Examples of discovered artificial lifeforms.
We show examples of the simulations discovered by our algorithm, Automated Search for Artificial Life (Paper, Website). In Lenia (top-left), ASAL discovers a diverse set of dynamic self-organizing patterns reminiscent of real cells. In Boids (top-right), ASAL discovers exotic emergent flocking behavior. In Particle Life and Particle Life++ (bottom-left), ASAL discovers dynamic open-ended ecosystems of agentic patterns. In Game of Life (bottom-right), ASAL identifies novel cellular automata rules that are more open-ended and expressive than the original Conway’s Game of Life.
Introduction
For the past 300,000 years, Earth has had only one form of advanced intelligence on it: humans. With the recent advent of AI foundation models, some believe we are at the dawn of a new kind of intelligence. As AI continues to evolve, we may witness the proliferation of diverse intelligent lifeforms coexisting with us.
But how did we get here in the first place? What fundamental principles govern the emergence of all life and intelligence, whether biological or artificial? What might the open-ended evolution of the ecosystem of our AI agents look like in the future? Though we don’t yet have the definitive answers to these questions, we can gain insight by returning to the scientific field that laid the groundwork for exploring these questions: Artificial Life (ALife). ALife offers the tools and framework to study the dynamics of artificial lifeforms, fostering insights into their potential behaviors, interactions, and trajectories.
So what is ALife? At its core, ALife is the ambitious quest to recreate and understand the phenomena of life itself—how it emerges, evolves, and thrives. It’s not just about mimicking Earth’s biology but going beyond and creating completely alien worlds to understand the principles that underlie all possible life. ALife researchers craft virtual ecosystems, robotic organisms, self-replicating programs, and biochemical simulations to uncover the deep mechanisms of complexity, evolution, and intelligence.
Sakana AI has previously drawn ideas from ALife to develop better foundation models, resulting in our works on evolutionary model merging, LLM self-play, and autonomous open-ended discovery. But now we want to go the other way: can foundation models help the study of ALife? Bridging this two-way road will be essential to getting more capable, natural systems and for understanding them as well. Regardless of whether or not you think foundation models will lead to the next generation of artificial lifeforms, they have already started revolutionizing various scientific fields. In fact, the recent Nobel Prize was awarded for radical advances in protein discovery, driven by a foundation model. They are also being used to predict the climate, do AI research itself, and prove mathematical theorems, so why not apply them to help in the search for artificial lifeforms?
In collaboration with MIT, OpenAI, The Swiss AI Lab IDSIA, and Ken Stanley, we are excited to release our new paper, Automating the Search for Artificial Life with Foundation Models (website).
In our paper, we propose a new algorithm called Automated Search for Artificial Life (“ASAL”) to automate the discovery of artificial life using vision-language foundation models! Our proposed approach aims to (1) find simulations that produce a specific target behavior, (2) discover simulations that keep generating novelty forever as you run it, and (3) illuminate all the different simulations that are possible.
Finding lifeforms that produce a specific target behavior: Here, we present examples of Artificial Life simulations, discovered by ASAL, which were found only with specified target prompts. Read on to see more results below in this blog post!
Because of the generality of foundation models, ASAL can discover new lifeforms across a diverse range of seminal ALife simulations, including Boids, Particle Life, Game of Life, Lenia, and Neural Cellular Automata. ASAL even discovered novel cellular automata rules that are more open-ended and expressive than the original Conway’s Game of Life. Additionally, the use of foundation models allows us to quantifiably measure previously qualitative phenomena, in a human-aligned way.
We believe this new paradigm may reignite ALife research by overcoming the bottleneck of manually designed simulations, thus advancing beyond the limits of human ingenuity.
What is Artificial Life?
ChatGPT-generated image with the prompt “the open-endedness of evolution on earth producing all the diversity of life”.
Natural evolution has produced the vast diversity of life on Earth, from bacteria to whales and humans. Artificial Life (ALife) seeks to understand this open-ended process of life by recreating it through computer simulations. While inspired by biology, ALife’s ambitions extend far beyond, exploring not only “life as we know it” but also “life as it could be.” Over time, ALife has grown to study the general emergence of complex behaviors from simple components, encompassing phenomena such as self-organization, collective intelligence, and open-ended evolution.
Using ALife, we can study some intriguing questions:
- What exactly is “life”?
- What lifeforms can exist in a hyper-realistic 3D world simulation?
- What does life even look like in a cellular automata?
- How does life emerge in a digital soup of self-replicating computer programs?
- If we slightly change the environment’s rules, is life still possible? When is it inevitable?
- What is required to ignite a “never-ending” algorithm similar to natural evolution?
The last question touches on perhaps the most fascinating property of natural evolution: open-endedness. The pursuit of truly open-ended systems capable of discovering interesting artifacts indefinitely remains so elusive that it is widely regarded as a grand challenge.
John Conway’s Game of Life (CGoL) is a simulation of a 2D grid with rules like “a dead cell with 3 alive neighbors becomes alive in the next timestep”. With only these simple rules, CGoL is able to simulate entire self-replicating “spaceships”. It can even simulate CGoL within itself!
Karl Sim’s Evolved Virtual Creatures is a masterclass in discovering artificial lifeforms in a 3D virtual world. Similar to real animals, these artificial animals have DNA that encodes the morphology and the brain of the animal. Then, the DNA can be evolved with a genetic algorithm to produce phenotypes capable of swimming, walking, and jumping.
Other ALife projects: Many more ALife simulations have been developed. Some model ecosystems of cells, while others simulate the evolutionary arms race of predators and prey as they forage for resources. Self-play, which bootstraps capabilities from scratch, is inspired by these evolutionary dynamics.
Our Method: Automated Search for Artificial Life (ASAL)
The most compelling ALife simulations took months, if not years, to hand-design. This is because the emergent behavior of complex systems is often impossible to predict in advance. Imagine the challenge of designing the periodic table from scratch, manually specifying the pairwise interactions of all elements. Finding the exact configuration that leads to interesting outcomes would be a nearly endless task. Now, imagine instead that you could simply define the number of elements—say, 100—and let an algorithm automatically discover the interaction rules that produce fascinating emergent simulations.
ASAL: Our proposed framework, ASAL, uses vision-language foundation models to discover ALife simulations by formulating the processes as three search problems. Supervised Target: To find target simulations, ASAL searches for a simulation which produces a trajectory in the foundation model space that aligns with a given sequence of prompts. Open-Endedness: To find open-ended simulations, ASAL searches for a simulation which produces a trajectory that has high historical novelty during each timestep. Illumination: To illuminate the set of simulations, ASAL searches for a set of diverse simulations which are far from their nearest neighbor.
This is the new paradigm we propose for ALife research. Once the researcher defines a space of simulations, or “substrate,” to search over, ASAL automates the search for interesting simulations that, when run, produce videos matching desired criteria, as evaluated by a vision-language foundation model. The criteria include:
-
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 itself produces novelty in the foundation model representation space as you run it, thereby discovering worlds that are persistently interesting to a human observer.
-
Illumination: Searching for a set of interestingly diverse simulations, enabling the illumination of all the possible alien worlds.
Here are some simulations ASAL discovered after specifying a single prompt, which is a text description of the image the simulation should create. In Lenia, prompts like “a self-replicating pattern” revealed dynamic structures mimicking biological processes, while Boids captured emergent behaviors such as “collective intelligence” and “a Fibonacci spiral in nature”. Particle Life produced visually compelling patterns like “cell division” and “a diverse ecosystem of cells”, highlighting ASAL’s ability to turn abstract concepts into concrete simulations that evoke both scientific and artistic intrigue.
Multiple prompts can be applied sequentially to find simulations producing a desired sequence of events. In the first simulation, ASAL discovered an update rule which allows a cell to split into two. The second simulation showcases ASAL’s level of control and its potential to eventually discover simulations displaying long and complex evolutionary trajectories. However, our work has not yet achieved this ultimate vision, and more progress is needed before truly fascinating worlds can be discovered solely through specified text descriptions.
For Open-endedness, ASAL discovered several simulations which are more open-ended than the famous Conway’s Game of Life (CGoL). If people have found spaceships and computers in CGoL, imagine what we could discover in these new cellular automata worlds!
ASAL can also illuminate the entire substrate to find a set of interestingly different simulations as shown here. This provides researchers with a general overview of what may be possible in a given substrate and serves as a step toward taxonomizing all potential life forms within the computational universe.
What Now?
ASAL is an exciting achievement, but there’s lots more to be done…by you! We open source our code on our GitHub.
Try your own substrate
We encourage you to apply ASAL to your own custom substrates you find interesting and explore what happens!
Here, we develop a novel substrate “Particle Life++” based on Particle Life, but which allows the colors to change as part of the dynamics rule, allowing for a combinatorial explosion in interaction dynamics. Out of the box, ASAL is able to find open-ended simulations in this substrate:
It would be very interesting to see ASAL applied to other substrates like ALIEN and JaxLife!
Room for creative exploration
After discovering some cool simulations, there is a lot of room for creativity and exploration. For example, here, we take the many “species” of Boids creatures discovered, and allow them to enter each other’s universes after a period of time. We can see some symbiotic relationships emerge, while some relationships are more invasive and destructive:
Similarly, we can also compare different cellular automata by pitting them against each other to see which one dominates the most territory:
Bigger picture
Overall, our ASAL framework will allow for the automated discovery of many new kinds of artificial lifeforms. In turn, this will help us understand the general principles of life and all complex systems, adding to our knowledge of concepts like emergence, computational irreducibility, assembly theory, and open-endedness.
In the bigger picture, we believe ALife is worth researching because there are a lot of important ideas from ALife that can be and should be incorporated into AI. For instance, the next generation of AI algorithms will likely incorporate concepts like open-endedness, self-organization, and collective intelligence in order to be more adaptive, creative, and continually learn. By bridging ALife and AI, we as a field have the unique opportunity to unlock a new era of natural AI systems.
Sakana AI
Interested in joining us? Please see our career opportunities for more information.
