US probes Tesla's Full Self-Driving software after fatal crash

That said, there is a night and day difference between FSD 12.3 that you experienced earlier this year and the latest version 12.6. It will still make mistakes from time to time but the improvement is massive and obvious. More importantly, the rate of improvement in the past two months has been much faster than before.

Yesterday I spent an hour in the car over three drives and did not have to turn the steering wheel at all except for parking. That never happened on 12.3. And I don't even have 12.6 yet, this is still 12.5; others report that 12.6 is a noticeable improvement over 12.5. And version 13 is scheduled for release in the next two weeks, and the FSD team has actually hit their last few release milestones.

People are right that it is still not ready yet, but if they think it will stay that way forever they are about to be very surprised. At the current rate of improvement it will be quite good within a year and in two or three I could see it actually reaching the point where it could operate unsupervised.

It got stuck in a side street trying to get to a target parking lot, shaking the wheel back and forth.

It's no better so far and this is the first day.

As I said, it still makes mistakes and it is not ready yet. But 12.3 was much worse. It's the rate of improvement I am impressed with.

I will also note that the predicted epidemic of crashes from people abusing FSD never happened. It's been on the road for a long time now. The idea that it is "irresponsible" to deploy it in its current state seems conclusively disproven. You can argue about exactly what the rate of crashes is but it seems clear that it has been at the very least no worse than normal driving.

Don't get me wrong without a concerted data team building maps a priori, this is pretty incredible. But from a pure performance standpoint it's a shaky product.

Oh dear.

Glad you're okay!

I'll believe it when I see it. I'm not sure "quite good" is the next step after "feels dangerous".

>And I don't even have 12.6 yet, this is still 12.5;

How am i supposed to take anything you say seriously when your only claim is a personal anecdote that doesn't even apply to your own argument. Please, think about what you're writing, and please stop repeating information you heard on youtube as if it's fact.

The is one of the reasons (among many) that I can't take Tesla booster seriously. I have absolutely zero faith in your anecdote that you didn't touch the steering wheel. I bet it's a lie.

12.5.6 Still fails to read very obvious signs for 30 Km/h playgrounds zones.

The current vehicles lack sufficient sensors, and likely do not have enough compute power and memory to cover all edge cases.

I think it's a matter of time before Tesla faces a lawsuit over continual FSD claims.

My hope is that the board will grow a spine and bring in a more focused CEO.

Hats off to Elon for getting Tesla to this point, but right now they need a mature (and boring) CEO.

I didn't say I didn't touch the steering wheel. I had my hands lightly touching it most of the time, as one should for safety. I occasionally used the controls on the wheel as well as the accelerator pedal to adjust the set speed, and I used the turn signal to suggest lane changes from time to time, though most lane choices were made automatically. But I did not turn the wheel. All turning was performed by the system. (If you turn the wheel manually the system disengages). Other than parking, as I mentioned, though FSD did handle some navigation into and inside parking lots.

Here's an unintentionally hilarious meta-post on the subject https://news.ycombinator.com/item?id=29531915

But the rate of progress I see right now has me thinking that it may not be more than two or three years before that threshold is finally reached.

How far are we now, 6 years later than I expected?

Dunno.

I suspect it's gonna need an invention on the same level as Diffusion or Transformer models to be able to get all the edge cases we can get, and that might mean we only get it with human level AGI.

But I don't know that, it might be we've already got all we need architecture-wise and it's just a matter of scale.

Only thing I can be really sure of is we're making progress "quite fast" in a non-objective use of the words — it's not going to need a re-run of 6 million years of mammilian evolution or anything like that, but even 20 years wall clock time would be a disappointment.

They may well best Tesla to this, though.

I’m not GP, but I can share video showing it driving across residential, city, highway, and even gravel roads all in a single trip without touching the steering wheel a single time over a 90min trip (using 12.5.4.1).

I suspect the free trials let tesla collect orders of magnitude more data on events requiring human intervention. If each one is a learning event, it could exponentially improve things.

I tried it on a loaner car and thought it was pretty good.

One bit of feedback I would give tesla - when you get some sort of FSD message on the center screen, make the text BIG and either make it linger more, or let you recall it.

For example, it took me a couple tries to read the message that gave instructions on how to give tesla feedback on why you intervened.

EDIT: look at this graph

https://electrek.co/wp-content/uploads/sites/3/2024/10/Scree...

Please stop presenting arguments of the form "I haven't seen problems so people who have problems must be extreme outliers". At best it's ignorant, at worst it's actively in bad faith.

Sounds like Tesla drivers have been at the Kool-Aid then.

But to be a bit more serious, the problem isn't necessarily that people don't think it's improving (I do believe it is) or that they will likely succeed (I'm not sure where I stand on this). The problem is that every year Musk says the next year will be the Year of FSD. And every next year, it doesn't materialize. This is like the Boy Who Cried Wolf; Musk has zero credibility with me when it comes to predictions. And that loss of credibility affects my feeling as to whether he'll be successful at all.

On top of that, I'm not convinced that autonomous driving that only makes use of cameras will ever be reliably safer than human drivers.

I would love for self-driving to succeed. I do long-ish car trips several times a year, and it would be wonderful if instead of driving, I could be watching a movie or working on something on my laptop.

I've tried Waymo a few times, and it feels like magic, and feels safe. Their record backs up that feeling. After everything I've seen and read and heard about Tesla, if I got into a Tesla with someone who uses FSD, I'd ask them to drive manually, and probably decline the ride entirely if they wouldn't honor my request.

> fortunately, it's clear as day to some of us that tesla approach will work

And based on my experience with Tesla FSD boosters, I expect you're basing that on feelings, not on any empirical evidence or actual understanding of the hardware or software.

That’s not a reasonable assumption. You can’t just extrapolate “software rate of improvement”, that’s not how it works.

The regressions are getting worse. For the first release anouncement it was only hitting regulatory hurdles and now the entire software stack is broken? They should fire whoever is in charge and restore the state Elon tried to release a decade ago.

Has it really? How much data is each car sending to Tesla HQ? Anybody actually know? That's a lot of cell phone bandwidth to pay for, and a lot of data to digest.

Vast amounts of data about routine driving is not all that useful, anyway. A "highlights reel" of interesting situations is probably more valuable for training. Waymo has shown some highlights reels like that, such as the one were someone in a powered wheelchair is chasing a duck in the middle of a residential street.

As they say, they take "full responsibility"

Even a system that is 99% reliable will honestly feel very, very good to an individual operator, but would result in huge loss of life when scaled up.

Tesla can earn more trust be releasing the data necessary to evaluate the system’s performance. The fact that they do not is far more informative than a bunch of commentators saying “hey it’s better than it was last month!” for the last several years — even if it is true that it’s getting better and even if it’s true it’s hypothetically possible to get to the finish line.

Otherwise every car brand will have to kill a whole heap of people too until they manage to make a FSD system.

This is important because I'd bet similar data on the use of standard, non-adaptive cruise control would similarly show it's much safer than human drivers. But of course that would be because people use cruise control most in long-distance highway driving outside of congested areas, where you're least likely to have an accident.

It relies on inferred depth from a single point of view. This means that the depth/positioning info for the entire world is noisy.

From a safety critical point of view its also bollocks, because a single birdshit/smear/raindrop/oil can render the entire system inoperable. Does it degrade safely? does it fuck.

> recognizes that training exceptional deep neural networks requires vast amounts of data,

You missed good data. Recording generic driver's journeys isn't going to yield good data, especially if the people who are driving aren't very good. You need to have a bunch of decent drivers doing specific scenarios.

Moreover that data isn't easily generalisable to other sensor suites. Add another camera? yeahna, new model.

> Tesla recently held a robotaxi event, explicitly informing investors of their plans

When has Musk ever delivered on time?

> his ability to achieve results

most of those results aren't that great. Tesla isn't growing anymore, its reliant on state subsidies to be profitable. They still only ship 400k units a quarter, which is tiny compared to VW's 2.2million.

> attract top engineering and management talent is undeniable

Most of the decent computer vision people are not in tesla. Hardware wise, their factories aren't fun places to be. He's a dick to work for, capricious and vindictive.

Tesla collects their own data, builds their own training clusters with both Nvidia hardware and their own custom hardware, and deploys their own custom inference hardware in the cars. There is no obstacle to them scaling up massively in all dimensions, which basically guarantees significant progress. Obviously you can disagree about whether that progress will be enough, but based on the evidence I see from using it, I think it will be.

Also, "years of death and injury" is a bald-faced lie. NHTSA would have shut down FSD a long time ago if it were happening. The statistics Tesla has released to the public are lacking, it's true, but they cannot hide things from the NHTSA. FSD has been on the road for years and a billion miles and if it was overall significantly worse than normal driving (when supervised, of course) the NHTSA would know by now.

The current investigation is about performance under specific conditions, and it's possible that improvement is possible and necessary. But overall crash rates have not reflected any significant extra danger by public use of FSD even in its primitive and flawed form of earlier this year and before.

One of these years, he is going to be right. And at that point, the fact that he was wrong for a long time won't diminish their achievement. As he likes to say, he specializes in transforming technology from "impossible" to "late".

> I'm not convinced that autonomous driving that only makes use of cameras will ever be reliably safer than human drivers.

Believing this means that you believe AIs will never match or surpass the human brain. Which I think is a much less common view today than it was a few years ago. Personally I think it is obviously wrong. And also I don't believe surpassing the human brain in every respect will be necessary to beat humans in driving safety. Unsupervised FSD will come before AGI.

You've been lied to.

As for whether it's worthwhile to solve that problem when having more sensors will always be safer, that's another issue...

On the other hand, I fear/suspect it is asymptotically, rather than linearly, approaching good enough to be unsupervised. It might get halfway there, each year, forever.

While it ought to be possible to solve for just RGB… making it needlessly hard for yourself is a fun hack-day side project, not a valuable business solution.

The report measures accidents in FSD mode. Qualifiers to FSD mode: the conditions, weather, road, location, traffic all have to meet a certain quality threshold before the system will be enabled (or not disable itself). Compare Sunnyvale on a clear spring day to Pittsburgh December nights.

There's no qualifier to the "comparison": all drivers, all conditions, all weather, all roads, all location, all traffic.

It's not remotely comparable, and Tesla's data people are not that stupid, so it's willfully misleading.

This report does not include fatalities. It also doesn't consider any incident where there was not airbag deployment to be an accident. Sounds potentially reasonable until you consider:

- first gen airbag systems were primitive: collision exceeds threshold, deploy. Currently, vehicle safety systems consider duration of impact, speeds, G-forces, amount of intrusion, angle of collision, and a multitude of other factors before deciding what, if any, systems to fire (seatbelt tensioners, airbags, etc.) So hit something at 30mph with the right variables? Tesla: "this is not an accident".

- Tesla also does not consider "incident was so catastrophic that airbags COULD NOT deploy*" to be an accident, because "airbags didn't deploy". This umbrella could also include egregious, "systems failed to deploy for any reason up to and including poor assembly line quality control", as also not an accident and also "not counted".

Musk in 2021: "Right now our highest priority is working on solving the problem."

But no. I want self-driving to succeed, and when it does (which I don't think is that soon, because the last 10% takes 90% of the time), I don't think Tesla or their approach will be the "winner".

If you treat the car like it's a student driver or someone else driving, disengagements will go do. If you treat it like you're driving there's also something to complain about.

Sensor platforms deployed at scale, that you have the right to take data from, are difficult to replicate.

Obviously it works some time on some roads, but not all the time on all the roads. A film with it when it works on the road it works is cherry-picking. Look up what the term means.