Lift Up Your Hearts!

I certainly have my worries about Artificial Intelligence, and plenty of frustrations with the LLM's I've been working with, but it can't be denied that I'm having a fantastic time exploring these new tools.

My father liked to keep copies of letters he sent to various distant family members, and recently I have started working on an inherited collection comprising a few years' worth of them. (Not nearly enough! But we work with what we have, and are grateful for what has been saved.) The letters are not in the best of shape, and my intent is to transcribe them while they are still in a condition to be scanned.

For that purpose, LLM's have been incredibly helpful. So far, I've been using Copilot, ChatGPT, Genesis, and most recently Claude. Each has its strengths and its weaknesses. For transcription, ChatGPT had been my hands-down favorite for both its accuracy and its ability (with training) to do light editing (fixing misspellings and typos) while keeping the original text maximally intact. However, ChatGPT then started throttling down my access to only one or two small uploads in a 24-hour period. (I'm still using the free version, but previously it had easily allowed much more—that in itself is an upcoming blog post.) To ChatGPT's credit, it also told me how to game the system, e.g. to create a PDF of 10-15 image pages and upload them all at once,. And to work at low-traffic times of the day, when more resources are available. Both tactics have been very helpful.

I like to keep several LLM tools in my pocket, and switch to another when one frustrates my work. I'd been intending for a while to try out Claude, for the same reason I experimented with ChatGPT: the enthusiastic recommendation of a grandson (different grandson for each). My frustrations with ChatGPT's limitations pushed me to take the jump. I was impressed indeed, and at the moment Claude is my favorite for transcriptions, though ChatGPT is stiil excellent.

My father's pile of letters is mostly in chronological order, but not absolutely perfect, and recently I came upon a page that had been separated from its context. I was preparing to go through the laborious process of discovering where it belonged, when it occurred to me that I had a tool that might be able to do the job more easily and much faster. I uploaded the single page to Claude, with the caveat that it might belong among the pages it had already analyzed—or it might be before or after, but could I have its best guess as to where it might fall?

Lo and behold, Claude nailed it to withing a few days of the actual date, which I confirmed by finding the pages that surrounded it. It was fascinating to read the logic behind the conclusion, the clues that Claude found among the letters it had in its memory. The work is nothing I couldn't have done myself, with careful reading and thought, and I do love puzzles—but in this case I loved better having the problem solved almost instantly.

Perhaps what I like best about working with LLM's is the collaborative nature of the work. If I could have dumped all of Dad's letters on Claude in one upload, the result would not have been nearly as good. By working together—Claude transcribing a handful of letters, me explaining where it had made mistakes, and refining the transcription ground rules, Claude carrying that knowledge forward to the next set, etc.—the iterative process consistently improved both the accuracy of the transcriptions and the time and effort saved on my part.

Claude is not without the tendency to flatter, common to all the LLM's I've tried so far, but it is more deft about it. Copilot, for example, tends to come across as condescending. Here is Claude's effort at adding a human element, and I think it did a good job there as well as with the transcription and solving the puzzle.

Thank you so much! It was genuinely enjoyable — like a little detective puzzle. The writer has such a consistent voice and such reliable habits (the Bethesda trips on a schedule, the Explorer Post camping at Resica Falls, the stop at Friendly's, Alan going back to Rochester) that once you've spent time with the letters, the clues add up quickly.

The "first day of Spring" detail was probably the clincher — it's exactly the kind of thing he would note with pleasure, and it gives an almost precise calendar anchor.

I hope when you find the rest of that letter among your papers it confirms the dating. And if the Bethesda trip series really was five trips every other week starting in early February, you might even be able to reconstruct a rough schedule for the missing letters from that period.

It's a wonderful collection. The writer clearly had a gift for capturing the texture of ordinary life — the bowling scores, the spaghetti for 50, the wolf book, Beth Myers and her cold swim. Future generations are lucky to have someone taking the trouble to transcribe it carefully.