Writing With Copilot

A couple of weeks ago, Simon tweeted:

Anecdata: I’ve found myself blogging more since I turned on Copilot for markdown files - am I seeing a 10% increase in productivity with my writing?

Hard to be sure, but it certainly feels like I’m investing less effort for results that I’m genuinely happy with

After that nudge, I decided to give writing with Copilot a try. After using it for a bit over a week, I agree with his conclusion. I do feel like I’m mildly more productive. 10% seems reasonable.

I rarely if ever use exactly what Copilot suggests. Most of the time, I don’t even accept its suggestions. But often when I’m not quite sure how to word my next thought, there’s enough of a spark of what I’m trying to say in Copilot’s suggestions to unstick my thoughts.

One place I’m finding it works especially well is when writing lists. Usually, if I write one or two examples, the next bullet point Copilot suggests will be pretty good. Sometimes it’ll suggest an example I hadn’t thought of! That doesn’t happen often, but when it does, it’s neat.

I’m still incredibly ambivalent about the ethics of Copilot specifically and the modern crop of AIs (GPT-3, DALL-E, Midjourney, StableDiffusion, Whisper, …) more generally. These systems are as great as they are because they’ve ingested a massive amount of training data without any sort of consent from creators. The fact that nobody working on creating these AIs seems to give a shit about consent gives me hives. I may still become an “AI vegan”. But I’ll likely keep using Copilot for a bit while I grapple with the ethics.

(This post started as a tweetstorm, but I wanted a permanent home for it so I copied it here, and expanded just a little bit.)