I'm an AI Luddite[1].

More accurately: I'm ignorant and triggered.

More poetically: I'm Mary Shelly watching the preadolescence of electrophysiology.

As a younger human, I was all in for disrupting all the ancient normality with decentralized open standard internetwork interoperability, net neutrality, strong cryptography, commodity hardware, open source software, open access academic publication, etc. And I knew some things about that stuff.

Now I'm the old man who doesn't understand what the kids are doing that's breaking everything.

I also have to (in a first world voluntary sense not unrelated to this metaproject) write some HaD project log entries about stuff I've been working on so that I can get on with doing more of it. And I was struggling to get my head around how to write about a knot of interrelated stuff that doesn't have an obvious beginning, middle, and end.

So I'm giving AI a go. Claude[2]. No idea what will come of it. I don't think it's saving any time but will attribute that to starting at zero proficiency. So far I've put a fair chunk of time in several sessions over several days into dumping in chunks of backstory up to the initial status quo. 

At this point I guess the benefit is that writing for no human makes it ok to dump out information in the form of walls of text that no human would ever read. Claude then reflects back its interpretation which is remarkably sophisticated but imperfect. I write back corrections to the imperfections and Claude reflects back its modified interpretation which is, so far, good on the second try. 

Here's one instance that struck me as remarkable --[3] i.e. an increment more remarkable than the general remarkableness of LLM outputs:

I had Claude read this, which is as far as I got trying to write current updates on my own. That page includes "(130 mm)²". Claude reflected back "130mm²". I got halfway into writing a clarification about length of sides of a square vs area then thought maybe I should back up and check first...

me:

Do you recognize "²" as mathematical notation?

Claude:

Yes — I recognize "²" as [...clip...]

Given the context, I suspect this is pointing at my earlier use of "130mm² footprint" — where I may have mashed together "a 130mm square footprint" (a square shape, 130mm per side) with "130 square millimeters of area" (which would be tiny, nowhere near mill-sized). Happy to take the correction if that's what happened.

I thought it was remarkably remarkable that the simple question prompted the LLM to recognize the ambiguity that I hadn't mentioned yet, anticipate the not-yet-stated correction, and also correlate the physical-space implications of the notational ambiguity with plausible physical scale of objects associated with the subject of text at a URL that I mentioned. (here the old-man-doesn't-understand problem of trying to think about writing what I'm thinking of when verbs like "prompt", "recognize", "anticipate" misrepresent what LLMs do; "correlate" maybe not so bad but then how does an LLM "correlate" between tokenized text and physicality of things?)

[1] as the word is currently used without reference to any actual history of Luddism

[2] Sonet 5 via claude.ai -- a little exploration at default "Medium" effort; mostly "High" effort on this task so far. "Thinking" enabled.

[3] It appears currently fashionable to disparage em dashes as a "tell" of AI generated text. Bah. Humbug.