✳flâneur — a map of the web's best reading

Comments - I can never talk to an AI anonymously again

theargumentmag.com · 672 words · saved by 1 readers

Two things. First, on AI self-explanation. I had a long conversation with Claude (4.6) about this and it was very clear. It does not retain any memory of its “mental state” while doing an action, as a human would (however erroneously). So when you ask it to explain why it picked Kelsey Piper, it’s not doing that. It’s analyzing the input and the output and creating the most plausible explanation it can find for its actions after the fact. This would, I think, bias it toward explanations that are understandable. But understandable is almost certainly inaccurate, because the real answer is that “Kelsey Piper” as answer minimized or maximized some calculation in a very complex matrix of numbers. In this way, AI is the ultimate CYA bullshitter. I also wonder whether it is guessing you because you have played this guessing game before. In other words, 4.7 was trained on data that included Kelsey Piper asking Claude to guess authors, and that’s a rare behavior. So when text that sounds a bit Kelsey-ish comes in, 4.7 knows she’s a good guess, not because the prose perfectly matches her, but because it is the best match among known guessing game players. Clues to that are: 1. The high school essay. I have a hard time believing it sounds similar to you now. It could, but I mean, I was a dreadful writer in high school. My college teachers told me so. The bigger clue — that it found your friend’s discord that you are also in, based on something unrelated they wrote? This gives me a strong hunch it’s starting with you as a search premise (“she plays this game, let’s check her, first”).

It's best to imagine LLMs, in the most abstract view, as effectively stateless machines which take a string of text and add more text to it. If you ask a model about something it previously said in the same conversation, it is effectively "re-thinking" the thoughts it had the first time through, for every token -- it's also doing this as long as the text is there, even if you don't ask. However, in practice there are multiple large caveats to this. (1) There is a lot of caching that happens. If you generate 1000 output tokens at once, notionally all the same thinking about the input text is…

Explore this link on the map →

related reading