The "model collapse" framing here is sharp. What's underappreciated is that this isn't just a technical problem about AI needing fresh data, it's basically revealing that value creation itselfmight be fundamentally tied to inefficiency and suboptimality. I ran into something similar when working with synthetic data pipelines last year—the stuff that actualy helps models generalize is the weird edge cases, the things people do for no clear reason. The Culture Ledger concept feels like acknowledging this without turning it into a quota system, which is probably the only way it could work. Tho I wonder if the gardening metaphor undersells the power dynamics—even gentle cultivation shapes what grows.
Thanks for the thoughtful reply, appreciated. Agreed, it is not a technical problem by any means any more - we know how to get data, train against it, work with transformers etc. It's interesting to think that in a model like this, wexactly what sort of human data would be helpful going forward - is it the volume of the common actions or the incidents of the unique ones that bring more value.
The "model collapse" framing here is sharp. What's underappreciated is that this isn't just a technical problem about AI needing fresh data, it's basically revealing that value creation itselfmight be fundamentally tied to inefficiency and suboptimality. I ran into something similar when working with synthetic data pipelines last year—the stuff that actualy helps models generalize is the weird edge cases, the things people do for no clear reason. The Culture Ledger concept feels like acknowledging this without turning it into a quota system, which is probably the only way it could work. Tho I wonder if the gardening metaphor undersells the power dynamics—even gentle cultivation shapes what grows.
Thanks for the thoughtful reply, appreciated. Agreed, it is not a technical problem by any means any more - we know how to get data, train against it, work with transformers etc. It's interesting to think that in a model like this, wexactly what sort of human data would be helpful going forward - is it the volume of the common actions or the incidents of the unique ones that bring more value.