Talking to AI is terrible

If you're a software engineer, it's likely that you're using AI for your work many hours every day. And the way you do it is writing or dictating prompts and reading outputs. If you're one of the more conservative engineers, you're probably also watching what actions it takes and what code it writes.

Every time a new model releases, you feel how good it is compared to the previous one, how different it is. Then a week or two passes. You've used it for 8 hours every day since then, and suddenly it feels absolutely bad.

The model doesn't understand what you're talking about. It does things you did not want it to do. It comes up with ugly solutions and designs. It can't solve the problem you gave it. And you're so fed up with its writing style and very specific tics.

Why does that happen? Well, humans remember things and update their behavior based on feedback day to day. Agents are much worse at that today. You can add a new line to AGENTS.md, but that doesn't solve new problems automatically, and old instructions make newer models worse. An agent could write a memory for itself, but today's memory systems are still far from working well.

And just as important, you've talked too much with it every day. Imagine continuously talking as a boss to your employee directly every day for 8 hours. You'll get tired of that person too. You'll start noticing their weird habits, traits and tics. And those things don't get instantly "fixed" in humans either.

So, how do you solve it?

The solution I see today is just talking to AI less. As little as possible. If you're a software engineer, try to delegate more instead of babysitting small steps. Watching every step your agent takes is probably the biggest reason why conservative engineers still have a bad opinion of using AI for work.

If you're building a product with AI under the hood, try to make it invisible. Move the steps and thinking to the background. Work on prompts and personality so that users don't see walls of filler slop. Make AI feel like a magic genie rather than a hyper-fixating autistic employee.

I expect memory to be mostly solved by the end of this year, but this still won't make it much nicer to talk to AI for 8 hours a day. It'll still feel terrible. So, don't get used to how it works now. Prepare for AI being even more autonomous and intelligent.