Listen to this article
AI may have been used to generate this audio.
Press Space or K to play or pause, and J or L to go back or forward 15 seconds.In May, Mitchell Hashimoto said he believes entire companies are currently operating under AI psychosis. The thread did 2,105 points and 1,272 comments, and buried in those comments is a fairly precise description of the mechanism.
An executive hears from a peer at dinner that their engineers are all using AI. A mandate appears. Token quotas get raised (one commenter described a $300/day allowance). Usage goes up, because usage is what is being measured. Nobody at any point writes down which problem this solves.
Somebody in the thread called it Dinner Driven Development, which is unkind and about right.
The serious part
Strip out the mockery and there is a real finding underneath, and it is about juniors.
Several engineers described the same thing. A junior ships a working solution and cannot explain why it works. The code passes review because it is correct. The gap only shows up months later when something adjacent breaks and nobody on the team has a model of that system in their head.
The same thing happens whenever you can obtain a correct answer without building the understanding that normally comes with it, so it predates AI. Calculators did a version of this, and so did Stack Overflow. This one is faster and much wider.
What a serious team measures instead
If you mandate AI use and then measure AI use, you have built a loop that cannot fail and cannot teach you anything.
The honest measurements are duller. Did the change ship? Did it survive contact with production? Can the person who shipped it defend the design in a review six weeks later, without the transcript?
That last one is the whole test, and it costs nothing to run. Ask a follow-up question the transcript would not have answered, and the answer shows whether they have a model of the system in their head.
To be fair to the other side
The thread also had people pushing back, and they were right too. On a large codebase, a big token bill can be entirely legitimate. Refactoring across hundreds of files costs that much.
The tell is whether anyone can say what the number bought, whatever the size of the bill.
The thread is here. Go and ask one junior on your team to explain their last merged PR without opening it.
One email when we publish. Research, product decisions, and what teams report back.
