Computing & AI · the experiment
Keep some learners
Language models fed on their own writing, a paradox from cultural evolution, people who ask instead of finding out. Four questions, from beginner to expert, on who pays for the knowledge models copy.
1 · State Rogers' paradox
Beginner
What does Rogers' paradox say about a population where copying spreads?
2 · Compute a plateau
The rule: knowledge kept at equilibrium q* = (1 − ρ) / (1 − ρ(1 − u)), where ρ is the synthetic share of the training data and u the chance a copied item is lost per generation. For Darwin, u = 0.15.
Intermediate
With half fresh data (ρ = 0.5), what share of rare knowledge does the lineage keep?
3 · Find the model's accuracy
In the paper's model, finding a fact out for yourself costs c = 0.3, out of a payoff of 1. Asking the AI is free, but right only as often as the AI is. People switch to whichever pays more.
Expert
At equilibrium, how often is the AI right when people use it?
4 · Count the learners needed
The share of people who must keep learning for themselves grows as 1/(1 − s), where s is the share of the training text written by the models themselves.
Expert
If s goes from 0 to 0.5, what happens to that share of learners?
Learners kept
—