Sample unit engineering, AI prompting, and poker
Where we learn some prompt engineering, some sample unit engineering (which is how to count things extra cleverly) and some non-engineering: a little about betting in poker.
Good morning! (Or whatever it is in your time zone.)
New articles
My Most Complicated Prompt So Far
Sometimes LLMs still surprise me in what they can do when it comes to munging symbols. This is one of those cases. Free advice on prompting for anyone who has – like me – lived under a rock for the past year.
Full article (1–4 minute read): My Most Complicated Prompt So Far
Sample Unit Engineering
The accuracy of an estimation derived from a random sample is often perfect; the problem lies in precision. It is common to believe that precision can only be improved by increasing the sample size, but this is not true. We will see an example where, by selecting sample units cleverly, we are able to double precision without increasing sample size at all!
The intuition around why that works is fairly valuable when reasoning about other statistical problems too. Hopefully the unicode plots in the article convey some of that intuition.
Full article (10–30 minute read): Sample Unit Engineering
Flashcard of the week
In The Theory of Poker, which I have started reading a couple of times but never finished, Sklansky notes that
When you have a good hand you should almost always bet, except in which two circumstances?
The answer hints at the counter-intuitive nature of poker.
You should not bet when the pot is tiny and/or your hand is and will remain the nuts.
Why not bet into a small pot? One of the effects of betting is shifting the conditional distribution of our opponents through selection. If we bet into a small pot, most of our opponents will fold (they have little to gain and a lot to lose). The ones that will not are the ones that think they have a good chance of winning. By betting into a small pot, we are eliminating the weak opponents and selecting the toughest ones. That's not what we want when we have a strong hand, because we put ourselves in a situation where it is very easy for us to make a mistake.
Why not bet when our hand will remain the nuts? (The nuts means the strongest hand at the table.) This is easier: because we want as many opponents in the pot as possible. If we know we have the strongest hand, we don't want to dissuade anyone from throwing in more money.
The reason we bet in every other case is to invite our opponents to make mistakes. The amount we should bet, according to Sklansky, is just enough so that our opponents make a mistake if they call, because that means we're happy either way: if they fold, we get the pot. If they call, they made a mistake and are costing themselves money in the long run, regardless of how this particular hand plays out.
Would you mind if I ran surveys?
Buttondown has survey support, which I am curious about trying at some point in the future. I'm convinced my readership holds a lot of uncoordinated knowledge that could be useful in aggregate.
But first – is that something you'd care for in the first place?