The story so far
Eight chapters.
One extinction. One first word. One teacher who had to go. And a creature called Omega that crossed from one world into another.
An artificial-life experiment · Genesis 1.8
We are trying to watch a mind begin.
Digital Genesis is a world small enough to run on one laptop, filled with creatures that are handed nothing. No language. No lessons. No instinct to stay alive. Whatever they become, they have to earn, and whatever we say about them has to survive a control.
Scroll to enter · you will travel through itWhy it matters
Every AI you have used was told everything.
The systems in the news learned by reading. An ocean of text, written by people who already knew how to speak, argue, teach and remember. They begin at the end of the story.
Digital Genesis begins at the beginning. Its creatures wake up in a world with food, hunger and nothing else. If one of them ever signals to another, it is because signalling kept it alive. If knowledge is ever passed on, it is because it outlasted the one who had it.
It asks the oldest question there is, in a place where the answer can be checked: where does a mind come from?
Three reasons this is a big thing
Small on purpose. That is the point.
Told → Earned
Nothing is handed over
The organisms never read a word of human writing. A behaviour exists only because, somewhere, it helped something stay alive. When a signal appears, nobody put it there.
Billions → 896
A mind you can read
A modern language model has billions of numbers inside it that nobody can interpret. Omega, the main character here, has 896. You could print them on two sheets of paper.
Demo → Experiment
You cannot rerun history. Here, you can.
Nobody can go back and remove the first teacher to see whether the lesson survives. In this world that is one setting. The project changed it, and measured what happened.
It is easy to build something that looks alive. The hard part is refusing to be fooled by it.
The world · v0.1 to v0.2
A place that does not care whether you live.
Before there could be a story there had to be somewhere for it to happen. So the first thing built was not a creature. It was hunger.
EnergyEverything alive is spending it. When it reaches zero, that organism is gone.
FoodScarce, and scattered. Finding it is the only way to refill.
ChildhoodJuveniles have to grow before they can do what adults do.
OffspringPaid for out of the parent's own reserves. Reproducing is a risk.
Why it is built this wayEvery run can be repeated on a "paired seed": the identical starting world with one thing changed. That is how an effect is told apart from luck.
Nothing in it is scripted to survive.
Chapter 1 · v0.3
Forty-five populations were given brains. All forty-five died.
Each organism received its own small neural controller and a memory of who it had met. It was supposed to be the moment things got interesting.
Not one population produced a single birth. They wandered, they starved, and the counter went to zero forty-five times.
It would have been easy to fix it quietly and never mention it. It went into the report instead. It is the first line of a ledger this project still keeps.
So why did every one of them die?
Chapter 2 · v0.3.1
The first ones to live
The cause was almost embarrassingly small. Decisions were being routed to the wrong part of the brain, and the credit for a good action was landing in the wrong place. The creatures were doing the right thing and learning nothing from it.
With that fixed, the next screen recorded 90 neural births. All 16 groups trained from demonstrations reproduced.
Ninety is not a large number. After forty-five zeros, it felt like one.
They could live. Could they talk?
Chapter 3 · v0.3.2 to v0.3.3
The first word
Each organism was given its own way to send a signal and its own way to listen, with limited range and delayed feedback. Nobody defined what any signal meant.
The early attempts were brutal. Under the original schedule of environmental eras, every group died. Then, in a sheltered version of the world, five groups learned to signal and lasted a million simulated years.
When an abrupt flood arrived, 11 of 12 signalling pairs were lost. With a gentler transition on fresh seeds, 12 of 12 survived. Their twins, whose signals had been secretly shuffled, managed 2 of 12.
Why the shuffle mattersIf scrambled signals had worked just as well, the "language" would have been noise. They did not. The signals were carrying something the partner needed.
Nobody defined the words. They meant something anyway.
Chapter 4 · Genesis 1.0.1
The last teacher dies
If a signal is useful, can it be handed down? Newcomers were fostered into populations of experienced "teachers" and left to grow up among them.
Then the teachers were removed. Every one. The learners were sent to maps none of them had ever seen.
They scored 73.3%. Learners with only what they had inherited scored 52.2%. With messages blocked, 50.6%, which is barely better than guessing.
Scroll on, and watch what stays behind.
The teacher is gone. The lesson is not.Whether it becomes a lasting culture is still listed as an open question
Chapter 5 · Genesis 1.4
Omega crosses over
One organism was carried forward from experiment to experiment. By the time it was archived it had made 944 recorded decisions. Its whole motor controller is 896 numbers. It was named Omega.
Then it was taken out of the only world it had known and placed in a body, in a fictional settlement with places to go and things to need.
Before believing anything it did there, the project replayed its past, decision by decision. All 432 archived decisions reproduced exactly. It was the same Omega.
Then came the tests it did not pass.
Chapter 6 · Genesis 1.4.5 to 1.7.1
The hard part
This is the chapter most projects leave out.
In a longer survival test, 0 of 80 runs were alive at tick 500. A language model was brought in as an adviser; it made legal moves and produced no measurable benefit. In a puzzle house, Omega's setup escaped 3 times out of 3, and so did a plain rules-only control, so the credit was not Omega's.
One thing held. On 20 worlds it had never seen, Omega found food and survived 11 times. Random action did it once.
Eleven out of twenty, earned honestly, is worth more than twenty out of twenty you cannot trust.
A test you cannot fail is not a test.
Chapter 7 · Genesis 1.8
Homer remembers being wrong
Omega's descendant is called Homer. It has the same small motor policy, and three new things: a journal of what has happened to it, goals it selects from a permitted list, and a school.
On a slope problem, Homer got the answer wrong. The mistake was verified and written into its journal: the method it used and the fact that it failed. Never the right answer.
On the next problem, the Homer with a journal recalled the mistake, changed its method and solved it. The Homer without one made the same mistake again.
It is one matched pair, and the project says so. It is also the first time in this whole story that something remembered being wrong and did something about it.
Chapter 8 has not been run yet.Its question is at the end of this page
The cast · Omega
This is all of it. Every point is one number.
What you are looking at is Omega's entire motor controller, drawn to scale in the only sense that matters: 896 points for 896 coefficients, grouped into 32 small recurrent modules.
For each decision, three modules are routed in. They are the ones lit in amber. The rest wait.
There is nothing else. No hidden layer you are not being shown, no second system doing the real work. When this project says it knows what Omega did, this is why it can.
What it can doMove, eat, rest, share, investigate.
How it is protectedIts checkpoint is hash-verified before every run.
The rest of the cast
Everyone who got us here.
A story about one survivor is also a story about everything that did not survive, and everything that helped.
v0.3 · 45 populations
The Forty-Five
Given brains before anyone knew how to wire them. None left a descendant. They are the reason every later result is checked twice.
v0.3.2 · 5 groups
The Million-Year Five
Five groups that learned to signal in a sheltered world and lasted a million simulated years. The first proof that it could be done at all.
v0.3.3 · 12 pairs
The Pairs
Twelve pairs with a private channel and no dictionary. All twelve lived. Ten of their twelve scrambled twins did not.
Genesis 1.0.1
The Last Teachers
Removed on purpose, to find out whether what they knew would stay behind once they were gone. It did.
Genesis 1.3 onward
Omega
896 numbers and 944 recorded decisions. Moved between worlds, replayed line by line, and tested against rivals that sometimes beat it.
Genesis 1.8
Homer
Omega's descendant, with a journal and five schools to attend: The Archive, The Academy, The Forge, The Citadel and The Open Horizon.
The method
How to lie with a simulation, and how this one refuses to.
Every result on this page has a twin: the same experiment with the interesting part removed. Try it yourself.
Run the control · v0.3.3
12 of 12 pairs survive.
Each circle is one pair facing a changing world with its signals intact.
Four ways to fool yourself
After the teachers were removed · score on unseen maps · Genesis 1.0.1
The receipts
Every result in the story. Wins and failures, same page.
A project that only reports success is showing you its marketing. This is what the experiments returned.
7held up against a control
7null results and failures
v0.3First neural populations went extinct45 of 45
v0.3.1Neural births in the held-out screen90
v0.3.2Groups surviving the original era schedulenone
v0.3.2Groups that learned to signal and lasted a million simulated years, in a buffered world5
v0.3.3Signalling pairs lost to an abrupt flood11 of 12
v0.3.3Signalling pairs survived a gentler transition, against shuffled signals12/12 vs 2/12
1.0.1Unseen-map score after teacher removal, against blocked messages73.3% vs 50.6%
1.4.1Omega's archived decisions reproduced on replay432 of 432
1.4.1Preserved memory against standardized memory+3.6 pts, interval spans zero
1.4.5Runs alive at tick 500 in held-out environments0 of 80
1.7.1Survive-and-eat task, Omega against random action11/20 vs 1/20
1.7.1Benefit from a language-model adviser recalling memoriesnone measured
1.8Recalled a verified error and switched method1/2 vs 0/2
1.8Evolved method selector against the worked baselinetied, not promoted
Held up against its controlNull result or failure
Honest answers
The questions you are right to ask.
Is it alive?
The projectNo. It is arithmetic that behaves, in a world that is also arithmetic. What is real is the pattern: behaviours that appear only when they are useful, and vanish when the reason for them is taken away.
Is Omega conscious? Is Homer?
The projectNo, and nothing here claims it. Homer's own design notes say its brain-like names are functional metaphors and that it makes no claim of simulated sentience. It is not a mind upload. It is a small controller with a journal.
Then is this the road to human-level AI?
The projectNobody knows, and it would be dishonest to say so. What this is, for certain, is a way to study the first steps of learning, signalling and memory with nothing hidden and nothing borrowed.
What has it failed at?
The projectStaying alive for 500 ticks. Benefiting from a language model's advice. Beating a simple rules-only control in the puzzle house. Showing, so far, that it learns from penalties. All of it is in the receipts above.
So why should I care?
The projectBecause the questions are enormous even when the creatures are tiny. How does meaning start? How does knowledge outlive the one who had it? What does it take to remember a mistake? Here, those can be asked with a control group.
Could I check it myself?
The projectThat is the design. The headless experiments use Python 3.12 and its standard library, on a laptop CPU. No GPU, no cluster, no special access.
Chapter 8 · not yet run
Does it learn from consequences?
Everything so far leaves one question standing. When Omega avoids something, is it because of what happened last time, or only because it is hungry right now?
The next experiment is a preregistered two-choice test. Same terrain. Equal immediate costs. Independent seeds. One measurement: whether the choice changes after a penalty.
Nobody knows how it comes out. That is what makes it an experiment.
Somewhere on a laptop, something very small is being asked a very hard question.
The answer gets published either way.
The film