In my conversation with Dmitry Merkulov I admitted something that still surprises me. I have the most advanced technology in the world in my hands, it does the work a hundred times faster, and I talk to it worse than to any living person. In fragments, irritated, with a line like "come on, just figure it out yourself".
This article is about why that happens and what I do about it. The first article in the series, an overview of the whole conversation, is here: How to think about AI.
Video
This part of the conversation starts at 20:16 in the English version. The Russian original starts at 20:20.
Crude, lazy and annoyed
All my life I read books, learned poems by heart, spoke at negotiations and on camera, and I thought I had a way with words. Then I noticed how I phrase tasks for a model. Crudely: "you get it, do it yourself", "make the call, remove the bad, keep the good". Sometimes I get angry at it, even though I know perfectly well it is a machine and being angry at it is as odd as being angry at a calculator.
What struck me most was the laziness. A task that used to take a month can now be done in a week, and all it takes is to pull myself together and say clearly what I want. But my brain resists, and the road gets longer again. Sometimes I am too lazy even to read what the model answered. It writes that it needs my judgement and my decision here, and I do not want to look.
Dmitry called this atrophy, and my reflection on the atrophy. I will not pretend it does not apply to me: many people do this, and so do I. Any tool that makes life easier gives us an excuse to stop doing what used to develop us. Dmitry's example was the car: our parents could keep a garage five kilometres from home and nobody minded, and now we park right at the door. In his view most of the downsides of any technology are about us, not about the technology itself.
The answer you got is the question you asked
Then Dmitry said something I now repeat to everyone I know. At the School of Great Books they have long said: the answer you gave me is the question I asked you. He has been using AI deliberately for more than a year, and he sees that most problems in ordinary tasks have nothing to do with the model and everything to do with how he asks. If the answer is wrong, the question was wrong.
It works with people too. Dmitry's example: if you ask your boss for a raise and he gives you more working hours, it is worth thinking about how exactly you asked. With a model today it is exactly the same.
Get used to each other instead of writing one perfect prompt
Friends often come to me, people who are professionals in their own fields, and complain: I installed it, loaded a big prompt, and the result is rubbish. My answer is almost always the same: everything you just told me over two minutes, tell it, and keep the conversation going.
A good, detailed prompt in a fresh chat will of course give a more sensible result than "make a beautiful website". But I have long felt that the point is not one perfect prompt. I talk to the model like a colleague, in an ongoing dialogue, and over time we get used to each other: I understand how it works, and it gets more and more context from me. I tell newcomers exactly that: you have only just met, you need to talk for a while.
Dmitry put it in philosophical terms. Getting used to it means understanding how it works, then handing the machine the dull work you do not enjoy and freeing up time for the work you do.
Voice input as speech practice
There is a good side to this story too. I used to speak in public a lot, decided I had mastered it, and did not practise for five years. It turned out the skill fades like guitar or chess if you stop playing.
Voice input brought that practice back. I dictate tasks out loud and every time I relearn how to put a whole thought into words. I live between three languages, Russian, Croatian and English, and at one table there can be a Spaniard, a Canadian, an American and a few Croatians. Switching between languages and still speaking clearly turned out to be exactly the skill that talking to a model trains.
What I do now
- Before a task I say out loud what I want to get and how I will know it worked. By voice it is faster than typing, and the thought comes out whole.
- If the answer is bad, I first ask myself what I actually asked for, and only then blame the model.
- I do not try to fit everything into one prompt. I start a conversation and refine as I go.
- I read what the model answers, especially when it directly asks for my decision. The decision stays with me anyway.
The next article in the series is about the blind archer strategy: why I run many projects at once and how a website about robots, built with no goal at all, now gets around 2,500 visits a week.
Originally published at ilinmaks.com.
Top comments (0)