[shituonui]

>Autopilot works great.

I don't think anyone claims otherwise on the older models. The problem is twofold:

1. They promised "Full Self Driving" and delivered ordinary driver assistance.

2. The vision-based approach just plain does not work.

[dham]

> The vision-based approach just plain does not work.

No. Vision is the one and only future for self driving.

Source: Been closely following this scene for years and own a Comma 2 (Open Pilot).

[Syonyk]

So, you're admitting that computer-based self driving cars won't be a thing?

Clever, but probably not what you think you're arguing for...

Silicon Valley has a regular, predictable, very tedious pair of arrogances when it comes to things like this.

1. "We are as gods in the synthetic world of the internet, because we know Code. Therefore we must be like gods in the external, physical world!" It's perpetually on display, and Tesla is an amazing case study here. After deriding those dinosaur, idiot, legacy automakers for their horridly slow, glacial, obsolete factories that mixed humans with robots on the assembly line, instead of being purely robot based and running at sonic velocities... the Model 3 and such are produced on something that's virtually indistinguishable from those "dinosaur" lines. GM, Honda, Toyota... they've all been trying for the "lights out dreadnaught" car factory since the 80s, and have all concluded that it simply doesn't work. It's cheaper, faster, and more reliable to have humans in the loop, and run the line at their pace, than to solve the problems that increasingly look like "full general purpose AI."

2. "A human visual system is just a couple shitty cameras and a neural network. How hard can it be?" The answer is, "Exceedingly hard." The human visual system works at a far better angular resolution in the core than Tesla's system does (the Tesla "narrow" camera has a 4x worse angular resolution than the human eye core, and the human eye is really good at scanning that high resolution core and merging details), and the neural network behind it is an awful lot more complex than anything our computers can emulate. My three year old makes fun of the "green moons" and "red moons" in the sky after showing him the video of a Tesla confusing a low moon for a yellow light, because he can tell the difference between the moon and a traffic light. And I guarantee his visual processing system is capable of an awful lot more than a Tesla, in practical real world situations. I don't trust him to drive anything faster than his Power Wheels, but... well, neither do I trust a Tesla to do that.

A multi-sensor fusion based system has some hope of working with current technology. Vision only? Not a chance.

[dham]

A few different things. IMO we don't really need full self driving, at least not in the next 10 - 15 years. The anxiety that it induces in city streets, pulling out at intersections, etc isn't even worth it. With Tesla your hand is already on the wheel, you're hovering the brake, just drive the car. It's really dumb when you think about it.

All we really need is better adaptive cruise control which is what Autopilot and Open pilot is. IMO it works flawlessly, takes the stress out of driving. It's really a solved problem at this point. I'd say 3/4 of all the miles I've driven on my Rav4 have been on Open Pilot.

[9935c101ab17a66]

You: 'vision is the only future for self-driving.' Also you: 'No one actually needs self-driving anyway! it's too goddamn complicated!'

Uhhhhhhhhhh