Link
Automating the Search for Artificial Life with Foundation Models
My thoughts
I was introduced to Artificial Life by learning about “The Game of Life”, also known as Conway’s Game of Life. This is a cellular automaton that solely depends on the initial conditions. You setup the a world with some initial conditions and rules for evolving and then let time start. The “creatures” you create then evolve based on some preliminary rules. It’s fascinating to watch.
Reading up on Conway’s Game of Life, I stumbled across “Alife” which is Artificial life. This is a field of study that looks at natural life, the processes of natural life, and its evolution.
The field of Artificial Life (ALife) spans three domains:
- “Soft” ALife: Software implementations
- ”Hard” ALife: Hardware/robotic systems
- ”Wet” ALife: Biochemical approaches
Similar to how AI / AGI look at intelligence and how to replicate it, Alife or A-Life look at life and how to replicate it.
The article linked is from Sakana AI, a Tokyo-based AI company. Sakana AI was founded by some of the authors of the original Transform paper, “Attention Is All You Need”.
They propose a new algorithm called “Automated Search for Artificial Life (“ASAL”)”, which automates the discovery of artificial life with a vision-language foundation model.
ASAL primarily operates in the “soft” domain but its principles could extend across all three.
They propose (quote from the Academic Paper):
a new paradigm for ALife research called Automated Search for Artificial Life (ASAL). The researcher starts by defining a set of simulations of interest, referred to as the substrate. Then, as shown in the figure above, ASAL enables three distinct methods for Foundation Models to identify interesting ALife simulations:
- Supervised Target Searching for a simulation that produces a specified target event or sequence of events, facilitating the discovery of arbitrary worlds or those similar to our own.
- Open-Endedness Searching for a simulation that produces temporally open-ended novelty in the Foundation Model representation space, thereby discovering worlds that are persistently interesting to a human observer.
- Illumination Searching for a set of interestingly diverse simulations, enabling the illumination of alien worlds.
Or more simply:
- Supervised Target Search: Finding simulations that produce specific desired behaviors
- Open-Endedness Detection: Discovering systems that continuously generate novelty
- Illumination: Mapping the space of possible simulations to understand the full range of potential artificial life forms
This means that, (quoting again from their Academic Paper):
Instead of probing for rules and interactions that feel right, researchers could refocus their attention to the higher-level question of how to best describe the phenomena we ultimately want to emerge as an outcome, and let the automated process of searching for those outcomes then take its course.
This represents a significant shift from manually designed systems to automated discovery, which were very brittle to build and could take a long time to build as well.
This then ultimately allows for the possibility of discovering new forms of artificial life that could help us understand the principles underlying the real “life and intelligence” we see today.
The classic example of Conway’s Game of Life gives us a way to look at what’s new about this research:
- Rule-Based Emergence: Like Conway’s system, but with automated discovery
- Pattern Recognition: Now powered by foundation models
- Behavioral Complexity: Systematically measured and explored
Ultimately, this approach could help AI and A-Life system and AI Agent design by:
- Enabling automated discovery of emergent behaviors
- Providing a framework for testing complex system interactions
- Creating more robust testing environments for AI agents
Finally, for developers interested in experimenting with ALife: