing
2. since this is in context of Tesla: tesla cars do have microphones and FSD does use it for responding to sirens etc.
(1) is true, but actually driving is definitely harder without hearing or with diminished hearing. And Several US states, including CA, prohibit inhibiting hearing while driving, e.g., by wearing a headset, earbuds, or earplugs.
Human inner ear is worse than a $3 IMU in your average smartphone in literally every way. And that IMU also has a magnetometer in it.
Beating human sensors wasn't hard for over a decade now. The problem is that sensors are worthless. Self-driving lives and dies by AI - all the sensors need to be is "good enough".
Human hearing is excellent. Good directional perception and sensitivity. Eyesight is the weakest sense. Poor color sensitivity, low light sensitivity, blindspot. The terrible natural design flaws are compensated by natural nystagmas and the brain filling in the blanks.
> The problem is that sensors are worthless
Well, in TFA the far more successful manufacturer of self driving cars is saying you're wrong. I think they're in much better position to know than you :)
I've long expected Waymo's approach to prevail simply because - aside from whether vision-only proves good enough to some standard - it will be easy to lobby for regulations that favor the more conservative approach.
But I also don't think we can take anything from what Waymo claims about the feasibility of vision-only.
Waymo has posted videos of accidents they've avoided purely because their lidar picked up on a pedestrian before their cameras saw anything.
A favorite of mine: https://x.com/dmitri_dolgov/status/1900219562437861685
They're very public about the data that makes them look good, but they went to court to keep their safety data from the public. (https://techcrunch.com/2022/02/22/waymo-to-keep-robotaxi-saf...)
That lawsuit was about trade secrets shared with DMV. And DMV advised them to file a restraining order against a third party seeking redacted info.
It was actually about all kinds of information about the safety of their cars, how the company handles collisions, how they decide to make human drivers take over, details of injuries caused by their cars, etc.
The DMV required them to submit their safety data before allowing the vehicles on the road. Waymo claims that the data the DMV needed about the public safety of their cars, and the emails they exchanged about it with regulators, were entirely "trade secrets" to keep them hidden from the public who understandably felt like they should be able to access that information since those cars are going to be on their streets.
I think past experience shows that the US prefers a wait and see approach - owning in part I think to it federal structure, where states compete for companies good graces and money, so if State A bans something, State B will allow it and gain an advantage in that area.
Moreover, why draw a hard line on vision only when there is existing technology is available to augment it? It's not like they have to develop 3 novel technologies.