Comment by tinco | Hacker News Reader

Comment by tinco | Hacker News Reader

Parent Full thread

So, genuinely curious, how can it happen that a system that's been developed over 12 years, can not deal with a common road obstruction? If something truly rare/dangerous happens I can imagine it not being able to deal. But it seems this sort of situation is not only easy to deal with, it's actually easy to predict.

Maybe not something you could predict in a product with just a couple years of time, with just a couple employees, but we're talking over a decade, and thousands of employees. Oh and in Phoenix Arizona, one of those ultra uniform grid cities the US is famous for.

What's stopping them from having their AI operate in simulations in literally every possible traffic situation? Every location a pot hole can be in, every type of traffic cone arrangement, every kind of jaywalking, every kind of vehicle approaching any kind of street from any kind of location, possible speed, reaction time, everything? How could a government approve any autonomous agent that is not tested in such a way?

My world view gets rocked a little everytime I see a headline like this. I want to believe these companies are ran by the smartest engineers on the planet, but then I read about a Tesla running into an emergency services vehicle, or a Waymo being confused about traffic cones and I can't help but wonder what the heck are they doing?

> What's stopping them from having their AI operate in simulations in literally every possible traffic situation? Every location a pot hole can be in, every type of traffic cone arrangement, every kind of jaywalking, every kind of vehicle approaching any kind of street from any kind of location, possible speed, reaction time, everything? How could a government approve any autonomous agent that is not tested in such a way?

Because this approach has huge limitations, principally that you’re now optimizing against a cost function that isn’t the real world. RL has a place for sure, but limitations.

It’s like saying that you’d expect someone who’s played 10k hours of Call of Duty to be equivalent to a Navy Seal. They’re just really good at playing the simulation, not the real thing.

I'm not saying they should optimize for these simulations, just saying they should test for them.

Also simulations are definitely a part of military training, and if they perform badly in the simulation they'll have to fix that before going on a mission I'm sure. I don't know much about Navy Seals specifically, but I've got friends who built full body VR suits for special forces training.

It's because they're trying to code the problem instead of solving it end to end. As you can imagine if you're trying to code driving scenarios you're never going to finish. Only Tesla and Comma.ai are on the right track. Think Alpha Zero vs Stockfish.

This problem has to be solved end to end. It can be done with just vision as humans can drive with just vision. It's literally just a software, machine learning and data problem. Just takes time

Except didn't Leela Zero (the closest thing we have to Alpha Zero anyone can actually use) beat StockFish, then StockFish was upgraded to use neural networks in a few edge cases and otherwise keep doing what it was doing and now StockFish is back on top? Seems like a Waymo-style approach is actually on top there.

Except Tesla is doing the exact same thing as it's the only practical approach at this time. The only place Tesla is extensively using general approaches is in building a model of the world from camera inputs. Other companies aren't doing that because they get much, much better results than Tesla simply by using lidar.

Reply on news.ycombinator.com