4 ms·

My experience is that for some people, the physical act of speaking is required for them to think through an issue. Think rubber-duck debugging, but applied to

My experience is that for some people, the physical act of speaking is required for them to think through an issue. Think rubber-duck debugging, but applied to most aspects of their lives. For whatever reason, typing does not seem to serve the same purpose for these people, or doesn't serve it nearly as well. I don't know if the important part is the actual verbalization or if it's the direct, low-latency synchronous conversation with another person, but folks like this strongly prefer and are much more productive in meeting settings. They are also often frustrating to interact with, but that's orthogonal...

wrs 3y ago
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!)
- mwigdahl 3y ago
  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.
teeray 3y ago
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.

wrs 3y ago

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!)

mwigdahl 3y ago
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.

teeray 3y ago

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.