Darwin's World
When an AI model's yes-or-no judgments are the only law of nature, evolution finds the adaptations its common sense expects. A sustained predator-prey arms race appears only when predators alone decide which prey breed, a rule chosen after the first rule produced none.
- 1What does evolution produce when an AI's common sense is the physics?What common sense expects: the tundra island grows thick coats and pale colour in 5 of 5 runs.
- 2Does anything emerge that the AI wasn't directly asked?Not when habitat fit also decides who breeds. When only predators decide, an arms race that per-animal scores cannot produce.
What does evolution produce when an AI's common sense is the physics?
What that common sense expects: on the tundra island, insulating coats spread from 21% of the population to 86–95% and pale colour from 18% to 71–89%, in 5 of 5 runs; the other six islands end at 16% and 8%.
Seven islands start with the same 300 random creatures, each written as 8 plain-English traits such as "a coat of hollow hairs that trap warm air". Every generation, Jev (TypeSafe's judgment model, which answers a yes-or-no question with a probability) judges whether each creature can find food, live in its habitat and survive the island's hard season, and whether one creature could catch and eat another. Survival, breeding, mutation and migration are ordinary code that never reads what a trait means.
How to read this picture
- Each mark is one creature, in its own colour. Shape is its coat: star = fur, triangle = feathers, oval = blubber, diamond = scales, hexagon = armour, circle = bare.
- A red flash is a creature being eaten. Red ripples mark a species going extinct.
- Tree of life: each branch is a species with 25+ members, coloured by its home island; outward is time, thickness is population.
- Hover a creature to read its 8 traits; click to pin it and highlight its family. Space bar plays.
Details: how it works, controls and limits
One generation, on every island
- Encounters. Each meat-eater meets 2 random neighbours. Jev answers "On this island, could the first creature catch and eat the second?" The kill happens with probability 0.7 × p².
- Survival. Jev answers three questions with the island's description: food, daily habitat, and the hard season (a nine-month polar winter, a two-year drought, an eruption). Fitness is the geometric mean of the three answers; survival probability is 0.12 + 0.88 × fitness1.5. Creatures die of old age after 5 generations.
- Breeding. Survivors refill the island to 300, parents chosen in proportion to fitness squared. A child takes each trait from one parent; each trait mutates with probability 0.06.
- Migration. 0.4% of creatures drift to a neighbouring island each generation (1.2% for swimmers and fliers).
Controls
Five independent runs. Different random founders and seeds, 80 generations each, about 450,000 creatures in total.
The other islands. Every island starts from the same founders, so each feature's share on the six other islands is the comparison for the island where it spreads.
What did not converge. Deserts did not evolve nocturnal burrowers: burrowing passed 50% in no run. Large bodies win everywhere (63% across islands), because Jev rarely judges that a small predator can catch a rhino-sized creature.
Limits
- Each island's fitness is a fixed function of one genome and one island, so evolution here is a slow climb to whatever Jev rates highest. Nothing appears that a single question did not already contain. Part 2 tests that critique.
- Jev has read about real animals. This tests whether its common sense is coherent enough to drive selection, not biology.
- One hand-written pool of 394 trait phrases and seven island descriptions.
- All 1.4 million island answers are committed (10 MB), so the five runs replay exactly with no Jev calls; a new run with other settings calls Jev (about USD 1.50 per run).
Does anything emerge that the AI wasn't directly asked?
Only under a rule chosen after the first one failed. When predators alone decide which prey breed, the most common prey design changes 6 to 13 times per run, against 0 to 5 in the controls. When Jev's habitat-fit score also weights breeding, there is no arms race.
Prey and predators each have 64 possible designs, built from two plain-English phrases: how prey escape and what happens when they are caught, how predators hunt and what weapon or sense they use. Jev answered all 4,096 questions "does this predator usually catch and eat this prey?" once per world; the simulation replays those answers for 1,000 prey and 200 predators over 600 generations. Because a prey design's chance of survival depends on which predators exist right now, and the reverse, there is no fixed best design unless Jev's answers collapse into one.
How to read this picture
- Dot area is the design's share of its population. Hover any cell to read the design in full.
- Cell shade: for prey, how often today's predators kill that design; for predators, how often that design kills today's prey.
- The bands below show which strategy dominates over time. A change counts when a new most common prey design holds for 10+ generations.
- Pick a world, a selection rule and a run above. Space bar plays.
Details: how it works, controls and limits
One generation
- Hunt. Each predator makes 15 attacks on random prey (5 when habitat fit also selects). An attack by predator i on prey j kills with probability equal to Jev's answer for that pair.
- Breed. Surviving prey breed back to 1,000; predators breed back to 200 in proportion to kills + 0.25. When habitat fit also selects, both are weighted by the square of Jev's habitat-fit answer ("could this animal find food and raise young here?").
- Mutate. Each of a child's two phrases changes with probability 0.01.
Three worlds change Jev's answers: an Earth river valley, a tidally locked planet (life in a ring of permanent dusk) and a planet with 3× Earth gravity. Jev sees one predator, one prey and the world description; it never sees a population.
Controls, same code and seeds
No-interaction table. Jev's 4,096 answers replaced by their best fit as "predator danger + prey vulnerability". It keeps everything per-animal scores can say and removes every predator-prey interaction.
Fixed predators. A fresh even mix of predators every generation, so prey face a fixed environment.
Shuffled answers. Jev's values randomly reassigned to pairs. It cycles more (13–24 changes) because noise has more interaction than Jev's structured table; the cycles under Jev follow sensible pairs, such as "flier, tiny" against "diver, tiny" while aerial divers hunt.
Limits
- The predators-only rule was chosen after the habitat-fit rule produced no arms race. Both rules are in the data and the figure; the constants were not tuned further.
- With habitat fit in the breeding weight, the critique holds: habitat fit outweighs predation, so prey climb to one design and stay (0 changes in 13 of 15 runs, the same final prey as both controls in 15 of 15).
- One set of 8 × 8 phrases, written by me, and three world descriptions. Different phrases would give a different table.
- Five runs per condition per world; populations of 1,000 and 200, so drift shows in the controls (up to 5 changes with fixed predators on Earth).