I started all this by saying PPLs were surprisingly visual and physical, so I'm concentrating on the interaction loop that I had in mind when I said those things: every time you edit a program, you see how it affects the distributions of data flowing through.

So in the attached GIF, there's a dataflow graph for a simple model where every few seconds I add an additional wire.

PPLs don't solve for a result—they converge . So I expected the visualizations to have a very springy response to changes. And that's what I get! In fact, the juiciness of the program output contrasts so much with the stilted UI that I might have to resubscribe to some game design blogs sooner than later. 😬

But this is good. I have ideas for where to take the language that are milquetoast but implementable by me in my spare time, at a pace that I think should allow me to explore complex models within my lifetime.

I've actually been keeping a lab notebook for this project. Tonight I decided to put it online, since it explains some of my design choices at a length that wouldn't be appropriate in Slack. alltom.com/pages/ppl-lab-notebook

I just published a new version of my thing, with focus not on the new version number but on the number of days the previous version lasted. Because I find these numbers in our statements subconsciously go from facts to bragging to metrics..

akkartik.itch.io/carousel/devlog/678890/new-version-after-51-days

I haven't been doing much coding lately (past ~4 months), so I'm doing a tiny side project to test out some new ways of working.

• I'm converting an old CLJS project to TypeScript. I really wanted to authentically live in CLJS, but after months and months of trying, I never got over pain caused by the tooling. (This was back in ~2016, and I hear things have come a long way since. I'd give it another shot, but I no longer have much free time to learn languages/ecosystems.) I tried to get GPT to actually just convert the whole CLJS codebase. It was a mixed success. This is why I chose TypeScript, rather than (my preference) some dynamically-typed language. As it passes through an intentional state of chaotic mush , I want as much structure as possible to be imposed upon my work. Once the conversion is done and I'm ready to resume work on the actual capabilities of the system, I'll probably gradually convert it to JS so that I can get away from needing a build step at all.
• Inspired by @Maggie Appleton, I'm using native apps for Whisper and GPT. When I want to know how to do something, I just ramble out loud. Eg: "Does JavaScript have any syntax yet for iterating over the keys and values of an object without using object.entries?" I've found that for certain kinds of question, it feels really fluid. It might not be any faster than switching to my browser, typing mdn.io followed by my query, and then scanning the page. But it often is, and even when it isn't it just feels really nice to get some help from the computer without having to touch the keyboard. (I have a weird relationship with my sense of touch — for instance, I really dislike touching anything made of metal and think we need to put all the metal back in the ground where it belongs, we've awoken a demon of unhindered conduction and are reaping the consequences, we're all damn fools — so it's nice to be able to converse directly with my computer without having to touch anything)

I've resigned myself to the fact that models in my language are going to get worse before they get better. There are things I love about this representation already, but I can already see it getting mired in "syntax".

I'll get to it. I want to work on model editing so badly that I'm just throwing in first drafts of the language features that I need in order to implement more interesting models as quickly as possible. I'm only a few away.

I added new entries in the lab notebook about why and how, and created this lovely GIF for you. This is a "trait attribution" model that infers whether an exam is fair based on an assumption about the base rate of exams being fair, and how many people pass. (Only one person took this particular test—and they failed. 😰) Cute and small.