That's Esoteric

Automating the Search for Artificial Life with Foundation Models (ASAL)

2 min read

Automating the Search for Artificial Life with Foundation Models (ASAL)

Authors: Akarsh Kumar, Chris Lu, Louis Kirsch, Yujin Tang, Kenneth O. Stanley, Phillip Isola, David Ha

This research introduces a novel framework called Automated Search for Artificial Life (ASAL), which integrates vision-language foundation models (FMs), such as CLIP, to automate the exploration and discovery of digital lifeforms. It addresses a core struggle in ALife: the historic reliance on manual design and trial-and-error to configure complex lifelike simulations.

Three Search Mechanisms

By leveraging the human-aligned representations inside Foundation Models, the authors propose three search methodologies applied to ALife substrates (the rules/initial states of a world):

  1. Supervised Target Search: Hunting for a simulation that yields a trajectory matching a specific text prompt (e.g., “a branching organism” or a temporal sequence of prompts).
  2. Open-Endedness Search: Searching for simulations that generate historical, non-repeating novelty over time, avoiding situations that statically plateau or dissolve into visual noise. Since FMs map to human intuition, novelty in the FM space maps directly to “interestingness” to a human observer.
  3. Illumination Search: Mapping out the entire parameter space of a substrate to find an array of maximally diverse simulations (an atlas of alien worlds).

Key Discoveries across Substrates

The ASAL framework proved wildly effective, discovering new emergent phenomena without manual tuning across multiple classic substrates:

  • Boids: Discovered new flocking patterns like snaking, cyclic circling, and hierarchical grouping.
  • Particle Life / Lenia: Found previously unseen self-organizing digital bacteria, “caterpillars”, and dividing cells by optimizing interactions natively through FM scores.
  • Life-like Cellular Automata: Systematically ranked CA rules to find cellular automata that exhibit endless open-ended evolution similar to Conway’s Game of Life.

The paper concludes that using Foundation Models allows researchers to quantify and optimize for complex, traditionally subjective traits like “open-endedness” and “diversity” automatically.

  • AI_Safety — ASAL demonstrates emergent, unscripted behaviors in AI agents; understanding emergence is critical to alignment
  • Mandelbrot_Set_Quasi_Black_Hole — both papers explore self-similar, emergent complexity: one in digital substrates, the other in mathematical/cosmological geometry
  • Emergence — meta-concept: ASAL as a case study in computational emergence

Graph View

Backlinks