The recent emergence of large AI models has reshaped many aspects of human society. Most these AI systems are built top-down, where neural architecture is artificially designed. These models excel at specific tasks but remain disembodied, fragile outside training distributions, and costly in energy. In contrast, the human brain emerged bottom-up through a single rule of natural selection, developing the ability to learn, adapt, and perform a wide range of tasks with astonishing energy efficiency. This contrast motivates Siliang Song’s research.
Inspired by Karl Sims’s 1994 work “Evolving Virtual Creatures”, Siliang Song aims to extend this concept into an open-ended, multi-species ecosystems where digital organisms possess richer bodies, neural systems, and behaviors. He will pursue two main goals. (1) Create a new AI paradigm in which neural architectures and functions evolve naturally alongside embodied digital organisms that live within virtual worlds governed by realistic physics and resource limitations. (2) Use this platform to gain insights into evolutionary principles that are difficult to examine through traditional methods, such as the emergence of self-assembly, modularity, and different reproduction strategies.
Specifically, Siliang will (1) build a digital evolution platform that integrates GPU-accelerated physics simulations with evolutionary algorithms, creating sophisticated virtual environments and selection strategies that foster the evolution of complex behaviors and neural controls; (2) design digital organisms with developmental “genomes” and rules that simultaneously specify morphology and neural control, allowing form and function to co-evolve; (3) introduce multi-species interactions with resource constraints to promote competition, cooperation, and coevolution, thereby driving continuous evolutionary innovations; and (4) gradually increase environmental complexity and expand the phenotypic design space to enable the evolution of increasingly sophisticated intelligence.
The project will explore a potential path toward an adaptive, general, and energy-efficient paradigm of AI, and will also provide a platform for testing fundamental principles of evolutionary biology.