Hacker News Reader

Comment by tomp | Hacker News Reader

You simply have no idea how AI works.

The key ingredient is data - the bitter lesson. It's not about better algorithms, but simply about algorithms that can process more data efficiently (e.g. transformers).

Tesla is one of the few companies that have a data flywheel - a fleet of (non-self-driving) cars collecting real-world data worldwide all the time at massive scale!

Now that is an insourmountable lead. (Along with good engineering, which, believe it or not, is still a competitive advantage - see e.g. German car companies unable to launch a single useful on-board computer, let alone a software-defined self-driving car.)

Google is one of the few companies that could compete, even without Waymo, because of YouTube.

View on HN

That’s the Muskophile babble. Cool story, where is it?

Reality is the Tesla product in this space hasn’t advanced in a decade. They hit a wall. They have promised that the next big thing is just over the next hill, but like the Roadster, hasn’t quite arrived yet.

I think the person you are responding too has too much hype but you are equally too negative. Without a doubt Musk overpromises and under delivers constantly, would never trust him or his words. Teslas FSD has been improving nicely especially over the last 3 years. To say it’s stalled for a decade is wrong.

The overall point is that automated driving limited to cameras can still improve, but that the threshold is probably very close to the quality, in terms of accident and mistakes, of human driving.

The point is that automated driving with a fleet of sensors can be an order of magnitude safer than just using cameras. The current slow improvements at Tesla do not correlate to an unlimited amount of improvements.

“The point” is the original statement I replied to is wrong. FSD has improved significantly in the past few years.

Does that mean I think it’s the future or they have won? Nope, not at all. I don’t like Musk but I also don’t let me feelings for him cloud the discussion.

Tesla has been getting closer to FSD the same way climbing up a ladder gets you closer to the moon. They can keep going the direction they're going, but at this point it's clear there's a pretty hard limit to how close they're going to get without significantly rethinking their sensor strategy.

I don’t think it’s clear to anyone who is going to be the winner 5-10 years from now. Waymo is rolling out and it definitely is a way to solve the problem but it’s still not widely available in the US. I cannot use it to commute to work daily.

I have no idea who is going to win or what that strategy will be, vision only, lidar, combination etc. but it’s false to make claims like yours. I get it though it’s easy to take your hate for an individual in an extension to other adjacent discussions. I don’t like the guy either but not worth making up silly claims.

If you're goin to call my post silly and condescend to me personally (in another post you called me a liar despite me citing sources to back up my claims), please respond with something more substantive than "the future is unknowable".

Please don’t play a victim card for no reason.

“Tesla has been getting closer to FSD the same way climbing up a ladder gets you closer to the moon”

This is a silly claim that perpetuates your dislike of Musk. FSD has absolutely gotten no where close to his claims. I have no idea if vision only is the right path and Tesla probably does play too fast and loose but their FSD program has absolutely been making improvements in, especially in the past few years. How about you back up your claim instead of using hyperbole?

I'm not playing a victim card, you just didn't make an argument. What am I supposed to say about your feelings about my post? You want to make an argument? Fine. You want to just call my post silly? Well you've done that, I don't think I need to add anything more.

I stated my reasoning, the sensor stack in insufficient. The evidence is they've been trying at it for close to a decade now, and have had to continually move the goal posts, missing their self-proclaimed milestones, delaying their shipping dates by years, reneging on their promises and projections, and leaving a trail of death and destruction in their wake (and that's not hyperbole); whereas Waymo went with a better sensor stack (the one roboticists have been saying was the way to go), and is actually shipping a true full self-driving service, one that won't run off the road if you fall asleep.

So we already know what does work (orthogonal sensors like LIDAR+RADAR+Camera) and what doesn't work (Camera only). We learned this in 2007 at the DARPA Urban Challenge, and it's been affirmed by Tesla's continued track record of failure in 2025.

Have you been using the Tesla magic full self driving (before they redefined “full self driving” to a not-full, not-self driver assist tool?

It’s far more likely that you’ll be commuting in a Waymo someday.

Like some other tech moguls, you can’t separate the man from the company. The company lacks meaningful governance, so for good or ill it’s an expression of Elon’s hopes and desires. Facebook is similar, but Zuck is more calculated and rational. Elon is obviously a man in need of help who is degrading before us.

Sure, they’ve made incremental improvements and have a better Level 2 system, it was promised (and sold as) Level 4 in 2006, and the company claims to be “approaching 4”, and now claims to be shipping robot taxis in production in a few months.

Meanwhile, Google actually delivered the robot taxi in SFO, and it’s amazing.

The sensor strategy really says it all. Why skimp on a minor part of the BOM, which enables a use case with 50-100x value?

The whole Tesla schtick is such a bizarre Lucy and the Football scenario, over and over again.

At the high level I don’t disagree but I also don’t know who is going to be the winner 5-10 years from now. The original statement I replied to is simply false that’s all I was saying.

Did you see last weeks video of FSD in Melbourne, even doing a hook turn?

It’s getting better. Slowly, but surely.

Teslas can be programmed to do a lot of things, that's not a question.

The problem with Teslas is given the deficient sensor suite Musk has insisted on, they can't be programmed to not do things, like not run though a picture of a road like Wile E Coyote.

That video was not FSD, but an old version of “driver assist”.

Given how quickly things are changing, I don’t think it’s useful to say iPhones are crap and always will be because you just got a 6 running super old iOS.

No, FSD suffers the same problem, because it has the same insufficient sensor stack. I already debunked that here: https://news.ycombinator.com/item?id=43614273

Things are not changing rapidly, this has been a failure mode on Teslas for almost a decade; it's been a problem at least since 2016 when Joshua Brown was decapitated by his Tesla running on AP. People keep saying "no, that was the last version. The new software/hardware version fixes it", but the looney tunes test demonstrates Tesla's sensor stack continues to be fundamentally insufficient.

And this is why: https://insideevs.com/news/658439/elon-musk-overruled-tesla-...

Musk thinks he knows better than his engineers how to solve this. And no, I don't consider Musk to be an engineer, he's a showman.

The link to your earlier post shows a HW3 car. If you keep watching the video it shows a second HW4 car stopping before the painted wall.

This is literally the definition of progress.

Tesla is currently developing HW5 which has 10x the inference capability of HW3. More progress.

You call this progress?

https://electrek.co/2025/05/23/tesla-full-self-driving-veers...

  My chin split open, and I had to get 7 stitches. After the impact, I was hanging upside down watching blood drip down to the glass sun roof, not knowing where I was bleeding from. I unbuckled my seatbelt and sat on the fabric interior in the middle of the two front seats, and saw that my phone’s crash detection went off and told me the first responders were on their way. My whole body was in shock from the incident.

  Wally said he was on Tesla FSD v13.2.8 on Hardware 4, Tesla’s latest FSD technology. He requested that Tesla send him the data from his car to better understand what happened.

This is not progress, these cars should not be on the road, they are a menace to society. I would never put my life or my family's lives in the hands of Tesla hardware and software. But let me guess, it'll be fixed in HW5, sure. We all just have to be the beta testers.

Lol. Don't bother defending your initial blatant lie, just move on to another one. The crash report shows that, prior to the car veering off the road, 2NM of torque was applied to the steering wheel disengaging FSD and that the turning force was applied until the car crashed. The brake was never pressed.

This person is just too embarrassed to admit they fell asleep at the wheel and the weight of their hand disengaged FSD. This could easily be cleared up if he showed the driver facing camera the car records in a crash.... But something tells me he won't do that.

You don't fact check things very well. That's two blatant falsehoods you've swallowed. What else are you wrong about.

Well if you applied Tesla’s rationalization for their failures, if the driver fell asleep before the crash, it clearly wasn’t his fault as he wasn’t in control.

How is what I said a lie??

My initial post said: "they can't be programmed to not do things, like not run though a picture of a road like Wile E Coyote."

Then someone said: "That video was not FSD, but an old version of “driver assist”."

Then I linked a video showing that in fact FSD was driving through the wall.

Obviously if you have to upgrade the hardware to not run through the wall, then they can't program their current cars to not do so. Because, as I've said since the beginning: their sensors stack is insufficient.

And that's the standard line with Tesla fans is constantly shifting the goal posts. It was supposed to be the case that in 2016, all Teslas sold would be equipped with hardware capable of self driving, because Elon told us that cameras were enough. And yet, they have required hardware refresh after hardware refresh after hardware refresh (now on V5) because he's been wrong time and time again. Because cameras aren't enough.

And what is this progress you speak of? Just special casing edge cases while leaving fundamentally flawed hardware and software on the road. Now it's 2025 and maybe finally HW5 will bring it? I guess we'll have to see because nothing Elon says can be trusted as fact.

> This person is just too embarrassed to admit they fell asleep at the wheel and the weight of their hand disengaged FSD

Ah yes, a driverless system that actually doesn't prevent you from crashing because you fell asleep. I don't know why you feel this alternative take makes Tesla's product look any better. Such progress. Much driverless. It's like a sailboat that sinks instead of floating.

Facts:

FSD will not randomly drive off the road and your post showed a human turning the steering wheel by accident.

Progress:

HW4>HW3 and HW4 cars can see painted walls HW5 exists and will be deployed soon.

Goalposts shifted by you = 2

> Facts: FSD will not randomly drive off the road

How do you know? The point of the video I posted from Youtube was in response to a chorus of people claiming the looney tunes test wasn't valid because FSD wouldn't be affected by it. And yet, they demonstrated FSD plowing through the wall. So Tesla had to respond by upgrading the entire HW stack, because a HW3 car apparently doesn't have the necessary hardware to see the wall. That's called moving a goalpost.

We have Teslas driving into walls, Teslas driving into firetrucks, Teslas driving under tractor trailers and decapitating their passengers. Presumably they weren't supposed to randomly do that either, yeah?

So with what confidence can you tell me they won't drive off the road? Has Tesla proved this? Is there independent testing that has verified it won't? Is there any guarantee Tesla can offer that they won't? Does Tesla even make this claim?

> HW4>HW3 and HW4 cars can see painted walls HW5 exists and will be deployed soon.

The entire existence of HW4 shows us that cameras are not enough. The entire thesis of Musk is that this should have been possible with the original Tesla sensor stack, so why are we on HW5?

  “The Model 3, Model Y, and S and X that we make today, will be capable of full autonomy, unsupervised.” If that sounds familiar, it is because he said almost exactly the same thing several times in the past, as far back as 2016. Musk reiterated this sentiment after the release of HW3/AI3 in 2019, and as we all know, Tesla has since released HW4/AI4, and has HW5/AI5 on the horizon.

https://driveteslacanada.ca/news/tesla-expects-to-roll-out-u...

You want goalpost moving, that's goalpost moving.

We get it. You have a bone to pick with Musk. I don’t like him either but don’t generate lies just to fuel your anger.

Where's the lie?

This is sort of a misguided take IMO - as if LiDAR and other sensor streams are somehow not large scale data while video is

The bitter lesson isn’t fundamentally even about data. the key ingredient is computation (which does scale with data in modern deep learning). There’s even a whole theme on search outperforming learning - until learning methods changed to leverage computation!

I think in a lot of applications, richer data yields better scaling performance. So the bet for self driving is that the complexity of multiple sensor fusion gives a much better constant factor or exponent in the power law scaling of performance with data and compute

Right this is why we periodically see Tesla's in FSD go absolutely bezerk in ways more sensor-laden competitors do not appear to.

https://electrek.co/2025/05/23/tesla-full-self-driving-veers...

This was literally 2 months ago, so miss me with the "you're wrong, the latest FSD solved it all this time bro, for real, trust me (the 100th time I've heard this)".

From the article:

> Despite its name, Full Self-Driving (FSD) is still considered a level 2 driver assist system and is not fully self-driving.

How the hell has Tesla not been sued into the ground by now?

Will we be able mid-term to rely on LLMs not hallucinating and causing crashes? Even if the probability is low, the thought that the AI might do something crazy because it's hallucinating is terrifying, so that might be a barrier for adoption. For the same reason, will a (fully) LLM driven car ever be allowed on Western streets? I have serious doubts regarding Europe, at least.

There are no LLMs driving cars at the moment

Large Language Models are never gonna drive cars, they’re plausible text-generating machines, not general-purpose computer intelligences

Almost every vehicle has multiple cameras in it today, plus cellular connectivity. Tesla doesn't have a major advantage here. Plus, the data isn't immediately valuable on its own, it needs to be labelled and fed back into the training. This is all very far from a data flywheel analogy.

Musk equates machine vision to human vision, but that is an over simplification, and the best MV algorithms and methods are still miles away from human capabilities. FSD is very reliant on depth perception, which is much easier to solve with sensors other than stereo vision.

I also don't get your German car company statement, Mercedes has been a technology leader in the automotive space across a number of fronts for quite a while.

It's not that car companies are unable, but rather, margins on new cars are pretty thin so they are trying to save every last penny.

The key ingredient is indeed data, and also, depending on the stack, hardware. If "true" level 5 self-driving is only possible with LIDAR and up-to-date HD maps, then it won't happen for some time. I foresee that unbounded "full" self-driving will either never happen or with severe boundary conditions only.

It will turn out there isn't a linear relationship between data volumes and successful self driving. There will be a wall somewhere that prevents it from being achieved regardless of data (assuming current AI techniques and hardware available in the foreseeable future). So what will happen is Musk and co will demand that roads be redesigned to accommodate their tech (e.g. cars network and human drivers are banned from major highways).

Maybe, but so far we don't really have any examples of good AI (or even barely working AI like GPT4o) that wasn't trained on massive amounts of data.

Is data enough? Maybe not, there's a lot of progress on RL now that can do wonders without even more data. Is data necessary? No evidence to the contrary, yet.

> what will happen is Musk and co will demand that roads be redesigned to accommodate their tech

Every road, in every country on earth?

No, that is never going to happen.

Transformer LLMs are good at generating text since they were trained on huge volumes of high quality texts. There is no reason to assume that feeding a neural net with unlabeled garbage data (i.e. video recordings from random cars) is going to lead to anything.

How would youtube help with driving?

Massive amount of "world" data - enough for Google to successfully train Veo3, a "video generation" model that is extremely good, and evidently includes a decent "world model" i.e. a model that can generate or, more accurately, predict the (short-term) "future" - like "what happens if you let go of an apple" (it falls down) etc.

This kind of "world models" that can understand physics enough to be able to predict short term future (like humans can) are crucial for any kind of real-world AI (i.e. robotics, including self-driving), because they constitute what could be termed as "physical common sense" (that humans have, but also animals, to some degree).

Is it enough for self-driving? No, you also need to understand road rules, communicate with humans (pedestrians and fellow drivers), etc. but it's a good, possibly necessary, step - it allows you to better handle many unpredictable (tail of distribution) situations: