Artificial-Life Doctrine
From the AI Research Compass wiki
Artificial-Life Doctrine: Intelligence is a late product of life-like processes, so we should grow it from self-organising, self-replicating systems rather than train it.
Beliefs
The Artificial-Life Doctrine thinks the field started at the wrong end. Brains are not where intelligence begins; they are what billions of years of self-maintenance, reproduction and adaptation eventually produced. So instead of fitting a giant function to a static dataset, it studies cellular automata that grow and repair shapes, digital organisms that compete for CPU time, and soft simulated creatures that must keep themselves alive. Get the life-like dynamics right, it argues, and cognition will be a byproduct you did not have to design.
It sits far toward the World side, just above the middle: intelligence is rooted in bodies, metabolism and environments, not in symbols. Its height, just above the middle, reflects a split: it prefers many interacting local units over one central model, with no master network, only rules that every cell follows, yet it wraps no tools or pipelines around them either. Its beyond-revolutionary shade, the darkest on the chart, is earned honestly, since it rejects not just the architecture of current AI but the whole train-on-a-dataset framing.
Arguments
- Neural cellular automata trained with purely local update rules learn to grow a target pattern and regenerate it after damage, showing robust global behaviour from local computation.
- Self-repair, open-ended adaptation and robustness to novel conditions are exactly where today’s large models are weakest, and they are life’s specialities.
- Life is the only process known to have produced general intelligence, so studying its dynamics directly is the most evidence-based route available.
Critiques
- Decades of artificial-life research have produced fascinating toys but no system with competitive performance on any practical cognitive task.
- Evolution took billions of years on a planet-sized computer; there is no clear argument that simulated life can compress that timescale enough to matter.
- “Life-like” is a slippery success criterion, which makes results hard to compare and easy to over-interpret.
Representative works
Relationships
Friends
How to get it
Land on this cell in the full or pro test: x ≈ +0.95, y ≈ +0.27, paradigm ≈ +1.00.