Hacker News Reader

Comment by mitthrowaway2 | Hacker News Reader

It's a matter of time. White collar professionals have to worry about being cost-competitive with GPUs; blue collar laborers have to worry about being cost-competitive with servomotors. Those are both hard to keep up with in the long run.

View on HN

The idea that robots displace workers has been around for more than half a century, but nothing has ever come out of it. As it turns out, the problems a robot faces when, say laying bricks, are prohibitively complex to solve. A human bricklayer is better in every single dimension. And even if you manage to build an extremely sophisticated robot bricklayer, it will consume vast amounts of energy, is not repairable by a typical construction company, requires expensive spare parts, and costs a ridiculous amount of money.

Why on earth would anyone invest in that when you have an infinite amount of human work available?

> but nothing has ever come out of it

Have you ever seen the inside of a modern car factory?

A factory is a fully controlled environment. All that neat control goes down the drain when you’re confronted with the outside world—weather, wind, animals, plants, pollen, rubbish, teenagers, dust, daylight, and a myriad of other factors ruining your robot's day.

I'm not sure that "humans will still dominate work performed in uncontrolled environments" leaves much opportunity for the majority of humanity.

Factories are highly automated. Especially in the US, where the main factories are semiconductors, which are nearly fully robotic. A lot of those manual labor jobs that were automated away were offset by demand for knowledge work. Hmm.

> the problems a robot faces when, say laying bricks, are prohibitively complex to solve.

That's what we thought about Go, and all the other things. I'm not saying bricklayers will all be out of work by 2027. But the "prohibitively complex" barrier is not going to prove durable for as long as it used to seem like it would.

This highlights the problem very well. Robots, and AI, to an extent, are highly efficient in a single problem domain, but fail rapidly when confronted with a combination of them. An encapsulated factory is one thing, laying bricks, outdoor, while it’s raining, at low temperatures, with a hungover human coworker operating next to you—that’s not remotely comparable.

But encapsulated factories were solved by automation using technology available 30 years ago, if not 70. The technology that is becoming available now will also be enabling automation to get a lot more flexible than it used to be, and begin to work in uncontrolled environments where it never would have been considered before. This is my field and I am watching it change before my eyes. This is being driven by other breakthroughs that are happening right now in AI, not LLMs per se, but models for control, SLAM, machine vision, grasping, planning, and similar tasks, as well as improvements in sensors that feed into these, and firming up of standards around safety. I'm not saying it will happen overnight; it may be five years before the foundations are solid enough, another five before some company comes out with practically workable hardware product to apply it (because hardware is hard), another five or ten before that product gains acceptance in the market, and another ten before costs really get low. So it could be twenty or thirty years out for boring reasons, even if the tech is almost ready today in principle. But I'm talking about the long run for a reason.

Reply on news.ycombinator.com