The only currently known self driving system uses that same sensor suite (neural nets and cameras).

This is wrong. Waymo uses both lidar and radar, which Tesla opted to remove. Also, look at Waymo's cars. Those things have sensors strapped all over in ways that give them lots of range and visibility. Tesla will never do that because it compromises the aesthetic.

https://waymo.com/waymo-driver/

They certainly referred to humans.

Brains aren’t neural nets and eyes aren’t cameras.

Why? The whole idea of neural nets is based on mimicking the brain. It's even in the name: neural (for neuron).

Yeah, and the name of "Full Self-Driving" tells you that it can drive "by itself" i.e. without human intervention.

But names aren't everything, and artificial neutral networks are "inspired by" but in truth drastically different from human brains.

How so?

They're different in practically every way. It's easier to enumerate the things they have in common: NNs borrowed the idea of using a connected network of functions whose outputs feed into each other's inputs.

That's it. That's the total resemblance between the two. The brain isn't just an NN implemented in biology, it has whole systems that aren't accounted for in digital NNs, like hormones and neurotransmitters, and even the system of connected neurons doesn't work the way digital NNs implement it.

Neural networks model the brain exactly as well as objects in OOP model cells: not very well at all. They're inspired by biology, nothing more.

Neural networks as used in AI are inspired by the brain in much the same way that OOP was inspired by the way cells work--neither one is an attempt to faithfully recreate the actual operations of an extremely complex (and only partially understood!) biological system.

I've never met anyone who knew anything at all about neural nets who thought they were mimicking the brain.

If you're talking about humans, I think you're underestimating the importance of the less obvious senses:

- sound

- touch

- equilibrioception (balance)

- proprioception

- and surely there are more

Think of the visceral difference between driving a car in real life, vs driving a car on a screen in a video game.

The problem with driving in video games is using a keyboard/mouse or controller. Driving with a steering wheel and pedals is pretty easy. Even more so if you have the monitors to give you a realistic field of view.

driving a car in a video game with a steering wheel is easy because it's an experience designed from the ground up with that interface in mind. driving games happily do shit like change the fov to make the user think they're going faster, etc.

being easy to drive with a steering wheel in a driving game, and being easy to drive a real car with a steering wheel (with internet level latency and packet loss at play, mind you) are very different things

>being easy to drive with a steering wheel in a driving game, and being easy to drive a real car with a steering wheel (with internet level latency and packet loss at play, mind you) are very different things

Sure, but I don't think it's that different than driving a real car looking out a monitor sized windshield. You might lose some braking and cornering without your inertial perceptions. But you'd still be able to driver around easily.

Either way I think a speedometer is an assumed input for a self driving car and from that you can calculate almost everything needed related to proprioception.

And they often employ tricks to improve the handling, like burying the center of mass below the vehicle.

Pretty much all simulator games have to cheat. Because most of us can't hop into a Formula 1 race car or the cockpit of an F-22 and drive/fly it with no real training. (Much less race other drivers or fight in a dogfight.)

Traction control software uses touch (knowing how much grip tires have via slip), proprioception (knowing the steering angle), equilibroperception (accelerometers).

Don't forget taste; when you can taste an oncoming car you know you have trouble.

It also takes almost 2 decades to train them (individually), not likely sustainable for a car seller.

Teslas don’t have LiDAR, which is used in all the best self driving systems.

He was talking about humans.

Humans have stereoscopic vision

This is true. But most of the things perceived while driving are outside of the stereoscopic depth perception ability of humans. IIRC that stops around 20 meters or so.

Unless I've been driving very differently from you, most of the things that I care about the precise distance of are well within 20 meters.

And moreover, it's lucky at best if my Model 3 acknowledges a car 20m away on the preview. It struggles with cars a few feet from me on the diagonal at stoplights.

I didn't downvote you, but as I see it human stereoscopic vision is kind of irrelevant.

If you close one eye, do you really feel like your ability to determine depth is significantly impacted?

And even then, the stereo effect is basically negligible at the distances relevant for driving.

A human missing one eye would have no trouble driving, except maybe for the reduced field of view.

Humans don't have lidar either.

And birds don't have jet engine, yet nobody is trying to replicate flight by wing flapping as a bird.

My friend, I have news for you :-)

https://www.youtube.com/watch?v=vBq9NcITh6o

https://www.youtube.com/watch?v=Gtn4PpZEB8I

https://www.youtube.com/watch?v=9R19QQiM5Mw

I hope not because it makes them even more wrong. Humans have a number of different sensors we use while driving.

In which case he’s even more wrong. Humans use a whole mess of different senses while driving, including hearing, the inertial sensitivity in the inner ear, touch to feel vibrations and from the car and the wheels on the road. Plus we have a huge amount of contextual information about the meaning of what we are seeing from life experience outside driving, which no Tesla that currently exists can ever have.

It’s a clever bit of snark, but absurdly wide of the mark. If that’s actuary what the Tesla engineers think, no wonder they’re failing by their own criteria so completely.

The claim isn't that humans can drive as well using the Tesla cameras as they can in person, just that they can. That seems obviously true.

The (not explicitly made) sub-claim is that an AI can make up for the lack of audio, etc. by being smarter than a human and faster than a human, better able to multi-task than a human, and completely non-distractible. That's debatable, but not impossible.

...And furthermore, the neural nets and cameras Tesla uses are vastly inferior to our brains. Just because you can argue that a neural network of some kind uses the same basic structures as our brains doesn't mean that it can come within a light-year of what our brains can actually do.

If you are referring to humans, we have a couple orders of magnitude more connections, our neurons can achieve the same functions at 1 order of magnitude fewer numbers, and our brain has better inductive biases.

Our ‘cameras’ are also so much better than any camera that could be put in a car (by most though not all metrics) that there is no real comparison.

By which metrics are human eyes outperforming “any camera that can be put in a car today”? This seems unlikely in almost any domain, given the much wider dynamic range camera sensors can capture - cameras can see into spectrum we simply can’t (infrared, UV…) and operate at much lower levels of light than a human eyeball while retaining full color vision using really cheap tech. They also don’t get tired or worse with age, or forget to wear their glasses, which is nice.

This strikes me as a pretty odd statement to make, personally!

“There is no real comparison” - for the benefit of the less informed, please make the comparison, assuming you are able.

The human eye can perceive 21 stops dynamic range, much better than regular cameras. Event cameras might solve that issue, but they're not used other than in research at the moment.

Maybe in a single still capture? Let’s not forget cameras easily combine multiple exposures into a single capture for truly enormous dynamic range, and can do so at high frame rates, and can go well beyond 21 stops in doing so. The human eye is stuck with the same ~21 stop range regardless.

If you simply use a pair of digital sensors with a 15 stop exposure offset between them (seems fair - humans have two…), thats ~30 stops in a single shot if we assume best we get is 15 from a digital sensor. Again though with high-speed exposure blending and one sensor this is not really necessary in a lot of cases.

The practical reality is digital capture can exceed 21 stops and you don’t need particularly fancy equipment to do it. Two decent cellphone grade sensors (~12-14 stops) would be enough if you don’t want to do single sensor blends and would work well for real-time video applications.

> https://en.wikipedia.org/wiki/Multi-exposure_HDR_capture

The biggest difference between human vision and cameras is the fovea. Half of our optic nerves are concentrated in the visual area the size of our thumbnail with an outstretched arm. To replicate human vision you have to have a high resolution camera, downsample the image and then grant the AI access to high resolution imagery when requested.

They can also "see" much further than the average human can [1].

The only way I could imagine that we are superior to phone cameras is stabilisation, something that could be resolved with vertical integration that informs the sensors and image processing units about forces being applied to the vehicle (though this is coming from somebody outside the field so take it with many grains of salt).

[1] https://www.youtube.com/watch?v=kKMBx6x-lOI

The claim was impossibility with the sensor suite. It may well be impractical. In the long run, there's no better way to be wrong than claiming impossibility.

If you think humans are equivalent to cameras and artificial neural networks, you are severely wrong.

It is pretty trivial to disprove that: A human could no doubt "remotely" drive using the current Tesla cameras. So the sensor suite might be more challenging than alternatives, but obviously not impossible.

You could however argue that the current computer isn't powerful enough to run the software required for vision based L3/L4.

>A human could no doubt "remotely" drive using the current Tesla cameras.

I have doubts about that. I wouldn't want to remotely drive that without much better cameras at the very least. Advanced planning to be in the right lane or avoid obstacles is hard when the details are just a couple pixels.

It’s a trivial claim, advanced without evidence. There’s no reason to believe that a human could drive a car based only on those limited visual inputs. Humans use all our sense to drive.

And even if it were true, so what? Last I checked Tesla didn’t have software which can perfectly emulated a human agent.

> And even if it were true, so what? Last I checked Tesla didn’t have software which can perfectly emulated a human agent.

GP was arguing against:

> The feature is impossible to implement with the sensor suite that comes with the car.

_For that statement_ it's immaterial whether Tesla does something right now, but rather whether it's in principle possible. Modulo your previous objection humans provide a counterexample to that (very general) statement.

It is plenty possible to implement with the sensor suite that comes with the car.

The cameras on a Model 3 were quite mediocre in 2018. Subaru puts a lot higher quality cameras on their cars than Tesla does.

Then why haven't they?

It's possible to cure cancer, so why haven't we?

It is? Where’s your evidence?

How do you know? It's definitely hard to prove that something is impossible.