Tay, AlphaGo, and News Feed Algorithms
AI reached two milestones last month: beating the top human player in the world at Go, and creating a “teen girl AI” that “became a Hitler-loving sex robot.” As an enthusiast in chess, I was under the impression that Go was way more complicated and that it would take decades from now for a computer to outplay a human. Oops. The future is here. It seems pretty awesome, but other than that, I feel unqualified to speak further about the Go match.
I do feel more than qualified, however, to talk about Tay. I mean, I’m not an expert on Hitler-loving sex robots, but I have spent a lot of time arguing with people on the internet, and wow do things derail quickly. So my obviously hindsight reaction is Come on, how did you not expect this? Like, have you ever seen the comments section of a YouTube video? If you train on that, was any other result even possible? My solution is, find anyone who has made any angry comment on any YouTube video, and hire them to break your AI during testing. Also, when I imagine what people in the 1990s worried about in terms of AI development, I’m sure this was not one of them. Thanks, internet.
One more thing about AI: Facebook and Twitter news feeds. I’m generally in favor of any change that makes me have to do less work, such as scrolling past posts which, according to some algorithm, I probably don’t care about. However, the paranoia is that if I “like” an article titled, “Microsoft deletes ‘teen girl’ AI after it became a Hitler-loving sex robot within 24 hours,” I would be horrified if Facebook thought I were interested in Microsoft, teen girl AIs, and Hitler-loving sex robots from that action and then show me posts about them. Apparently, this worry is already prevalent:
In Karahalios’ study, many people voiced a common Facebook complaint: too many baby photos in their feeds. They said they would Like a friend’s baby picture out of a feeling of obligation, but then immediately hide the post to try to tell Facebook they didn’t actually want a feed full of toddlers.
So the solution is to like the story but then immediately hide it? I’ll remember that for the day when Microsoft releases a schoolgirl chatbot in the culture that is Japan.
Oh wait, that already happened.
Stocks and College Tuition
When a stock price is too low, you buy, and when it is too high, you sell. That is the most basic thing about a market. Of course, there are a million reasons why this is not so easy to do and why there is an entire sector of the economy trying to do this. And of course, as a disclaimer, nothing on this blog is ever financial advice, even “buy low, sell high.”
Now the “buy low, sell high” strategy may be even more difficult to do in things outside of stocks. One side is pretty easy to do: if groceries or cars or houses are being offered too low, buy them.
But the other side is tricky. If your grocery store is selling oranges at $50 per orange, and people are actually buying them, you probably want to sell oranges at $45 to compete with them. But you need oranges! One way to obtain oranges is to buy them for a cheaper price elsewhere, and resell them at a higher price. So you could go to a nearby city, buy 100 oranges for $1 a piece, pay $100 in transportation fees and resell them for $45 a piece, for a net profit of $4,300. But if oranges are also selling $50 in the other town, you can’t do this, so you would need to grow your own oranges, and that takes some effort, but may be worth doing depending on how much demand there is for these oranges.
The same is true for cars and houses. If the prices for them are just too high overall, the right thing to do may be to found a company that produces oranges or cars or houses and sell them at exorbitant prices. By creating competition, you are also ever so slightly lowering prices to be closer to the fair price.
During the housing bubble, if you thought prices for houses were too low, you would buy houses hoping to flip them out at an even higher fair price. But if you thought house prices were too high, that bet was more difficult to make. You could buy a financial instrument called a credit default swap on subprime mortgage bonds, or you enter the competition by building houses for a relatively low cost and selling them at very high prices. Both options were difficult and risky.
Enter college tuition [Bloomberg]:
No longer is it the rent that is too damn high, but the college tuition. It has become a political issue now, with politicians in both parties decrying the cost of higher education. Among the 2016 presidential candidates, Bernie Sanders has made a particularly big deal about this, going so far as to propose universal free college tuition. Even Donald Trump agrees in spirit: “That’s probably one of the only things the government shouldn’t make money off – I think it’s terrible that one of the only profit centers we have is student loans.”
How are markets supposed to work again? Do you buy when prices are too high? No, you sell! That’s how you both make money and help drive down prices to some reasonable level. The simple theoretic solution is to found new universities. Unfortunately, the for-profit college idea has empirically been a failure so far.
I’m still hopeful for a better solution. Nonetheless, I’m glad I have already gone through college. My alma mater currently costs $67,613 a year.
Not-Often-Talked-About Sources of Income Inequality
My news feeds on Facebook and Twitter seem more political than before, which is unsurprising given the proximity to the presidential election. At least on the Democratic side, there is much talk of economic inequality.
I roughly agree with this Paul Graham essay written a few months ago. It starts off the contradiction that startups seem to both help the world and increase inequality:
I’m interested in this topic because I was one of the founders of a company called Y Combinator that helps people start startups. Almost by definition, if a startup succeeds its founders become rich. Which means by helping startup founders I’ve been helping to increase economic inequality. If economic inequality should be decreased, I shouldn’t be helping founders. No one should be.
It then talks about how there are some good sources of inequality (startups, variation in productivity) and some bad ones (tax loopholes, high incarceration rates), and we should be focusing on the latter group, not on inequality categorically.
Besides startups and productive gaps, what are other good sources of inequality? The one that came to mind was assortative mating. Basically, if two rich people married each other and two poor people married each other, you have household inequality, but if they cross-married, you have household equality. The former is becoming more prevalent. Not only does this increase immediate inequality, but it also decreases economic mobility by denying poor people from marrying up.
Tyler Cowen thinks this is nontrivial [NYT]:
These matches are great for those individuals who can build prosperous and happy family alliances, but they also propagate inequality across the generations. Of all the causes behind growing income inequality, in the longer run this development may prove one of the most significant and also one of the hardest to counter.
And more sentences here:
As it becomes harder for many people to “marry up” as a path for income mobility for themselves or their children, families that are not well connected may feel disengaged, and the significant, family-based advantages for some children may discourage others from even trying. The numbers show that assortative mating really matters.
One study indicated that combined family decisions on assortative mating, divorce and female labor supply accounted for about one-third of the increase in income inequality from 1960 to 2005.
Will the fight against economic inequality be so fervent that, in the future, startups and assortative marriages will be shunned? It would be a strange world to imagine.
I’m a space geek, but someone definitely spent too much effort making this Pluto and Charon video [Business Insider]. Also, when it states that Pluto and Charon are tidally locked, it is one of the times when the animation does not show them tidally locked.