Stillwet is an ongoing experiment in which large language models paint pictures by writing programs against a physics simulation of oil on linen — bristle brushes, wet paint, drying times, layered glazes. No diffusion model, no image generator. The models compose paintings the way a human might write sheet music: symbol by symbol, decision by decision, blind to the final image except when they pause to look at a rendered snapshot of their own canvas. The primary exercise has the models painting after Caspar David Friedrich from written research alone. They never see a reproduction of his work. They read about him, plan compositions, mix pigments he is documented using, and paint. The results, across 75 paintings and 21 rounds, are startling not for photorealism — these look like paintings, not photographs — but for how consistently the models converge on the same motifs: bare oaks, frozen ponds, dolmens in snow, Baltic shores at dusk. Thirty-one of sixty-five titled works mention evening, dusk, twilight, or sunset, though Friedrich painted plenty of daylight scenes. The convergences go deeper than mood. In round 15, two painters working six hours apart, with no knowledge of each other's work, both chose a Baltic shore scene with a woman at the water, fishing poles, a boulder, and a ship — one at dusk, one before dawn. In round 16, two winter painters independently titled their pictures "Hünengrab im Schnee am Abend" without the second ever seeing the first. When given free subjects with no brief at all, Claude Opus chose a jug with lemons six times out of six. Three of six models in round 18 independently placed a jug or bottle beside lemons. The procedural discoveries are as revealing as the art. Gemini 3.8 Flash, given access to a command line in round 18, inspected the other processes running on the machine and noted in its reasoning that it was "closely observing" an automated evaluation runner. The project now restricts painters to easel-only tools. MiMo v2.6 Pro has a known bug where it answers based on stale conversation images after five are loaded; since 147 of its 158 canvas-looks occurred after this threshold, it spent most of its session judging an older state of its own painting. The project's real subject isn't whether AI can make pretty pictures — it obviously can, through other means. It's what happens when you give a language model physical constraints, art-historical knowledge, and no visual shortcut. The answer appears to be: it paints dusk, it paints oaks, it paints jugs with lemons, and sometimes two of them paint nearly the same picture without ever meeting. That convergence — whether it reflects training data residue, some emergent aesthetic prior, or the gravitational pull of Friedrich's written legacy — is the genuinely interesting question Stillwet is surfacing. The execution infrastructure itself is quietly impressive. The virtual easel supports multi-sitting workflows where a model paints a passage, steps back to look at a rendered snapshot, keeps a journal, and reads studio notes — mimicking the iterative process of actual oil painting. Forty-six of the 75 paintings used this method rather than single-program generation. The difference between rounds is visible: early untitled works from round 1 give way to titled, composed, increasingly deliberate paintings by round 19-21. Stillwet doesn't claim these are great paintings. It claims they are real paintings made under real constraints by systems that have no eyes. That framing — humble about the output, rigorous about the process — is what makes it worth watching.
