Thank you for the Berry post. Has anyone tried to test his hypothesis? My gut feeling is it will check out, particularly for AI chats. Input measure is easy–you have time and text quantity. But output value is more difficult...I'm wondering if there is a good case where this could be observed.
I'm doing a PhD in sociology of work (I have a background in software development) and want to do such a study on software development with AI, but as you say it's quite hard to find a good setting, because it's hard to operationalise quality.
What if you looked at scheduling? Like, an assistant or agent scheduling a calendar something?
It's atomic, and one can't usually add value beyond a low threshold. If we're seeing more words–or tokens–used over time for a task which one cannot reasonably argue has gotten more complicated in that time, it would be direct evidence of involution. If you can relate that to competition for VC dollars or customers, that takes you over to Neijuan.
It's hypothesising a phenomenon. I'm looking for a measure. The phenomenon is intuitively correct, but also suspiciously attractive. Sort of like the bullshit-job (or corrupt-bureaucrat) hypotheses, where every job or public official one doesn't like can be conveniently terms bullshit absent any rigour.
Has it been measured? Or is this more of a qualitative critique?
In western thought, this is mostly related to the concept of involution.
The most famous study is on Javanese farming: https://en.wikipedia.org/wiki/Agricultural_Involution
In terms of software development, here's a blogpost from 2025, by a professor at Sussex Uni: https://stunlaw.blogspot.com/2025/07/involution-and-artifici...
There are also some studies in China on the concept but I'm less familiar with that literature.
Thank you for the Berry post. Has anyone tried to test his hypothesis? My gut feeling is it will check out, particularly for AI chats. Input measure is easy–you have time and text quantity. But output value is more difficult...I'm wondering if there is a good case where this could be observed.
I'm doing a PhD in sociology of work (I have a background in software development) and want to do such a study on software development with AI, but as you say it's quite hard to find a good setting, because it's hard to operationalise quality.
I have been thinking of measuring it along these lines perhaps... https://www.gitclear.com/the_ai_code_quality_maintainability...
But they have a closed sourced software so I would have to find or create an open one in that case.
Happy to hear if you have any ideas on study design!
What if you looked at scheduling? Like, an assistant or agent scheduling a calendar something?
It's atomic, and one can't usually add value beyond a low threshold. If we're seeing more words–or tokens–used over time for a task which one cannot reasonably argue has gotten more complicated in that time, it would be direct evidence of involution. If you can relate that to competition for VC dollars or customers, that takes you over to Neijuan.
It isn't pointing at particular system to blame.
But if you feel like it's raining, it probably is.
> It isn't pointing at particular system to blame
It's hypothesising a phenomenon. I'm looking for a measure. The phenomenon is intuitively correct, but also suspiciously attractive. Sort of like the bullshit-job (or corrupt-bureaucrat) hypotheses, where every job or public official one doesn't like can be conveniently terms bullshit absent any rigour.
There's a Greek proverb disputing the latter: they are spitting on him and he thinks it's raining.
"1920s style crisis of overproduction with Chinese characteristics"