Becoming a dictator

· tools, llms

The first time I tried speaking to a computer was around 2000, in Microsoft Word. It was hopeless, and I went back to the keyboard.

This post, by contrast, was talked into existence. I dictated to an LLM, it asked me questions, and it drafted the post from my answers. A few weeks ago I wouldn’t have done that. Now I barely type to LLMs at all.

What triggered this was how much of modern work now revolves around LLMs. Ask about anything in depth and you get an answer with several points, so you write a long reply, and get another long answer back. I type quickly, but after a while it’s a slog. There’s a reason we jump on a Zoom call to talk rather than trade long messages in Slack.

What I didn’t expect was the effect on how I talk. Dictation is a stream. Mid-sentence, I can’t say “go back six words” or “forget that”. I can tidy up afterwards, but not while I’m speaking. So I’ve slowed down. I’m more intentional. Rather than ramble to an LLM, I like to pause and think, which sometimes means 30 seconds of silence. Thankfully the LLM doesn’t seem to find it awkward.

A quiet place to work helps. Dictating in a noisy café may not go so well, and in an office I’d worry about distracting people.

Having the LLM talk back is another matter. For technical work I want something I can reread, so the arrangement suits me: I talk, it writes.

I never thought I’d become a dictator, but here we are.

Further reading

  • Simson Garfinkel, Enter the Dragon, MIT Technology Review, 1998. How dictation software reached ordinary PCs, around the time I first tried it.
  • Alec Radford et al., Robust Speech Recognition via Large-Scale Weak Supervision, OpenAI, 2022. The Whisper paper, or everything you always wanted to know about 80-channel log-magnitude Mel spectrograms but were afraid to ask.