Lea Desandre, Claude, Categories …

We have just booked tickets to see the wondrous Lea Desandre at Wigmore Hall on Tuesday 16 September, singing Dowland and Purcell. It should be an amazing concert. If you haven’t ever seen her perform then this candlelit concert from Rouen three years ago has recently re-appeared on YouTube: watch it while you can. (Louise Ayrton, the terrific first violin for Ensemble Jupiter in the film, who is also playing in the upcoming Wigmore Hall concert, read philosophy here at Trinity about ten years ago …)

I’ve been doing some proof-reading on the categories book, now with AI help from Claude as well as ChatGPT (I am temporarily subscribing to “pro” plans on both). In both cases I’m using a prompt like “Can you please proofread a LaTeX file for typos, ungrammatical English, and other mistakes. The file compiles correctly, so I don’t need comments on details of the LaTeX coding.” And then, after I’ve got a first suggested list of actual mistakes, I ask “Do you want to add any other, more stylistic, comments on that file?”.

I’m giving Claude and ChatGPT files that are late drafts which are already pretty polished. Claude is notably easier to work with and presents its feedback in a much more attractive and user-friendly way. While ChatGPT is astonishingly bad at actually accepting files to correct. Repeatedly I have uploaded a file, asked it to report back the chapter title in the first line, and it gets it wrong: it is obviously reading someone else’s file that is on category theory, but on different topics. ChatGPT is highly resistant to being told it is reading the wrong file and keeps telling me I must have uploaded an unintended file. Frustrating. So I have to chunk up a chapter, and simply paste portions into the chat window.

Those irritations apart, ChatGPT perhaps marginally shades Claude. But it is quite definitely worth using both because, except for gross typos like repeated or omitted words, there is almost no overlap in what they report as mistakes. Both over-report supposed errors; but both spot enough real infelicities to be pretty useful. (Their additional more “stylistic” suggestions in response to my second prompt are, however, in each case almost entirely unhelpful — too often, what I get are more like de-stylising suggestions of how to make the prose as banal or flat-footed as possible.)

As I read through again, checking a ridiculously long book from beginning to end, I find can live with the technical expositions and motivational chat, by far the bulk of the book. But I now realize that the author of some of the more conceptual/philosophical commentary in early chapters needs to have a good talk to the author of some of the later commentary. I’m going to need to do rather more work than I had hoped to make the text’s line consistent (overall more correct by my current lights). But first, a needed holiday to recharge the batteries.

The post Lea Desandre, Claude, Categories … appeared first on Logic Matters.