Things learnt Y2026 W34

[0xe2, 0x80, 0x94] is an em-dash (—) in utf-8

Now I know. I have been seeing this in markdown files in terminals decoding ascii by default (kind of rare now, I grant you). Definite hallmark that it was written by AI. I will add this to my lexicon of byte sequences to know, similar to [0x0d, 0x0a] for line-feed, carriage return in http headers and [0xEF, 0xBB, 0xBF] for the byte order mark at the start of utf-8 encoded files… ([226, 128, 148] is — in base 10)

Docker’s sbx puts a kind-of-crappy AGENTS.md file in the parent folder of the sandbox

There is an AGENTS.md file that appears in the context of agents in the docker sandbox which is placed at the directory above the one you mount. There doesn’t seem to be any way to stop this from being put there, having played with the aiFilename settings on the docker sbx kit. I really hate things that be infecting contexts externally, particularly this one that explicitly tells the agent to update the file with changes that it thinks are appropriate (😱) meaning it may constantly change… These systems are stochastic enough, we don’t need more randomness with their system prompts changing (the reason I use Pi as my agent harness. I will need to put something in docker entrypoint.sh to find that supplied AGENTS.md file and replace it with a static version of my own design.

Revisiting topics from a new perspective can be very useful

More reinforcement of a belief than something new, but anyway… I borrowed this book on Bayesian statistics from a good friend over the weekend. A lot of this came up in my degree and I do use the concepts variously, but going through basics from a different viewpoint is always useful. Funny things I’ve internalised, like the relationship between odds and probabilities, is covered early on and provides a new way of looking at stuff that cements knowledge.

In particular the description of combining non-mutually-exclusive¹ events struck me as never having been taught to me as removing double-counting before. 🤷‍♂️

$$P(a \cup b) = P(a) + P(b) - P(a \cap b)$$

Clankers summaries as HTML = 😁

If you’ve prompted them to give an overview of a codebase… Or anything, really!… Get them to output that overview as an HTML file and serve it to you, This is much nicer than scrolling back through the session or reading through a bunch of crappy markdown files.

1. Mutually inclusive? that’s not a thing… ↩︎