Where Does Initiative Come From?
You’re minding your own business when you suddenly feel like reading Viktor Pelevin. Why Pelevin? Why now? On an ordinary afternoon it hardly seems like a question. You get up and grab the book. Try building that into a program, though, and the first thing you have to decide is what gets it started.
Anyone who’s written a scheduled job knows one way to do it. Have a function wake the agent at a random time and invite it to pick something to do. Leave the choice entirely to the model. It might even decide to read. My code still picked the moment it got a turn.
Take away the suggestion. Give the agent access to books and old conversations. Run it occasionally and let it do nothing. Now I’d like to know why, of all the things it could have done, it picked a book. Maybe the author’s name came up in yesterday’s chat. If we can find that line and rerun the experiment without it, we’ll have the beginnings of an explanation.
In an earlier note about initiative, I wondered whether AI could get by without that first nudge. Thinking through the program, it’s easy to overlook the same difficulty on the human side. How do I know the urge to read arose all on its own? I might have spotted a familiar cover this morning and forgotten about it half an hour later.
With an agent we have a rare luxury. We can go back to the same state and remove one line of conversation. I can’t rerun that afternoon on myself. I picked up the book. Good luck figuring out exactly why it was that one.