Weird AI Projects — Sept. 7

Weird AI Projects — Sept. 7

This week’s batch is unusually good. The common thread isn’t “AI makes content”; it’s AI becoming part of the operating logic of an object, interface, body, or simulated world. These are the five I’d spend time with.

1. Solaris — an interface that generates itself while you’re using it. Runway announced Solaris on August 31. It calls the system an “Interface World Model”: instead of an app having predefined screens and menus, the interface is generated continuously in response to what the user does. That’s a genuinely different conception of software—not AI generating an app once, but software whose UI may effectively exist only for the current moment. (source)

This is probably the most important thing in this week’s report for you. Imagine UMS having capabilities—play media, inspect schedules, search the archive, control KMAT, check system health—but essentially no fixed interface. Asking “what hasn’t aired lately?” might cause a little programming dashboard to materialize; asking “what’s wrong with channel 8?” might instead create diagnostics controls.

2. moi — your coding agent builds its own computer around itself. The open-source moi project on GitHub gives Claude Code, Codex and other agents a visual workspace they can actually modify. The agent can create persistent widgets, dashboards, canvases and full application views connected to APIs, local files, commands and MCP servers. Unlike a transient generated UI in chat, those interfaces remain in the workspace as software. (source)

The weird conceptual leap is excellent: instead of building software with Codex, give Codex an empty operating environment and let it build whatever software it needs around itself. This is small and early—the repo currently shows only dozens of GitHub stars—but I’d watch it precisely because it feels like an experiment rather than a polished AI product.

3. Parallel Forests — AI designed from a plant’s point of view. Artist Yuma Kishi built Parallel Forests, shown in Tokyo this year. Small edge-computing devices distributed through a nocturnal forest sense plant bio-data and environmental changes. Kishi calls the system Botanical Intelligence (BI) rather than artificial intelligence; algorithms adapt to incoming biological/environmental signals and translate them into sound and light that are directed back toward both the forest and humans. (source)

I love the premise because it’s not “put ChatGPT in the woods.” It’s asking: what would an intelligent computer be like if plants rather than humans were the primary users?

4. Generative Muscle Stimulation — the AI’s output is your body moving. Researchers led by Pedro Lopes and Yun Ho published this at CHI on April 13. A multimodal system looks at the user’s surroundings, hears what they’re trying to accomplish and generates electrical-muscle-stimulation instructions that physically guide their body through a task. Instead of requiring a separately programmed routine for every action, the system generates movements contextually while constraining them using knowledge of muscles, joints and feasible motion. (source)

The associated Ars Electronica artwork Embodied AI takes the idea somewhere stranger: an AI communicates not through words but through involuntary bodily movement, including a “hallucinated” dance. (source)

That gives us:

prompt → reasoning → your arm moves

rather than:

prompt → reasoning → text

That’s one of the strangest AI interfaces I’ve seen recently.

5. Pathology Inducing Machine — knobs that physically mess with an AI’s brain. Raymond Chen and Jason Salavon built a physical control surface attached to an image diffusion system. Knobs, switches and electrical signals deliberately destabilize parts of the generation process and push images away from the model’s learned distribution. Instead of hiding latent space behind a friendly prompt box, the installation turns it into something resembling an unruly analog synthesizer. (source)

This one immediately suggests a project: an AI image synthesizer built like a piece of 1970s laboratory equipment. No text prompt after initialization—just physical knobs controlling concepts, abstraction, corruption, memory, influence, composition, etc., with a tiny CRT-style display. That’s considerably more interesting than another Stable Diffusion frontend.

Two persistent-world projects worth watching

WorldLines is explicitly trying to make persistent AI worlds rather than chat characters. World agents and character agents operate independently; NPCs have memory and continue living while you’re absent; characters are portable files that can move between worlds; and its roadmap includes multi-agent multiplayer. Version 0.2.9 shipped in June and its next phase includes agent-driven world creation. (source)

There’s also The ShAIre, a Godot 4 experiment in which LLM-driven inhabitants live inside a 3D fantasy valley. Characters forage, mine, cook, craft, build shelters, research recipes and maintain relationships, with an LLM handling their higher-level goals and reasoning rather than scripting each action. (source)

These are much closer to the “world keeps going after you leave” idea we discussed than conventional NPC chatbots.

The project I’d steal an idea from this week

It is moi + Pathology Inducing Machine.

You already have a little ESP32 device running and displaying PNGs, and you’re working toward SD-card audio playback. Rather than viewing these embedded projects solely as devices controlled by software, there’s an interesting inversion:

The physical object becomes the AI’s user interface.

Give an agent an intentionally tiny hardware vocabulary:

DISPLAY(image)
DISPLAY(text)
PLAY(sound)
LED(state)
KNOB(value)
BUTTON(event)

Then don’t design the application.

Give the agent a fictional purpose and let it decide what those extremely limited capabilities mean.

For example, make a little 1980s-looking HM-1 FIELD INTELLIGENCE TERMINAL with a screen, speaker, rotary encoder and three buttons. Instead of you designing menus, the agent gradually invents its own interface within those constraints. It could create symbols, sounds, status screens and interaction conventions—and persist them to storage.

That’s a much stranger experiment than “ChatGPT on an ESP32,” because you’d essentially be watching an AI invent a tiny operating system for a computer that you built for it.