Large language models have become so ubiquitous so quickly that it’s easy to forget how new they are. I still feel like I’m barely up to speed, but I have friends who are using them daily. And the more I poke at them, the more I use them, so I want to document how I’m using them now.

Asking questions, getting examples

The main thing I’ve been doing, mostly with Claude and ChatGPT, is asking questions about code. I have a little side project where I want to record audio from a web browser, but I’ve never done that, so I asked Claude how it works.

I’m building a web application that will let users record themselves reading a short text passage. I want the user to record a short audio file and save the file to cloud storage.

Claude built an entire React app to answer the question. I poked around, asked to see a SvelteKit version, tweaked things. I don’t know if I’ll use any of it, but it’s useful just to see how it works.

In another thread, I asked about using Postgres in SvelteKit, because I’ve never done that. It gave me a bunch of library choices. Could I have answered this question with a Google search and Stack Overflow? Definitely. But this is an easier way to put the pieces together.

Lists of things

When Joe Biden dropped out of last year’s presidential election and Kamala Harris became the Democratic nominee, I realized she might be the closest candidate to my age (at the time of the election).

Curious if this was true, I asked Claude for a list of all major party nominees and dates of birth. A few rounds of prompts later, I had a spreadsheet, which I could easily check against other sources.

I did something similar when I was curious about long-running movie francheses. In this cases, I used ChatGPT. Once I had a format I liked, I just gave it new franchises as I thought of them. That all went into another spreadsheet.

One thing I learned on that little side project: I needed a way to quickly see if an LLM is hallucinating, so I asked ChatGPT to include an IMDB link for each movie. A few of those URLs didn’t lead anywhere, because ChatGPT had invented sequels. That was a good low-risk lesson.

Translations, in production

I have one place I’ve used an LLM in production: generating translations for DocumentCloud.

When we rebuilt DocumentCloud’s front end last year, most of the site’s existing translation strings became obsolete.

I wrote a small Python script to take each string from the site’s JSON file for English text and run each line through Ollama, using Llama 3.2.

I wanted to use a local model for this, because I was sending each string individually and figured that would be faster. I also wanted to keep costs down.

This took more prompt iterations than I expected, especially around dealing with format strings. But we got there, at least for the romance languages. It helped that I can read Italian and muddle through Spanish and French (especially with Google Translate to help). German, Russian and Ukranian will need a different approach, or a different model, but this is a start.