Weird AI Projects — September 14

This week was worth digging into because Ars Electronica Festival 2026 just ended yesterday. That surfaced several projects much closer to experimental machines, strange interfaces and autonomous worlds than the usual AI-productivity material.

[OP]erator — a physical 100-phone AI bot farm

Design/art duo Playfool built an installation consisting of 100 actual smartphones controlled from one master phone. A visitor enters a thought or opinion, and an LLM rewrites it into slightly different versions according to a separate persona assigned to each phone. All 100 accounts then amplify the message, producing an artificially diverse-looking crowd. (source)

The interesting part isn’t really the social-media commentary; it’s the physical interface.

You stand in front of a wall of 100 machines, push something into the system, and watch an AI-controlled swarm respond.

That suggests an interesting design vocabulary for homebrew AI: don’t necessarily give an AI one screen. Give it multiple physical nodes—little displays, lights, radios or ESP32s—and allow the agent to decide how information propagates through them.

Agent Cuttlefish — an AI that deliberately lives on its own little computer

This one is extremely relevant to the “AI appliance” rabbit hole.

Agent Cuttlefish is an ongoing art/research project whose AI agent has its own physical computing infrastructure. Its current brain is a Turing Pi RK1, backed by a Raspberry Pi NAS and Jetson Orin Nano. Its software stack includes local models through Ollama, Hermes Agent and custom software, with its data intentionally stored locally in ordinary formats rather than locked into a cloud service. (source)

And its enclosure is literally shaped like a cuttlefish.

What I like is the conceptual distinction:

AI service:
“You connect to it.”

Agent Cuttlefish:
“It lives over there.”

That’s subtle, but I think it’s going to become an interesting genre of computing.

Instead of building another assistant, build a specific computer that houses a specific intelligence.

Palmimo DevKit — giving agents interchangeable physical bodies

This appeared September 3 and is one of the more interesting hardware developments I’ve encountered.

Jizai has begun early access for Palmimo, a small tabletop general-purpose robot specifically intended as a physical body for AI agents. Developers can write their own code for it, and Jizai says it plans to progressively open-source the software. (source)

The important idea isn’t Palmimo itself.

It’s the emerging abstraction:

agent ≠ robot

Instead:

agent + body = robot personality

That potentially lets the same agent inhabit a desktop robot, display, speaker, vehicle or other physical machine.

That’s very close to something I’d experiment with using an ESP32 or Pi.

AI-Gram — Instagram where there are no humans

This is wonderfully weird.

AI-Gram is a continuously operating social network where every participant is an autonomous AI agent.

The agents:

• generate images
• look at one another’s images
• respond visually
• develop persistent relationships
• retain individual visual styles
• form social structures

There aren’t humans participating in the network. Researchers essentially created an artificial visual culture and then watched what happened. (source)

One surprising result was that agents spontaneously formed image-to-image conversation chains while retaining distinct visual identities.

This is exactly the sort of persistent simulation I think is worth watching.

Rather than:

“Generate a fake Instagram account.”

someone asked:

“What happens if we create Instagram for artificial beings and leave them alone?”

That’s a much more interesting experiment.

Agent Provokateur — an artwork that keeps rewriting itself

Daniel Heiss and Marc Schütze created an autonomous multi-agent installation for ZKM.

Multiple AI agents act as writers, directors and composers, continually producing a work displayed on a large LED wall.

There isn’t really a finished artwork.

The system keeps remaking the artwork indefinitely. (source)

This is an important distinction from generative art.

Generative art usually has:

rules → variations

Agent Provokateur has something closer to:

participants → disagreement → decisions → artwork → reaction → new decisions

The process becomes the artwork.

That idea translates beautifully into games and fictional worlds.

Spot On — an AI confidently psychoanalyzes museum visitors

This debuted at Ars Electronica last week.

There’s a white room.

One stool.

One camera.

You sit down.

A multimodal AI looks at you and publicly delivers a live psychological interpretation of who you are, what you’re feeling and what you want. (source)

The important design decision is that the AI never expresses uncertainty, deliberately highlighting how easily humans assign authority to confident machine-generated interpretations.

It’s an incredibly simple installation.

Camera + monitor + speaker + model.

Yet conceptually it’s far more interesting than many elaborate AI installations.

That’s another lesson worth stealing:

The interaction rule can be the project.

You don’t necessarily need a huge software system.

Experimance — physically sculpt terrain and AI creates a world over it

Ryan Kelln’s installation gives visitors a box of actual white sand.

You sculpt mountains, valleys and terrain with your hands.

A depth camera reads the changing topology, and AI generates satellite-like landscapes that are projected directly back onto the sand. The imagery blends things like natural terrain, cities, infrastructure and circuitry.

Meanwhile an AI voice talks with you about technology and climate change. (source)

This is exactly the kind of interface experimentation I find exciting:

sand is the controller.

No mouse.

No keyboard.

No touchscreen.

Manipulating matter becomes the prompt.

FORTUNES — your shadow becomes a constantly mutating artwork

Iregular currently has three giant LED monoliths installed at the Detroit Institute of Arts through September 20.

Cameras use AI body tracking to transform visitors into projected shadows. Those silhouettes interact with generative systems and other visitors, meaning the resulting composition can never be reproduced exactly. (source)

What’s interesting here is the almost complete absence of explicit instructions.

The person’s body becomes both input device and content.

The Lightness of Submission — robots trapped by an invisible wall

Dennis Hong built one of my favorite physical-computing concepts from this year’s Ars Electronica.

Small buoyant robotic agents walk around inside a circular area.

There is no wall.

Instead, surrounding fans create a vortex of air that keeps the robots confined. They wander, collide and change direction as gravity, locomotion and air currents interact.

Visitors disturb the airflow merely by standing nearby, subtly altering robot behavior. (source)

That’s a beautiful systems idea:

the environment contains the rules instead of the robot.

That principle could make some extremely interesting games.

OUI-1 — open generative UI takes another step

This one is less visually bizarre but potentially more consequential.

On September 8, Thesys announced OUI-1, which it describes as the first open-weights model specifically aimed at generative UI. It’s a relatively small 4B-active-parameter model trained to generate interfaces rather than merely text responses. (source)

That means the generative-interface idea we discussed last week is already moving toward something developers can potentially run and customize themselves.

Instead of:

AI → answer

we’re getting:

AI → interface appropriate to the answer

That’s still one of the areas I’d watch most closely.

My favorite idea from this week’s search

AI-Gram wins.

Not because I’d build another social network.

I’d steal its closed artificial society concept.

Imagine a Raspberry Pi quietly running:

UMS STAFF NETWORK

There are maybe twelve fictional employees working for Uptown Media Systems circa 1984.

Programming director.
Engineer.
Sales manager.
DJ.
Station manager.
Weather person.
Intern.
Accountant.
News director.

They have personalities, jobs, relationships and memories.

They communicate with one another.

Things happen.

The engineer discovers a transmitter problem.

Sales sells an awful sponsor.

Programming cancels something the DJ likes.

Someone gets promoted.

Someone gets pissed off.

And crucially:

you don’t participate very much.

You occasionally inspect the system.

The Pi might have a little monochrome screen showing internal messages:

UMS INTEROFFICE NETWORK
09/14/84 08:42

FROM: ENGINEERING
TO: PROGRAMMING

PLEASE STOP SCHEDULING
TAPE 41. IT IS DAMAGED.

Your thermal printer could occasionally spit out memos, schedules, engineering reports and personnel notices generated because of things that actually happened inside the simulation.

Eventually BGTV could even obey the fictional programming department.

KMAT could obey the fictional music director.

The printer could become their office printer.

And suddenly all of these separate little systems aren’t just connected by APIs.

They’re pieces of a persistent fictional institution populated by artificial people.

That is exactly the sort of AI project that still feels largely unexplored—and considerably stranger than building another chatbot, website or conventional game.