We drive text adventures by Perl script, and we infer Poisson rates backwards
In order to have an LLM play text adventures, we need a way to programmatically drive them. We figure it out! Then we learn how to infer a Poisson rate through just the probability of no event. It's useful, sometimes.
Hello! I hope your week is going well.
New articles
Interacting With Text Adventures Through Perl
If we want to get an LLM to play text adventures, we will need a way to programmatically interact with a text adventure interpreter (the program that runs text adventures). This turned out to be slightly more complicated than I had desired, because text adventures do not reliably signal when they expect user input, but the Perl event system IO::Async made it as simple as it gets.
Full article (10–20 minute read): Interacting With Text Adventures Through Perl
Flashcard of the week
One of those rules of thumb that are easy to forget but sometimes really useful.
If the data are Poisson with rate m, what is P(X=0)?
The answer to this isn't actually the interesting bit. The interesting part is flipping it around. Sometimes we observe a Poisson process, but we are unable to get exact counts. We might only be able to learn whether there were no events, or at least one event, in any given instance. Can we still infer the rate of the Poisson process? Turns out we can!
The answer to the question in the flashcard is
e^-m
and driving this backwards, if, say, 12 % of instances had no events, then we solve the equation
0.12 = e^-m
which gives m = -ln 0.12 = 2.12, i.e. the rate parameter of the corresponding Poisson distribution ought to be around 2.
Premium newsletter
The next premium newsletter will contain at least a review of the non-fiction Fatherhood: A History of Love and Power, as well as fiction The Book of the New Sun – my latest read.
The previous newsletter reviewed Code That Fits In Your Head, a bunch of links and a preview of a future article. To get access to it, a premium subscription is currently only $2/month, cancelable at any time, no questions asked.
That money also gets you the past issues, with a handful of longer book reviews/summaries, 22 brief book recommendations, forecasting/world modeling tips, and more. To upgrade, click the subscription link at the top of this newsletter and fill in your email again.
Your opinions
I cannot improve without feedback. Reply to this email to share your thoughts on any of the topics above, or anything else!Hello! I hope your week is going well.