Comment by beat | Hacker News Reader

Comment by beat | Hacker News Reader

According to a libertarian friend of mine, if you raise the minimum wage, jobs will get offshored or replaced with automation. I told him good luck with his new robot McDonald's in India.

Basically, any job that can be automated will be automated, as soon as it's feasible. This has been going on for millennia, ever since the first genius learned to herd goats rather than hunt them. People just have a very narrow definition of "automate". We don't plow our fields or harvest our crops by hand anymore, leading to massive reductions in farm employment over the past 100 years. But we don't think of tractors as "robots", so it gets ignored.

I have this saying that drives my whole career's thinking... "Have computers do what computers do well, have humans do what humans do well." Computers are wonderful about performing precision tasks perfectly, at incredible speeds. But they are terrible at making decisions in the face of incomplete data. They can be programmed to sort of simulate it, but it's not the same. Meanwhile, humans are incredible at making correct decisions using incomplete or incorrect data, but it's almost impossible to get them to do the same thing the same way twice in a row.

So the jobs that will last are the ones that require decision-making in the face of incomplete/incorrect data. The jobs that will be automated are the precision, repeatable tasks.

I think you overestimate what humans are capable of doing (comparatively) well. By your definition, computers should be terrible at Air Traffic Control, Driving Cars, Logistics Planning, and even playing Chess. But computers are incredibly advanced at those tasks...performing well above their average human counterparts, and in some cases even better than the best human specialists.

Humans have developed a large skillset through evolution, and it is a skillset that includes highly sophisticated heuristics for problem solving, communication, and physical interaction. Even 15 years ago, the top Ops Research and Artifical Intelligence researchers and their computer models couldn't outperform an ordinary truck driver on the Traveling Salesman problem on average. Now, there isn't a human in the world that can outperform their models. 15 years ago, there wasn't a single robot in the world that could compete with an ordinary human in the task of walking. Now we have Boston Dynamics.

The idea that humans are special at making decisions with incomplete data is short sighted. It assumes that computers can't improve at the task. They are certainly a far ways away from making general purpose decisions with incomplete data, but there are several current systems out there that make very sophisticated (even unparalleled) decisions in specialized domains, and they do it with incomplete data.

That isn't to say that your saying is wrong. It is just to say that it isn't a given that something that humans do comparatively well today may not be that way in the future.

> By your definition, computers should be terrible at Air Traffic Control, Driving Cars, Logistics Planning, and even playing Chess.

All of these are problems with complete information.

Chess is a solved game, however you can't analyze the whole problem quickly enough to cover it all.

Driving a car, a computer can have access to all relevant data--far more than a human. Sure, new data can be introduced, but the computer can pick it up and handle it as quickly as a human.

Poker is a better example of a problem with incomplete information--and computers do NOT excel at it. They do play excellent heads-up limit, which is more-or-less a solved game. They play much-less-excellent heads up no-limit, and they certainly don't approach new situations in a novel manner. The same goes for full table games.

I would expect computers to be very good at all the tasks you mentioned, and mediocre (but trainable) at those like poker. I expect that this, too, will change in the future, when we manage to invent better ways to handle incomplete information, especially if there's a rapid feedback loop.

I think those problems don't have complete information.

Air Traffic Control employs certain models of behavior of planes and air flows, which are not exact. Understanding of situation could be different among participants, so are actions. Driving Cars doesn't have information about all surroundings (say, behind the bush) - and definitely not about plans of other agents in the area. Logistics have to deal with uncertainties of transportation - you can't be sure a particular truck going through Oregon won't be delayed by a storm. Chess isn't a solved problem (checkers is).

Computers do not play chess well. They play chess badly, but incredibly fast. Deep Blue was capable of analyzing 200 million board positions a second, but it could be defeated regularly (if not a majority of the time) by a human being. I'm pretty sure Kasparov wasn't calculating 200 million board positions in his head!

Pick a harder problem - the game of Go. Computers have only recently been able to beat a reasonably talented child or inexperienced adult at the game. When the data set becomes too large to brute-force (Go has many more possibilities than chess), computers can't hold a candle to human intuition.

Sure, it's possible that at some point, computers will be able to produce better results than humans at some surprising real-world tasks. But that doesn't mean computers will do them well. Doing it well versus doing it very fast are two entirely different things. If you can't understand the distinction, you can't understand where humans will continue to outperform computers.

Deep Blue was using search-tree algorithms, not learning algorithms. Has anyone tried putting modern machine-learning engines up against professional chess players?

Given the incredibly well-documented history of professional chess, learning during play isn't efficient. If you're going up against a player, just program every game they've ever played professionally into the system to study their behavior. You can be sure the human players do something similar.

Heck, I'm fairly certain that every master game ever recorded was shoved into Deep Blue and other high-end chess programs.

What does it matter if Kasparov had sophisticated heuristics that let him bypass considering a few hundred million positions? Deep Blue won, even with its dumb brute force approach. A self driving car may have to use billions more calculations per second in order to drive a car than I do. And yet, they have driven hundreds of thousands of miles without accident, which is more than I can claim for myself.

The crazy thing about human heuristics is that they can be learned. Deep Blue was created more than 15 years ago, but in the time since, Houdini can do even better than Deep Blue, and it can do it on a dual core laptop. How? By learning heuristics from Grand Masters.

> making correct decisions using incomplete or incorrect data

If there's a problem that fits the description it's medical diagnosis , and it's been known for quite some time that machines are better at this.

Reply on news.ycombinator.com