Perintö

Small learners are born, explore a layered world for one short childhood, and die. Before they explore, each hears twelve sentences from an elder. The world has laws: in each layer one move leads down from most places. A learner can notice a law in what it has seen, and elders can say it. Watch whether a few spoken laws carry a world that no single lifetime can map.

Four populations share one world. Every curve is computed live in your browser from the same code as the measured runs.

12
5%
4

How deep adults find food (layer², max 49)

Inside one adult's head

holds the true law fact found itself fact heard wrong

True laws held (of 7)

Trust: chance a child accepts what it hears

Measured before this page was built

Headless runs of the same code, 12 sentences per child, mean food depth over generations 51–60. Predictions were committed before each run.

RunPopulationGen 1Gen 51–60True laws
v1, trust a gene from 0.2Alone14.614.22.1
Facts only13.614.32.1
Laws + facts13.815.02.3
v1.1, children trust, fresh worldsAlone13.213.42.2
Facts only12.813.82.0
Laws + facts13.121.04.9
Laws + exceptions13.418.94.5
Laws + exceptions, aware listener13.420.34.8
Facts only, 40 sentences12.816.33.3

The world they live in

  • 8 layers of 8 places, 4 moves from each: 256 rules. In each layer one move leads down from 75% of places. The rest are exceptions with a different down move.
  • Food is worth layer². 10% of steps slip back to the surface. Childhood is 120 steps in trips of 12.
  • A learner adopts a law after seeing the same move lead down from k places in a layer. k is a gene. Elders speak their laws first, one sentence each, then facts.

Ledger

  • Frozen v1 failed all four predictions. Trust started at 0.2 and a 12-sentence culture was too thin to select it up, so the culture never started. Gene and culture each waited for the other. Press "Make trust a gene" to watch it stall.
  • Post-hoc v1.1, fresh worlds: with trusting children, laws beat facts by +7.2 (6 of 6 worlds) and 12 law sentences beat 40 fact sentences.
  • Exceptions did not pay. Sending "where my law is wrong" did worse than random facts. A listener told that the facts were chosen as exceptions recovered part of the loss (+1.3, 5 of 6), short of the +1.5 predicted.
  • Not new: cumulative culture (Tomasello 1999), iterated learning and the compression of language (Kirby, Cornish & Smith 2008), gene–culture coevolution (Boyd & Richerson 1985), the Baldwin effect (Hinton & Nowlan 1987), pedagogy and inferring why a teacher chose an example (Shafto, Goodman & Griffiths 2014).