Inspiration
Every AI system today is optimization toward a target a human defined — minimize a loss, maximize a reward, predict the next token. It converges, then it stops. But life, culture, and science have no terminal objective. They never converge; they perpetually invent new problems and new solutions. That property is open-endedness, and no AI system has it. Aleph is a bet that open-endedness — not scale — is the missing ingredient for both artificial life and general intelligence. So we did the one thing optimization cannot survive: we removed the objective entirely.
What it does
Aleph is a 2D world of digital organisms, each carrying a tiny neural-network "genome." They sense their surroundings, act, eat, reproduce (clone + mutate), and die when their energy runs out. There is no fitness function anywhere in the code — nothing computes a reward or a score. Selection is entirely emergent from survival. You press play and watch what a world invents on its own.
With no objective of any kind:
- Adaptation emerges. Mean organism lifespan rises from ~1 tick (random founders) to 1,000+ ticks, and 150 founders grow into a stable population. Foraging and patch-following strategies appear that we never programmed.
- Novelty never stops. A behavioral-novelty archive (after Lehman & Stanley's Novelty Search) grows to 560+ distinct behaviors and does not saturate — and it keeps climbing after the population reaches equilibrium. Innovation is decoupled from growth. That is the signature of open-endedness.
- Diversity holds around 0.5 (Jensen–Shannon divergence) instead of collapsing to a single strategy.
How we built it
Pure Python + NumPy, a few hundred lines, no GPUs, no LLMs, no cloud. Organisms are rows in structured arrays; the neural controller is a vectorized MLP forward pass; there is no gradient descent anywhere — learning happens only across generations, through mutation and differential survival. The world is fully deterministic from a single seed, energy is conserved through one accounting choke-point, and 14 tests cover determinism, energy conservation, and the information-theoretic metrics. One command runs it: python -m genesis.run --seed 31.
Challenges we ran into
The system is bistable and seed-dependent: populations either ride the carrying-capacity ceiling or collapse, with a razor-thin boundary between the two. Getting a smooth intermediate equilibrium turned out to be a genuine property of the model, not a tuning miss — a lesson in itself about self-organizing systems. We also predicted communication would emerge and it didn't (see below), which forced us to build careful controls rather than trust an eye-catching chart.
Accomplishments that we're proud of
A working, tested, fully-measured open-ended evolutionary system that demonstrates real adaptation and unbounded behavioral novelty with zero human-defined objective. And — just as important — a hypothesis we rigorously falsified and reported honestly instead of hiding. We instrumented emergent communication (mutual information, Miller–Madow bias-corrected, plus a signal-ablation control) and ran a pre-registered 20-seed sweep. It did not hold up: silencing signals was as good or better than keeping them (ablated larger in 7/20 seeds, signals-on in only 2/20). Communication is not load-bearing in Aleph, and we can prove we checked properly.
What we learned
"Emergence" claims are cheap; measurement is hard. Two findings came out of our failed communication experiment and are the most interesting part of the project: communication is selected only when signals carry non-redundant information, and high mutual information is necessary but not sufficient — even signals correlating with context at 0.7 bits conferred no survival benefit under the causal ablation test. Correlation is not communication. The most valuable thing we built wasn't the simulation — it was the control that stopped us from believing a result that wasn't real.
What's next for Aleph
Open-ended genome topology so network complexity itself can grow; multi-species co-evolution; a predation/danger communication channel where heeding a warning is directly life-saving (redundancy becomes impossible by construction — the most promising route to genuine emergent signaling); and scaling the world to hunt for major evolutionary transitions.
Log in or sign up for Devpost to join the conversation.