Eliezer Yudkowsky asks about automation | Hacker News Reader

Eliezer Yudkowsky asks about automation

marginalrevolution.com

43 pointsgwern16 comments

I'll cop to being a bit ideological here (mostly in the form of not fully justifying the statements here), but I don't think automation is displacing labor... yet. I think this time that the explanation that we grossly increased regulation, taxation and even more so the fear of future taxation, government spending and benefits, and just in general have significantly raised the both the direct cost of hiring someone, and the reasonable expectations of future cost of hiring someone is a sufficient explanation. I find the idea that raising hiring costs can be done without resulting in lowered hiring to be, shall we say, something that can only be rationalized by a brain in "politics" mode, rather than "rational" mode.

(The test to this theory is that if we elect a regime that tones all of this back, and the economy comes roaring back 1980s-style, then this theory was essentially correct. If we do that, and it does not come roaring back, then my theory here is incorrect. I believe in treating politics scientifically, at least, my personal politics. My highest allegiance is ultimately to what works.)

However, lest this sound like a raw ideological rant, I do think it is near inevitable that either by the 2020s or the 2030s at the latest, that this automation displacement will be a real problem. We're looking at a world where trucking is gone as a viable job. Where fast-food will probably be rapidly in the throes of total roboticization. Where what farming jobs are still available will be rapidly falling to yet more sophisticated robots that have human-like dexterity and ability to detect ripeness. Taxis, gone, in favor of heavily capital-based self-driving car rent pools. I think this whole automation argument is premature today, but it's going to be a real problem in a relatively small number of election cycles. How is anyone who is not suitable for college and has no useful artistic talent combined with enough work ethic to use it profitably going to make it? I don't know.

>(The test to this theory is that if we elect a regime that tones all of this back, and the economy comes roaring back 1980s-style, then this theory was essentially correct. If we do that, and it does not come roaring back, then my theory here is incorrect. I believe in treating politics scientifically, at least, my personal politics. My highest allegiance is ultimately to what works.)

Except that the economy only "roared back" in the 1980s for a narrow subset of the population. If we're on a labor-replacement curve where rehiring and growth are going to just keep getting less and less egalitarian, then an increasingly large majority of the population has the exact same problem as if jobs were just disappearing outright. The "selectivity" only gives you personally a fear more years to bury your head in the sand before you're the one standing in the dole queue.

(Of course, on Hacker News we're even more goddamn privileged, since so many of us are programmers doing AI-complete jobs.)

We'd kill today for any of the 1980s employment numbers during its height. A generally middle-class lifestyle was still possible on relatively low skill jobs, the idea that unions would price themselves out of work was still laughably absurd, etc etc. I really don't think this attempt to project 2010 political rhetoric that far back in time is practical.

It may not have been entirely egalitarian, but that's a chimera anyhow, so I would consider that a very weak criticism. Failing to achieve impossible goals is not a useful criticism.

I recommend reading gwern's response, as it refutes the major question in Eliezer's thesis:

http://lesswrong.com/lw/hh4/the_robots_ai_and_unemployment_a...

When someone asks "what's different now?" the answer is sometimes "thresholds, sweet spots, and tipping points." So often these things are overlooked in so many areas of life; things don't always turn out the same way _because the state of things (environment, economy, culture, laws, development, wathever) is not the same._

So in other words, the theory is that unemployement due to increased automation is still not a systemic problem - it will just take a number of years (perhaps decades) for the economy to readjust. The somewhat obvious corollary to this theory would be that in the meantime, you want to make sure you're employed in the field which is causing the disruption. This is where the fruits of the increased automation will be seen first.

I can accept this theory. Personally, I believe we'll eventually get to a point where the economy will have to do a dramatic shift in the direction of greater taxation of the richest and more leisure time for most people ("most work is automated"), but we might not be close to that point yet. As of right now, these two outcomes are indistinguishable.

I think we'll eventually get to that point as well. We might even need some form of base salary for everyone. But until we get to that point I think we're gonna see a transition to more creative work. Design, art, music, movies, ect. are all things that will probably take the longest for computers to get as good at as humans. Even when computers do get better at creative work, I think there will still be a large market for art created by humans and not some AI.

Or do we just need fewer people in the world? Once much is automated, you no longer need a lot of people to sustain the luxurious life styles of the rich (with fewer people: of everybody).

"Fewer people in the world" solves a lot of employment problems - we're a dime-a-dozen right now.

But at some point in the future, the time it takes for us as a culture to adjust might be drastically out of line with the rate of technological change. It seems like both are speeding up, but if change outpaces the capacity for dealing with that change, we're in trouble.

I don't think we're at that point yet, though. Current societal failures such as unemployment and poverty can still be reasonably explained by historical means, by looking at political and other non-technological events.

So we should just let people suffer, rot and die for a few decades while the wonderful and loving Invisible Hand of the Free Market adjusts us to its brave new world!

</dripping sarcasm>

Here is a counter-example to the general validity of the "it was all right in the past" argument.

Think about the iteration x_n+1 = a * x_n * (1 - x_n)

Usually one rushes to examine a > 2, where there are interesting phenomena of bifurcation, period doubling, and chaos. Go back and look at a < 2. The iteration converges and defines a function f(a).

It is a rather strange function. For 0<a<1 f(a)=0. Then for 1<a<2, f(a)>0. Imagine if this were some kind of theory of technological unemployment, with the parameter a being the level of technology. Technology grows and grows 0.2, 0.5, 0.9, without any problem of technological unemployment, but there is a threshold at a=1, and boom! at a = 1.1 there is lots of technological unemployment.

The most general argument, that technology didn't lead to unemployment in the past, so it will not do so in future, is at heart a claim that functions like f never arise naturally. But clearly they do!

On the other hand, showing that an argument is invalid doesn't help with working out whether the conclusion is true. Indeed, the counter example also serves as a challenge to believers in technological unemployment. There are toy models in which things are zero for a good while, then a threshold is crossed and things start to happen. Is there any model of technological unemployment with a threshold arising naturally out of the dynamics?

Judging by how often this meme gets upvoted on HN, a lot of entrepreneurial-minded people are thinking about it.

My initial reaction, conditioned by years of reading startup news - and it's probably the same reaction many of us have to these articles - is "how would someone make money off this trend".

I'd like to see how someone can make money off reversing this trend of human unemployment.

Meaning to say, can entrepreneurs create profitable businesses by finding ways, or creating markets, for average humans to stay genuinely productive and feel valuable and independent? A world where those basic needs aren't being met isn't going to be safe for anyone, except maybe those rich enough to maintain their own (robotic?) security force.

We see a lot of activity in the "train to be a hacker" arena, but as has been mentioned, probably not everyone who wants to make a living is fit for that type of work.

There are comments in this thread about how it's better to be on the side that's winning than the side that's losing, and there's a sense of resignation.

The reason automation won't replace a human workforce is because it's far easier to program a person than to program a computer.

There's a cost to automating any task: the time spent designing and testing the system. So if it doesn't make economic sense to automate a task, then employment will occur instead.

People are also conservative by nature, so society tends to shun radical new ideas. That conservatism is why it's usually necessary to start your own company in order to put such ideas into practice: almost no one believes it'll work until it's shown to work, so it's rare for big companies to make huge changes to their established business model.

Most new companies aren't huge successes on the scale of Google because success is distributed according to a power law. There are only a few big winners, and big winners are the only companies that can influence a whole society. Therefore it seems mathematically unlikely that most jobs will disappear to automation because most companies won't automate unless they're economically forced to. So even if some work could theoretically be done by a machine, most companies will choose to employ people right up until competitive pressure forces them to embrace automation or go out of business. And that competitive pressure is necessarily infrequent due to the rate of new and hugely-successful companies being infrequent.

> There's a cost to automating any task: the time spent designing and testing the system.

Good point, but, once something's automated, it's automated (and cheaper) forever.

There's a cost to automating any task: the time spent designing and testing the system. So if it doesn't make economic sense to automate a task, then employment will occur instead.

That's true, but the cost of automation is declining. What happens when Watson-like systems are available to the average consumer? What happens when NLP becomes good enough to understand standard English? Eventually the cost of programming a computer will drop below the cost of training a human. What happens then?

If you want to take a more systematic approach then you could look at the ratios of intellectual to manual labour over time (they're not strict categories, but approximations would do) then you could also make an estimate of the rates at which both kinds of labour are being replaced by automation.

In the limit everything turns into fixed capital, but that wouldn't work unless the machines themselves become independent economic actors and displace humans entirely.

Reply on news.ycombinator.com