Techbros: self driving cars are inevitable!
Also techbros: prove you are human by performing a task that computers can’t do, like identifying traffic lights.
@goatsarah That's a straw target. They're training them to be better than _all_ humans, not just as good as _a_ human.
AI automation won't come about from a machine being able to identify some specific traffic light, it'll come when a machine can identify more traffic light signals far quicker than an average human, and on average drive safer per mile - not perfect, but better than an average human. That's maybe already happened. But there are always more traffic lights to make them better.
@ingie Ever seen a Captcha that uses Vienna Convention traffic signs, as used in most of the world?
I’d say they’re training them to be better at people who’ve never been to America working out what American street furniture looks like.
And that’s the problem with this sort of AI training. It is almost certainly not optimising for what you think it is.
@goatsarah Not sure what you meant with the American versus Viennan Convention bits. The training is to allow them to learn things that aren't easily trained on standards.
Tho I totally agree with your latter point. It's not really important for automated driving as it's not AI driving, as the AI model isn't in control, it's part of a feedback system of sensors to a conventional autopilot system.
Like in an airplane... which can already land themselves safely if needed.
@ingie I mean that they're all trained on American stuff.
5% of the world population.
So when they unleash the result on the other 95% of us, it's going to have no fucking idea what it's doing, because it doesn't look a damn thing like the training data.
@goatsarah ... but i fully agree, we should never, at least not in any time scale I can see currently solvable, put an AGI actually in charge of driving *anything*. As that's indeed when the optimisation - the alignment problem, as they call it, rears its very ugly head. (an AGI will never have the same philosophical world view as a human)
We haven't got AGIs yet... but that's a very italicised yet. That's that moment we need to really be very very cautious of.
@goatsarah related: I've found Rob Miles' videos on Computerphile very good on the subjects - he's a researcher into such things at Nottingham Uni.
I found his stuff very eye opening to things I hadn't even considered as "threats" in that sense - his research is a good reference of concerns that really need to be made more widely known. Most of his interviews tend to be a little technical, but not excludingly so. I think this is a good example of such worrying concepts:
https://www.youtube.com/watch?v=3TYT1QfdfsM