Yoshua Bengio

Unless we hit a wall scientifically, like some obstacle that prevents us from making progress to make AIs smarter and smarter, there's going to be a time when they'll be doing more and more, able to do more and more of the work that people do.

And then, of course, it takes years for companies to really integrate that into their workflows, but they're eager to do it.

So it's more a matter of time than, you know, is it happening or not?

So the jobs that you can do behind a keyboard.

Robotics is still lagging also, although we're seeing progress.

So if you do a physical job, as Geoff Hinton is often saying, you should be a plumber or something, it's going to take more time.

But I think it's only a temporary thing.

Why is it that robotics is lagging compared to doing physical things compared to doing more intellectual things that you can do behind a computer?

One possible reason is simply that we don't have the very large data sets that exist with the internet, where we see so much of our cultural output, intellectual output.

But there's no such thing for robots.

Yet.

But as companies are deploying more and more robots, they will be collecting more and more data.

So eventually, I think it's going to happen.

And by the way, going back to the question of

catastrophic risks, an AI with bad intentions could do a lot more damage if it can control robots in the physical world.

If it can only stay in the virtual world, it has to convince humans to do things that are bad.

And AI is getting better at persuasion in more and more studies.

But it's even easier if it can just hack robots to do things that would be bad for us.

So that's for the national security risks that are coming with the advances in AIs.

C in CBRN, standing for chemical or chemical weapons.