Long before “generative” became a product category, Karl Sims was letting populations of blocky virtual organisms mutate until some of them figured out how to swim, walk, jump, follow targets, and fight over a green cube.
The centerpiece for anyone arriving from the Ancient Web is Evolved Virtual Creatures, Sims’s 1994 work. The creatures were not animated one limb at a time by an artist. Their bodies and control systems were encoded, varied, tested, selected, copied, combined, and mutated across generations inside a simulation.
Give the population a task and evolution did the rest.
The results still look strange because evolution has no obligation to produce a sensible character rig. Some creatures crawl. Some tumble. Some discover motions that are effective without looking remotely like anything a human animator would have designed.
That is the interesting part.
Sims was working at the intersection of computer graphics, artificial life, and evolutionary computation when those fields still required an explanation before the demonstration. His site preserves not only videos but the technical trail behind them: the SIGGRAPH 1994 paper “Evolving Virtual Creatures,” earlier work on artificial evolution for graphics, procedural models, particle animation, and interactive evolution.
The surrounding career makes the archive even better. Sims studied computer graphics at the MIT Media Lab, worked at Thinking Machines Corporation, founded visual-effects software company GenArts, and created interactive museum installations including Genetic Images and Galápagos. The latter let museum visitors become the selection pressure by choosing which forms reproduced.
A quarter-century of AI marketing has made it easy to forget how old the underlying urge is: build a system that can search a possibility space humans would never explore manually, then see what crawls out.
Sims was doing that with creatures in 1994.
The page matters in 2026 because it leaves the papers, demonstrations, and artwork together. You can move directly from a bizarre walking pile of blocks to the actual research describing how it was generated. That continuity is increasingly rare. Modern summaries tend to clip the video, call it an early AI experiment, and sever it from the engineering that made it interesting.
CacheRat’s 1,967 Ancient Web Domains research list includes sites like this because sometimes an old homepage is not nostalgia at all. It is the lab notebook for an idea that became fashionable decades later.
The creatures are still weird.
The future just caught up with them.