Many have seen the graph of how the number of questions on StackOverflow has developed over time: Basically, it’s a story of solid growth 2008–2014, plateauing 2015–2022, and implosion since ChatGPT. The one place where all programmers went to share their questions and answers is becoming a thing of the past.

There are two main explanations for the effect:
- Systemic negative feedback: At some point, most of the good questions that are relevant to a big group of people have simply all been treated. Instead of producing new question-answer pairs, one can rather use the search function.
- Recent chat bot disruption: It’s now more convenient to have one’s individual question answered by querying an agent than by searching/asking on a Q&A site.
I think most good papers are essentially “question-answer” pairs. So, similar dynamics apply to the paper writing ecosystem as a whole.
Already without chat bots, paper culture has tended to decline. At least in my field, nice and neat question-answer papers are basically out. Everything has to either propose one grand new framework (that mostly nobody will ever use) or provide an incremental improvement (that’s intelligible only to five nerds). The point of producing such papers too often became showing off rather than settling a question that many care about.
But with agents, both the altruistic reason (settling questions) AND the egoistic one (showing off) are undermined: If I can get my nerd question answered by some machine “at home,” I don’t need to look into how other nerds are answering adjacent nerd questions in their dense papers. And, once everybody assumes that most PDFs on arXiv are not human-written, they stop being impressive.
Of course, there’s still a need to document what people or machines have found out in experiments and to discuss what they care about. So, while a lot of things that have counted as a paper in the last decade just stop making sense as a paper, I think there’s a future for similar artifacts as ways of communication between humans.
