Monthly Archives: July 2019

Pluribus Skepticism

Is Facebook’s new poker AI really the best in the world?

Facebook released a paper and blog post about a new AI called Pluribus that can beat human pros. The paper title (in Science!) calls it “superhuman”, and the popular media is using words like “unbeatable”.

But I think this is overblown.

If you look at the confidence intervals in the FB blog post above, you’ll see that while Pluribus was definitely better against the human pros on average, Linus Loeliger “was down 0.5 bb/100 (standard error of 1.0 bb/100).” The post also mentions that “Loeliger is considered by many to be the best player in the world at six-player no-limit Hold’em cash games.” Given that prior, and the data, I’d assign something like a 65-75% probability that Pluribus is actually better than Loeliger. That’s certainly impressive. But it’s not “superhuman”.

I don’t know enough about poker or the AIVAT technique they used for variation reduction to get much deeper into this. How do people quantify the skill difference across the pros now?

I’m also a bit skeptical about the compensation scheme that was adopted – if the human players were compensated for anything other than the exact inverse of the outcome metric they’re using, I’d find that shady – but the paper didn’t include those details.

Thoughts?

Defensive Randomization

Machine learning is common and its use is growing. As time goes on, most of the options that you face in your life will be chosen by opaque algorithms that are optimizing for corporate profits. For example, the prices you see will be the highest price under which you’ll buy, as based on an enormous amount of data about you and your past decisions.

To counter these tendencies, I expect people to begin adopting “defensive randomization”, introducing noise into your decision-making and forcing corporate algorithms to experiment more broadly with the options they introduce to you. You could do this by simple coin flip, or introduce your own bots that make random (or targeted exploratory) decisions on your behalf. For example, you could have a bot log in to your Netflix account and search for a bunch of movies that are far away from Netflix’s recommendations for you.

One possible future is for these bots to share data between themselves — a guerilla network of computation that is reverse-engineering corporate algorithms and feeding them the information that will make your life more humane.

This is related to:

[mildly inspired by Maximilian Kasy’s Politics of Machine Learning]