A research project, explained
Almost every artificial agent is handed a goal. Life never was. This project asks which features of the early Earth are needed for agents to appear without a goal being installed, and tests that in simulated worlds, with the predictions written down first.
Read this first. This is a hypothesis tested in toy worlds. Most of the results so far are negative, and this page says so. Nothing here shows how agency arose on Earth.
The idea
Sunlight is concentrated energy. Left alone it turns into heat and spreads out. That difference, concentrated against spread-out, is called a gradient, and the universe is full of them.
Anything that happens to drain a gradient well, a flame, a hurricane, a plant, tends to persist while the gradient lasts. Plants drain sunlight. But plants never use all of it, so something else can live off what they leave: a grazer, then a hunter, then the things that break down what is left.
The hypothesis: at every stage, whatever uses the leftovers best persists, and that chain eventually produced living things that sense and choose. We call those agents. The project does not assume this is true. It asks which ingredients of the chain are doing the work.
Five conditions
The simulations keep five features of the early Earth that might be doing the work. Each can be turned off, one at a time and in every combination. If agents still appear without a feature, that feature was not needed.
The design rule
Entropy here means how much energy a system turns into spread-out heat. A busy ecosystem produces more of it than bare rock under the same sun.
The rule: creatures are selected only by whether they stay alive and make copies of themselves. Entropy production is measured by the experimenter and never handed to the creatures as a reward.
This does not mean the two are unrelated. Energy helps you survive, so a creature that takes in more energy also produces more heat. What the rule forbids is a separate prize for entropy.
Life does not. Living things carry learned internal signals tied to energy and balance, such as hunger and satisfaction. None of them refers to entropy.
Rewarding entropy directly also puts the answer in the question: if you pay for it, finding it proves little. And as we tested, it can reward the wrong behaviour, burning energy for its own sake. See the experiments on rewarding entropy below.
If the aim is to build agents rather than explain them, rewarding it is a fair design choice. We tested that too.
The simulated world
Creatures live on a wrap-around grid. They can move, stay still, pay energy to open a well, eat, copy themselves and die. Everything that goes in is accounted for: the energy that enters the world equals the energy creatures take up plus the energy lost as heat, and the code checks this on every step.
Spread thinly everywhere, regrows slowly. Easy to eat, never much. This is the bottom rung.
Rich, but locked. One creature can pay energy to open one. Others on the same square share the payoff, which makes opening one a gamble.
Very rich, and they only open if several creatures pay at the same time. They appear where small wells have been drained.
The experiments
Every experiment had its predictions written down before the confirmatory run, in a file that was committed to the repository first. Each prediction has a verdict. Open any card for what we did, the numbers, and what the result does not show. Bars show the estimate, with a line for the 95% interval.
Jeremy England's idea
The physicist Jeremy England argued that when a system is pushed by an energy source and can rearrange itself, rearrangements that absorb and dissipate more of that energy are more likely to stick. No selection or reproduction is needed for this. He called it dissipative adaptation.
Where the theory stands. A related 2013 bound on how much heat a self-copier must give off has been challenged: a 2024 analysis argues it is not physically meaningful in general. Dissipative adaptation itself has support from simulations and a few experiments, and a recent review describes the link to agent-like behaviour as speculative.
How we tested it. No births and no deaths. Creatures re-draw their instincts only through a rule that keeps a change more often when it absorbed more energy. The dial shows what that kind of bias does in the simplest possible setting.
Result: the biased creatures took in more energy than unbiased ones, which is what the rule builds in, but they did not climb to the big wells. So the rule alone produced nothing that looks like agency.
What we learned so far
Selection by survival helps a little. Entropy production rose by 2.3 units over random parents, a small but clear effect.
Our central prediction was wrong. We expected scarcity, inheritance and death to be required to reach the big wells. Removing any of them made reaching them more common.
Paying for entropy yields a furnace. Optimising it directly beat our best hand-written forager by burning energy, and giving creatures a learned reward produced no sign of agency by our tests.
Nothing about Earth. These are toy worlds. Whether they capture what mattered is the main assumption.
Nothing against bigger genomes in general. Our large-genome world killed most runs before learning could start, so the test was uninformative about what a large genome can do.
Nothing final about agency. Our two tests of agency are crude. One of them was distorted by how it handled the random comparison group, and the other had no positive control.
Check our work
One preregistration file per experiment, committed before the confirmatory run, with an outcome section added afterwards and a log of any deviation.
Every run's numbers are in the repository as CSV, with the report generated from them.
The world, the experiments and 53 tests that check the energy accounting. Long runs are reproducible from the repository's Actions tab.
Plain answers