I wonder if using a low-latency conversation with a reasonably effective LLM as a “front end” for documentation will be a big productivity boost for folks like this.

I have been using ChatGPT as a rubber duck (who is also a pathological liar!) for working through code issues with good success. What if you just ramble on for five minutes and ask it to summarize what you just said in README form? (Hmm, I’m going to try it!)

LLMs already make amazing rubber ducks! I would bet you're right -- when we have really good voice UIs for LLMs so it's even more like having a real conversation there are a lot of people who are going to start unlocking considerably more value from them.

I've had some success using Gemini 1.5 to take a recorded Teams meeting of a debugging session (with screen share), extract an audio transcript with Whisper, upload _both_ the video and the transcript, and get a summary of what was done. I'm still working on how to get the right amount of detail and organization without losing the high level flow, but even in a basic state it's better than anything I've had previously.

This. Some people are just not great writers, but when you ask them “what do you want to say?” they say it well. The next step, “okay, just write that”, never happens. They just use the 50-person all-hands weekly sync to do the saying part.

What we really need are something like internal podcasts.