Hacker News Reader

My favorite case so far of the '5%' that you mention happened on my Tesla irt:(object recognition) and I still laugh about it to this day.

I was driving down the road as normal, 4 lane divided highway that's a bit hilly. Suddenly my car starts having what I can only describe as a panic attack saying I'm running a stop sign and blaring alarms.

It was detecting a giant 40ft tall red circle sign a bit away as a stop sign...

That's interesting, because my car never has blared an alarm for that. I live in a part of the country where, in certain parts of semi-rural and hilly areas, they've decided to introduce 4-way stops instead of red lights, or to slow down traffic on a straightaway. As a result, every so often I would not realize there's a 4-way stop (luckily google maps now shows them) and stop a little late. This happened 2-3 times in the last year or two and never did the Tesla make a peep.

Replying to my own comment because I just remembered this as well...

I definitely saw a case of them overfitting their neural networks lately.

Going over a single lane bridge that's an exit ramp, the car started decoding the "other side" of the concrete barrier as oncoming traffic lanes...when there was nothing there.

Excellent example: "reflective cistern tank with a reflection of the back of a truck transporting stop signs"

People tend to forget just how hard the edge cases in vision are!

It’s weird to me people forget how bad vision can be for driving and Exocet Tesla to somehow be better then our own eyes.

How often a do you encounter situations like bad fog/sun set/rain at night where it’s a total struggle to drive and you slow right down to a crawl and even then only do alright because of a ton of inference?

i think Tesla deciding to go vision only will be regarded of one of the greatest blunders is self driving history.

The counterargument to this is that since humans reach acceptable safety levels with vision only, it must be possible to do self driving with vision only. That said, augmenting vision with other methods does seem like a no brainer for better performance.

We have a lot more signal than vision only. For example audio, the “feel of the road”, like feedback on the steering wheel and traction that we physically experience. Most of all we have actual intelligence and reasoning - not just pattern recognition.

Feel of the road is easy enough to get with the traction control hardware that most modern cars have.

Do you drive much? It would be QUITE the stretch to say most drivers have "actual intelligence and reasoning."

Pattern matching for driving is probably better, frankly. You don't have people who are stressed out, pissed off, inattentive or in a hurry doing risky stuff on the road.

Like all drivers I also hate other drivers, but "actual intelligence and reasoning" means "I can tell that person is gonna try and cross the street" or "I can tell that's a picture of a bicycle instead of an actual bicycle". This stuff is hard just by vision alone--humans use a lot of context to reach these conclusions (bicycles don't float, they don't hang out on the side of advertising boards, they aren't 30 ft. tall) that computer vision machines don't have.

> It would be QUITE the stretch to say most drivers have "actual intelligence and reasoning."

The accidents caused by people with a lapse in judgement are much more memorable than the billions driving safely and preventively every day.

The counterargument to THAT is that human safety levels aren’t acceptable. They are tolerable perhaps, but I wouldn’t call the number of accidents and fatalities we have today acceptable.

The average human isn't acceptable, but in principle a system like this should still be better than the best human just because it has more cameras than humans have eyeballs and they can arranged so there are no blind spots, and even with just two cameras placed (for no good reason) inside the cabin should be able to reach the performance of the best human all of the time.

Current AI isn't that, but in principle it could get there.

Though human accidents are seldom due to the inherent limitations of vision.

That counterargument only holds if Tesla can build software that can approximate the human brain. I think it's laughable to expect they can do that, at least on any reasonable timeframe.

Even if they could, a goal of self driving should be to do better than a human driver. Avoiding technology that can "see" in ways a human cannot is just short-sighted, and a huge missed opportunity.

And all that still even ignores the fact that many common environmental conditions make driving only with human eyes very unsafe. Think fog or heavy rain. A car relying only in cameras to drive in those situations will be next to useless.

Why is it laughable that Tesla can build software that approximates the human brain? There is software that is better at chess, better at go and better at poker than humans. Why is driving so special?

Agree with your other points.

i agree with the principle of your comment, but pure games like the ones you mentioned seem like they would be much easier for a machine to solve. traffic and driving are more complex and require much faster response times, not to mention the high variability (and the literal speed) of factors

Two thoughts: how sure are you that the safety levels achieved by humans during bad vision would be considered acceptable for AVs? And secondly: humans have access to a reasonable (non-artificial) general intelligence.

> since humans reach acceptable safety levels with vision only

No AI (reasoning) exists yet, only Machine Learning. It will take decades if not centuries.

Even ignoring that "AI" is generally accepted as a term of art:

What do you mean by "reasoning", such that there is no example of a ML system that does this?

Nope, not one. And not on any car for sure. https://en.m.wikipedia.org/wiki/Automated_reasoning

From actual AI wiki article: “highly mathematical-statistical machine learning has dominated the field”. This will never be able to drive as good as humans.

None of what you have written answers my question (my question wasn't a yes/no one), your quotation isn't in the link, the link is full of examples of things which constitute automated reasoning, and there's no reason for me to doubt the possibility of ML ever achieving what is currently being done by the large number of tiny warm bags of chemicals that is an organic brain.

> The counterargument to this is that since humans reach acceptable safety levels with vision only

100+ car pileups in Southern California checking in to provide a counterexample.

The patchy fog in Southern California on I-5 can go from "not too bad" to "can't see your own hood" in a matter of seconds. Radar is going to catch hazards WAY before a human will.

My thinking is similar. Removing ultrasound may in the end be more of a legal decision than a purely technical one. I suspect neither humans nor ultrasound can deliver real safety under fog conditions or blizzard conditions; so it may be best to clearly fold under truly difficult conditions and cut lawsuits vs Tesla for fog crashes off at the pass. If drivers want to drive in fog; they will be entirely responsible for the results and can hardly argue otherwise.

This leaves the question of moving to radar, but for precise resolution well ahead of the vehicle you need microwaves and a lot of power, I would guess - which reduces the vehicle's range. For all I know you might parboil passersby, too. One old Mig had a radar that would kill and roast rabbits on the runway as it took off, but that's a much different use case, of course.

Or perhaps it will be their advantage in the short term.

Imagine a foggy condition that causes a 50 car pile up on the highway. Which is more likely to avoid the collision, a Tesla that slowed down because it couldn't see or a Waymo/Cruise blasting down the highway at 65 mph because it's Lidar can see through the fog?

Lidar can't see through fog (or snow/rain to enough of a degree), which is one reason tesla has avoided it. Do you mean radar? In the case of radar, I would hope that it becomes a base features of all cars eventually to avoid/mitigate rear endings by preemptive braking.

Many modern lidars absolutely can see through fog/snow/dust/rain, albeit degraded.

Blackmore, and Aeva can see through fog and dust that others can’t. Most sensors can see sufficiently through rain and snow.

If there's a car in front of you, and you're following at normal following distances, it might not be enough. That person will keep going decently fast, and your radar car will follow. The lead car comes to an almost instant stop when it hits, and while radar might pick up the hit, it won't be able to stop as fast as a car hitting a stationary car.

so? With a radar you can follow at a safe distance to brake safely because you can “see” it through the fog. With vision you either slow down to near zero, or follow close enough to see, which doesn’t give you enough time to stop.

Do you understand how radar and stopping distances work? Are you another armchair self driving car expert?

As someone who also follows the progress online and watch the presentations by Cruise, Tesla, etc, I agree that Cruise is well ahead of Tesla.

It feels like Tesla’s main strategy is to add more data, more compute power, more simulation, and hope for “convergence”. Maybe that will work, but right now it feels like Cruise’s technology feels more mature and thought through.

Is it? Remarkably Tesla’s compute bill does not seems to be growing exponentially with the amount of data it is supposedly collecting.

Point is I would not take anything Elon says at face value. He’s a marketer who is constantly bending the truth.

Have you actually watched the presentations? They are not the NSA storing all data from all cars on a giant hard-drive.

They are deploy code to cars to search for potential interesting, take lots of sample, curate and create a test and training set from that data.

Also how do you know what the Tesla compute bill is? Given their investment in GPU clusters and their own development of Dojo their cost may well have grown exponentially.

> He’s a marketer who is constantly bending the truth.

He is also a CEO of two major companies that have a good track record of achieving interesting technology and growth. So just denigrating him to a 'marketer' says more about yourself.

They’re not though.

As of today I can’t buy a production car with Cruise. You don’t get points for building something that’s theoretically superior but not an actual product. It’s the same story with companies like Apple that wipe the floor of wannabe hardware companies with theoretically better specs. Like Apple, Tesla actually ships.

Cruise is going the uber route for their service. The robo-cars are already available in some areas

Not to the public. Or at least it’s “invite only”.

And Tesla FSD is public and not invite only? And the company that has actual driverless cars that you can hire isn't "shipping", but the the car you aren't allowed to take your hands off the wheel is?

Hilarious.

I wasn’t comparing to nor defending Tesla’s self driving system. I was just clarifying that Cruise’s cars are not “available” to the public. I was surprised to learn this, so I thought other might have missed this fact as well.

What happens if you drop a Cruise car at a random place in North America it has never seen before?

I believe that Waymo and Cruise have more capable platforms that can more accurately measure the environment at those boundary conditions, where as Tesla's camera only approach more closely mimics human perception. But where do they stand in terms of datasets used for training their models?

It seems like Tesla has a huge advantage in terms of training data by leveraging a fleet of millions of vehicles.

Which is more valuable, experience or technology?

Having ridden the self driving vans from Waymo in Arizona several times, it really does feel like stepping into the future. Although they only cover a specific geographical area, they have really refined the riding experience within it.

George Hotz is 'in the industry' and seems to disagree with you. I have heard others say the same. It seems people who work at Waymo/Cruise are totally convinced by those things.

Different approaches lead to different paths to solutions, I am not convinced that either will be successful and not convinces Waymo/Cruise is ahead.

Unlike those, Tesla actually makes money and uses the technology stack in more limited forms.

If Tesla wants to be serious about self driving cars they should be seeking regulatory approval and running pilot programs with test drivers. Without regulatory approval it’s just a toy, albeit an incredibly dangerous one.

95% is a ridiculous exaggeration. How well can these cars do in the winter? You know, that season we have that can easily keep the ground covered with snow and/or ice for 30% of the year in many cities. I lived in Chicago and it was genuinely difficult to drive for 3-4 months of the year. How well can these cars do during the very rainy hurricane season in Florida?

Do people who work in this industry actually think they've solved 95% of driving scenarios because their software can manage driving in sunny California, Nevada, and Arizona?

> Everyone in the industry knows that Tesla is nowhere near the tip of the technology. What Tesla does is _fantastic marketing_.

Between one and two years ago, that was my perception too. But the rapid progress I've seen with Tesla FSD Beta over the past couple of years, and over the last year in particular, has forced me to change my mind. (Note: I'm talking only about Tesla's beta software. Tesla's production software is behind by dozens of versions and is without a doubt technologically inferior to Waymo and Cruise.)

Less than two years ago, I would have said FSD Beta could only deal with the "first 95%" too. Now, my perception is that FSD Beta routinely handles the first > 99% and fails only on < 1% of situations. Moreover, the failures have become more graceful -- e.g., the car will stop at intersections perceived as risky and ask the driver to confirm go-ahead by pressing the accelerator. If FSD Beta continues to improve, sooner or later it will cross the threshold at which it becomes safer than most human drivers.

Of course, IF I see new evidence that contradicts my perceptions, I'll change my mind again. There's no shame in changing our minds when the facts disagree with our views. FWIW, I'd love to see videos of Cruise and Waymo vehicles, filmed by tens of thousands of regular consumers driving autonomously on fully unrestricted roads, with zero editorial input from Cruise or Waymo.

With cost of lidar plummeting, I wonder how long it will be before Musk changes his mind.

Musk over and over again has shown that if something that he wants is expensive he will invest as much money as required in it to develop and to build it in large quantities. If Musk had thought lidar cost was the big problem, they would be building a lidar giga-factory right now. But they don't because not using lidar wasn't about cost.

He might be right or wrong about that, but he wouldn't just ignore lidar because its currently on on the market cheaply.

Tesla is far ahead on their ability to train a myriad number of heavy AI models (computer vision) at the same time.

That asset will probably be worth far more than the self driving system.