Engineering reliability is primarily achieved through redundancy.

There is none with Musk's "vision only" approach. Vision can fail for a multitude of reasons --- sunlight, rain, darkness, bad road markers, even glare from a dirty windshield. And when it fails, there is no backup plan -- the car is effectively driving blind.

Driving is a dynamic activity that involves a lot more than just vision. Safe automated driving can use all the help it can get.

Just imagine Tesla would have subventioned passive LIDAR on every car they ship to collect data. Wow that dataset would be crazy, and would even improve their vision models by having ground truth to train on. He’s such a moron

Your comparison to hallucination is spot on.

LLMs have shown the general public how AI can be plain wrong and shouldn't be trusted for everything. Maybe this influences how they, and regulators, will think about self driving cars.

> And why not just use LIDAR that can literally see around corners in 3D?

Based on what I've read over the years: it costs too much for a consumer vehicle, it creates unwanted "bumps" in the vehicle visual design, and the great man said it wasn't needed.

Yes, those reasons are not for technology or safety. They are based on cost, marketing, and personality (of the CEO and fans of the brand).

I’ve always wondered about LiDAR - how can multiple units sweep a scene at the same time (as would be the case for multiple cars driving close together, all using lidar)? One unit can’t distinguish return signals between itself and other units, can it?

I use FSD in my Model S daily to commute from SF to Palo Alto along with most of my other Bay Area driving. It does a better job currently than most people and it drives me 95% of the time now I haven't had the phantom braking.

I'm in a 2025 with HW4, but it's dramatic improvement over the last couple of years (previously had a 2018 Model 3) increased my confidence that Elon was right to focus on vision. It wasn't until late last year where I found myself using it more than not, now I use it almost every drive point to point (Cupertino to SF) and it does it.

I think people are generally sleeping on how good it is and the politicization means people are under valuing it for stupid reasons. I wouldn't consider a non Tesla because of this (unless it was a stick shift sports car, but that's for different reasons).

Their lead is so crazy far ahead it's weird to see this reality and then see the comments on hn that are so wrong. Though I guess it's been that way for years.

The position against lidar was that it traps you in a local max, that humans use vision, that roads and signs are designed for vision so you're going to have to solve that problem and when you do lidar becomes a redundant waste. The investment in lidar wastes time from training vision and may make it harder to do so. That's still the case. I love Waymo, but it's doomed to be localized to populated areas with high-res mapping - that's a great business, but it doesn't solve the general problem.

If Tesla keeps jumping on the vision lever and solves it they'll win it all. There's nothing in physics that makes that impossible so I think they'll pull it off.

I'd really encourage people to here with a bias to dismiss to ignore the comments and just go in real life to try it out for yourself.

>why not just use LIDAR that can literally see around corners in 3D?

LIDAR requires line-of-sight (LoS) hence cannot see around conner, but RADAR probably can.

It's interesting to note that the all time 2nd most popular post on Tesla is 9 years ago on its full self driving hardware (just 2nd after the controversial Cybertruck) [1].

>Elon's vision-only move was extremely "short-sighted"

Elon's vision was misguided because some of the technologists at the time including him seem to really truly believed that AGI is just around the corner (pun attended). Now most of the tech people gave up on AGI claim blaming on the blurry definition of AGI but for me the truly killer AGI application is always full autonomous level 5 driving with only human level sensor perceptions minus the LIDAR and RADAR. But the complexity of the goal is very complicated that I really truly believe it will not be achieved in foreseeable future.

[1] All Tesla Cars Being Produced Now Have Full Self-Driving Hardware (2016 - 1090 comments):

https://news.ycombinator.com/item?id=12748863

Camera only might work better if you used regular digital cameras along with more advanced cameras like event based cameras that send pixels as soon as they change brightness and have microsecond latency and\or Single Photon Avalanche Diode (SPAD) sensors which can detect single photons. Having the same footage from all 3 of these would enable some fascinating training options.

But Tesla didn't do this.

Cannot agree more on this phantom braking.

I rented a Tesla a while back and drove from the bay to the death valley. On clear roads with no hazards whatsoever, the car hit the brakes at highway speeds. It scared the bejeesus out of me! Completely off put by the auto drive and derailed plans to buy a Tesla.

The around corners thing, when I saw demos of it seeing the vehicles the driver can't even see ... I wanted it for my non self driving car ... it's just too big of an advantage to skimp out on.

> maybe there are filters to disregard unlikely objects as irrelevant, which act as guardrails against random braking

The filters introduce the problem of incorrectly deleting something that really is there.

Whats the oddest thing about the wiper tech is that we've had the tech for automated wipers since at least the 70s. As a kid my neighbor's Cadillac had it.

tl;dr: you can use optics to determine if there's rain on a surface, from below, without having to use any fancy cameras or anything, just a light source and light sensor.

If you're into this sort of thing, you can buy these sensors and use them as a rain sensor, either as binary "yes its rained" or as a tipping bucket replacement: https://rainsensors.com

Elon's vision-only move was extremely "short-sighted" (heheh)

Careful. HN takes a dim view of puns.

"There are many reasons but that drives it home for me multiple times a week is that my Tesla's wipers will randomly sweep the windshield for absolutely no reason."

Self-starting wipers uses some kind of current/voltage measure on the windshield right - unrelated to self-driving? It's been around longer than Tesla - or are you just saying it's another random failure?

I upvoted this just for "short-sighted."

How does lidar see around corners?

This and that. FUD this FUD that. Tesla have communicated clearly why "adding" LiDAR isn't an improvement for a system with goals as high as their are. Remember, no vision system yet is as good as humans are with vision, so obviously there's a lot to do with vision still.

Check this for a reference of how well Tesla's vision-only fares against the competition, where many have LiDAR. Keep it simple wins the game. https://www.youtube.com/watch?v=0xumyEf-WRI

>Elon's vision-only move was extremely "short-sighted"

It wasn't Elon's but Karpathy's.

Not sure it was actually Elon's move though, I heard it was mainly a decision taken by Andrej Karpathy

I think Elon's prediction was that LIDAR was too expensive and will stay too expensive. In a sense he was right, LIDAR prices did not drop and I wonder why that is?

Are you referring to autopilot or FSD? Phantom braking is a solved problem since the release of V12 FSD. As soon as a vision based car is safer than a human, it's flaws don't matter because it will save lives.

Supervised FSD is already safer than a human.

Lidar is great for object detection. But it's not great for interpreting the objects. It will stop you crashing into a traffic light. But it won't be able to tell the color of the light. It won't see the stripes on the road. It won't be able to tell signs apart. It won't enable AIs to make sense of the complex traffic situations.

And those complex traffic situations are the main challenge for autonomous driving. Getting the AIs to do the right things before they get themselves into trouble is key.

Lidar is not a silver bullet. It helps a little bit, but not a whole lot. It's great when the car has to respond quickly to get it out of a situation that it shouldn't have been in to begin with. Avoiding that requires seeing and understanding and planning accordingly.

I've defended some of Musk because I think what he did for Twitter was completely necessary (showing Jay Bhattacharya that the old regime had put him on a trends blacklist, and all the other people who got banned for no reason) but things like this (and Tesla's already been accused of killing people through crashes) are alarming (vision only as opposed to multiple telemetry) and it's kind of amazing he's in charge of something like SpaceX (are we about to witness a fatal incident in space?)

The wiper system has nothing to do with self-driving -- it's based on total internal reflection in the glass: https://www.youtube.com/watch?v=TLm7Q92xMjQ

The mistakes you describe are the issues of the AI system controlling the car, not of the cameras themselves. If you were watching the camera feed and teleoperating the vehicle, no way you'd phantom brake at a sudden bit of glare.

As someone who worked in this space, you are absolutely right, but also kind of wrong - at least in my opinion.

The cold hard truth is that LIDARs are a crutch, they're not strictly necessary. We know this because humans can drive without a LIDAR, however they are a super useful crutch. They give you super high positional accuracy (something that's not always easy to estimate in a vision-only system). Radars are also a super useful crutch because they give really good radial velocity. (Little anecdote, when we finally got the Radars working properly at work it made a massive difference to the ability for our car to follow other cars, ACC, in a comfortable way).

Yes machine learning vision systems hallucinate, but so do humans. The trick for Tesla would be to get it good enough to where it hallucinates less than humans do (they're nowhere near yet - human's don't hallucinate very often).

It's also worth adding that last I checked the state of the art for object detection is early fusion where you chuck the LIDAR and Radar point clouds into a neural net with the camera input so it's not like you'd necessarily have the classical methods guardrails with the Lidar anyway.

Anyway, I don't think Tesla were wrong to not use LIDAR - they had good reasons to not go down that route. They were excessively expensive and the old style spinning LIDARs were not robust. You could not have sold them on a production car in 2018. Vision systems were improving a lot back then so the idea you could have a FSD on vision alone was plausible.

If a human brain can tell the difference between sun glare and an object, machine learning certainly can.

It’s already better at X-rays and radiology in many cases.

Everything you are talking about is just a matter of sufficient learning data and training.

Premature? Is that what we call this now? It's straight up fraud!

Others are in prison for far less.

Not gonna happen as long as Musk is CEO. He's hard over on a vision-only approach without lidar or radar, and it won't work. Companies like Waymo that use these sensors and understand sensor fusion are already eating Tesla's lunch. Tesla will never catch up with vision alone.

So does anyone who previously bought it on claims that actual full self-driving would be "coming soon" get refunds?

This is fraud he went in front of investors and said multiple times it was around the corner.

He said consumers, just buy the car and it will come with an updated. It didn't.

This is a scam, end of story.

7 years of it.

>false advertising

I think you mean "securities fraud", at gargantuan scale at that. Theranos and Nikola were nowhere near that scale.

But, they're changing the meaning of FSD to FSD (Supervised). So that means they don't make any promises for unsupervised FSD in the future anymore. They'll of course say that they keep working on it and that stuff is progressing. But they don't have to deliver anymore. Just like they say to people getting into accidents that they should keep their arms on the wheel or else it's your own responsibility.

> Maybe they'll reach level 4 or higher automation

There is little to suggest that Tesla is any closer to level 4 automation than Nabisco is. The Dojo supercomputer that was going to get them there? Never existed.

What does Waymo lack in your opinion to not be considered "full self driving"?

The persistent problem seems to be severe weather, but the gap between the weather a human shouldn't drive in and weather a robot can't drive in will only get smaller. In the end, the reason to own a self-driven vehicle may come down to how many severe weather days you have to endure in your locale.

Do Waymos (without safety driver in the car) count as FSD?

> This looks to me like they are acknowledging that their claims were premature, possibly due to claims of false advertising, but are otherwise carrying forward as they were.

Delusionaly generous take. Perhaps even zealotry.

Normally the Board of Directors would fire any CEO that destroyed as much of the company's value as Musk has. But Tesla's board is full of Musk syncophants and family members who refuse to stand up to him.

Who was going to stop them from lying?

Same experience in a mix of city/suburban/rural driving, on a HW3 car. Seeing my car drive itself through complex scenarios without intervention, and then reading smart people saying it can’t without hardware it doesn’t have, gives major mental whiplash.

I would like to get my experience more in line with yours. I can go a few miles without intervention, but that's about it, before it does something that will result in damage if I don't take over. I'm envious that some people can go months when I can't go a full day.

Self driving is not the same as "autonomy". Musk lied to everyone with the Tesla self driving, the Boring Company, DOGE...wake up people...

> Based on the rate of progress alone I would expect functional vision-only self-driving to be very close.

So close yet so far, which is ironically the problem vision based self-driving has. No concrete information just a guess based on the simplest surface data.

On a scale from "student driver" to "safelite guy (or any other professional who drives around as part of their job) running late" how does it handle storrow and similiar?

Like does it get naively caught in stopped traffic for turns it could lane change out or does it fucking send it?

Honestly at that point of complexity I hope automakers just quit chasing FSD and go back to making actually good cars again.

Let the automated trucks figure it out if it’s an actual problem worth solving or we can just use trains or let truck driving be a decent middle class job.

SpaceX is a success despite Elon. Maybe setting an extremely lofty goal helped somewhat but Gwynne Shotwell and all the actual engineers at SpaceX deserve the credit for their success.

We live in an area with sort of challenging roads, and I strongly disagree.

There’s an increasing number of drivers that can barely drive on the freeways. When they hit our area they cannot even stay on their side of the road, slow down for blind curves (when they’re on the wrong side of the road!), maintain 50% the normal speed of other drivers, etc. I won’t order uber or lyft anymore because I inevitably get one of these people as my driver (and then watch them struggle on straight stretches of freeway).

Imagine how much worse this will get when they start exclusively using lane keeping on easy roads. It’ll go from “oh my god I have to work the round wheel thingy and the foot levers at the same time!” to “I’ve never steered this car at speeds above 11”.

I’d much rather self driving focused on driving safely on challenging roads so that these people don’t immediately flip their cars (not an exaggeration; this is a regular occurrence!) when the driver assistance disables itself on our residential street.

I don’t think addressing this use case is particularly hard (basically no pedestrians, there’s a double yellow line, the computer should be able to compute stopping distance and visibility distance around blind curves, typical speeds are 25mph, suicidal deer aren’t going to be the computer’s fault anyway), but there’s not much money in it. However, if you can’t drive our road, you certainly cannot handle unexpected stuff in the city.

We already have transponders on freeways. They’re technically passive reflectors, but they reflect a high proportion of incident EM waves, in the visible spectrum, and exist between lanes on every major road in the US. Also known as white paint.

Following roads and lane markers and signs and signals is the "easy" part of autonomous driving. You could do everything you say and it wouldn't result in something that is any better than the current state of the art. Dealing with others on the road is the main problem (weather comes in close second). Your solution solves nothing, I'm afraid.

How would you know which signals to trust and which to ignore?

Until very recently I worked in the FSD community, and I wouldn’t say I viewed it as a joke. I don’t know if I believed they would get to level 5 without any lidar, it’s pretty good for what’s available in the consumer market.

> In other words: you can take your mind off the road as long as you keep your mind on the road.

They literally did this with Summon. "Have your car come to you while dealing with a fussy child" - buried far further down the page in light grey, "pay full attention to the vehicle at all times" (you know, other than your "fussy child").

Tesla does not use stereo/binocular vision, that's not how humans perceive relative motion at that distance either, we would depend on perspective and parallax

Lidars have been reporting per-point intensity values for quite a while. The dynamic range is definitely not 1 bit.

Many Lidar visualization software will happily pseudocolor the intensity channel for you. Even with a mechanically scanning 64-line Lidar you can often read a typical US speed limit sign at ~50 meter in this view.

Silicon Valley tech workers and companies are not known for their morals.

This is not a heavily researched comment, but it seems to me that the Model 3 is relatively affordable, at least compared to available options at the time. It depends on your point of comparison: there is a lot of competition for sure. The M3 was successful to a good degree, don’t you think? I mean, we should put a number on it so we’re not just comparing feels. The Model Y sells well too, doesn’t at least until the DOGE insanity.

They went from no revenue to the 9th most valuable company in the world under him. No vehicle sales to having the best selling vehicles in the world.

They are still profitable, have very little debt and a ton of money into the bank.

Every company has hits and misses. Bezos started before Musk and still hasn't gotten his rockets into orbit.

Tesla haters tend to just move the goal posts.

Daily reminder that Telsa is not — nor was ever intended to be — a car company. Tesla is fundamentally an "energy generation and storage" (battery/supercapacitor) company. Given Tesla's fundamentals (the types of assets they own, the logistics they've built out), the Powerwall and Megapack are closer to Tesla's core product than the cars are. (And they also make a bunch of other battery-ish things that have no consumer names, just MIL-SPEC procurement codes.)

Yes, right now car sales make up 78% of Tesla's revenue. But cars have 17% margins. The energy-storage division, currently at 10% of revenue, has more like 30% margins. And the car sales are falling as the battery sales ramp up.

The cars were always a B2C bootstrap play for Tesla, to build out the factories it needed to sell grid-scale batteries (and things like military UAV batteries) under large enterprise B2B contracts. Which is why Tesla is pushing the "car narrative" less and less over time, seeming to fade into B2C irrelevancy — all their marketing and sales is gradually pivoting to B2B outreach.

Wait for them? or buy them?

Tesla has binding arbitration that prohibits class actions.

> But Robotaxi shows it is very meaningfully progressing towards true autonomy.

What? They literally just moved the in car supervisor from the passenger seat to the driver seat. That's not a vote of confidence.

And I don't think you can glean anything. There are less than 20 Robotaxis in Austin, that spend their time giving rides to influencers so they can make YT videos where even they have scary moments.

Gotta keep juicing it to get that $1T payout.

Am I the only one that noticed most of the targets are in nominal dollars, not inflation adjusted? Trump’s already prosecuting Fed leadership because they’re refusing to print money for him. Elon’s worked with him enough to understand where our monetary policy is headed.

The word "driving" has multiple, partially overlapping meanings. You're using it in a very informal sense to mean "I don't have to touch the controls much". Power to you for using whatever definitions you feel like.

Other people, most importantly your local driving laws, use driving as a technical term to refer to tasks done by the entity that's ultimately responsible for the safety of the entire system. The human remains the driver in this definition, even if they've engaged FSD. They are not in a Waymo. If you're interested in specific technical verbiage, you should look at SAE J3016 (the infamous "levels" standard), which many vehicle codes incorporate.

One of the critical differences between your informal definition is whether you can stop paying attention to the road and remain safe. With your definition, it's possible have a system where you're not "driving", but you still have a responsibility to react instantaneously to dangerous road events after hours of of inaction. Very few humans can reliably do that. It's not a great way to communicate the responsibilities people have in a safety-critical task they do every day.

Tesla set their own goal posts.

In 2016 Tesla claimed every Tesla car being produced had "the hardware needed for full self-driving capability at a safety level substantially greater than that of a human driver": https://web.archive.org/web/20161020091022/https://tesla.com...

It was a lie then and remains a lie now.

Is Full just a catch word for actually not full now?

Full Speed USB is 12Mbps, nobody wants a Full Speed USB data transfer.

Full Self Driving requires supervision. Clearly, even Tesla understands the implication of their name, or they wouldn't have renamed it Full Self Driving Supervised... They should probably have been calling it Supervised Self Driving since the beginning.

The [2016 Tesla promotional] video carries a tagline saying: “The person in the driver’s seat is only there for legal reasons. He is not doing anything. The car is driving itself.”

https://www.reuters.com/technology/tesla-video-promoting-sel...

Musk promised unsupervised driving being right around the corner so many times it became a joke.

https://youtu.be/B4rdISpXigM

In 2016 Musk said you’d be able to drive from LA to NYC without touching the steering wheel once “within 2 years”. He’s been making untrue statements about Tesla FSD for a decade.

The article is wrong. What happened is that they added some highway-capable ridehail vehicles, and only those vehicles have the safety person in the drivers seat. Frederic (the author) lives in Canada, he doesn't have access to any recent version of FSD.

Fish rots from the head.

Tesla's share price is all based on the Greater Fool Theory in the short run.

In the long run some of those promises might materialise. But who cares! Portfolio managers and retail investors want some juicy returns - share price volatility is welcomed.

Marketing choice of words aside, it's already really good now to the point that it probably does 95% of my driving. Once in a while it chooses the wrong lane and very rarely I will have to intervene, but it's always getting better. If they just called it "Advanced Driver Assist" or something, and politics weren't such an emotional trigger, it would be hailed as a huge achievement.

Great spin job. They didn’t lie, they’re just “pivoting their messaging”

Can you find any statement in the article that is false?