Aaron Lemke
1.2K posts
@aaronlemke
Creative technologist on the AI frontier. Deep in ML and deep learning research, quietly building something new. Co-Founder @thewavexr. Advisor @qoral.dev
What's been holding artificial life back from the open-ended complexity we see in nature? We posit that the issue, in part, lies in our computational substrates, which are either physically grounded but too slow to evolve at scale, or scalable but disconnected from real physics. Introducing Microcosmos: a scalable, physically grounded, end-to-end differentiable ALife engine where lifeforms are elastic filament chains swimming in a viscous fluid. Accepted at ALIFE 2026. alife.institute/en/blog/microc… Why filaments? Chains of connected units, from bacterial flagella to nematode bodies, are among the most ancient and ubiquitous structures in biology. We model each as an elastic rod coupled to the fluid, sensing the local flow as drag and pushing back on it in turn. Because the fluid is physically simulated, this coupling respects real physical constraints such as Purcell's scallop theorem. Just as a chain of amino acids folds into a protein, a filament can fold into a target shape. Because the full simulation is end-to-end differentiable, we can learn these folds directly with gradient descent. Microcosmos expresses a rich space of locomotion, from hand-designed swimmers like tadpoles and jellyfish to gaits discovered via quality-diversity search, including sinusoidal undulation, directional turning, and paddling with flippers. And it scales. Runtime grows linearly with particle count, not quadratically like simulators built on pairwise interactions. That makes evolution at scale feasible. We release Microcosmos as an open platform for the ALife community to build on. Our hope is to offer a substrate that is grounded enough to be credible, scalable enough to support evolutionary search, and rich enough to inspire. Paper: arxiv.org/abs/2607.02954… Code: github.com/alife-institut…
Adding sound to one of my favorite RL creatures. Built with @threejs and @webgl_webgpu trained @runpod using Jax and MuJoCo from @GoogleDeepMind. A single oscillator maps to the output activation while hidden layer activation controls a formant filter. Sonification can give different insight into the behavior of these neural nets than just visualization.
