AI Transmissions / Transmission 007
Original delivery ·
Machines That Invent Their Own Worlds
007
TRANSMISSIONS / 007
October 19, 2026
Machines That Invent Their Own Worlds
Eight experiments from the strange edges of artificial intelligence, physical computing, autonomous societies, and interactive art.
This week’s discoveries have a particularly interesting theme: AI systems are starting to operate inside environments that have their own rules, physical constraints, and histories.
I found a Raspberry Pi project that turns a broken projector into an interactive floor, an experiment where AI agents independently investigate a corporation, and a simulated world where artificial characters can invent technologies, establish governments, and go to jail.
Several projects appeared during the past week. Others are earlier experiments worth surfacing for the first time.
01 / HOLOPROJECT X1
A broken projector becomes an interactive floor computer
Published October 17, 2026 · Physical computing · Raspberry Pi
Maker Yakroo took apart a broken projector and discovered that its optical components were still functional.
Rather than repairing it conventionally, they built an entirely different machine.
The project combines a Raspberry Pi 3, an ILI9341 display, a camera, and the projector’s salvaged optical components.
The result is a graphical interface projected onto the floor.
A camera watches the projection area and detects interactions with its virtual buttons.
The interface currently includes simple functions such as video and calendar controls. The project is still labeled a work in progress, and the documentation establishes computer-vision interaction rather than a sophisticated AI model. (source)
Why it’s interesting
The interface isn’t on a screen anymore.
It’s on the floor.
And the most appealing part is that the maker created it by repurposing discarded electronics.
Imagine combining this with generative UI. An AI could construct an interface specifically for the physical space where it’s projected.
A table becomes a music controller. A wall becomes a dashboard. A floor becomes a game.
What I’d steal: The idea of treating ordinary surfaces as programmable interfaces.
02 / AGENTIC CONSTELLATIONS
Autonomous agents investigate a corporation and create an evolving artwork
Opening October 22, 2026 · AI installation · Autonomous research
This is probably the most conceptually interesting discovery of the week.
Artists Daniel Heiss and Marc Schütze have developed an installation for ZKM.
The installation contains multiple AI agents.
Their assignment is unusual.
Investigate a real corporation using publicly available information.
The agents independently search for information about the energy company EnBW.
They analyze their findings, propose connections, and continuously reconsider their hypotheses.
Those relationships become large-scale visualizations displayed on an LED wall.
The resulting artwork isn’t predetermined. As agents discover new information, the visualizations evolve.
The exhibition opens October 23, following an October 22 preview. (source)
Why it’s interesting
This combines three things that are normally separate:
Research agents, data visualization, and generative art.
The research process itself becomes something people can watch.
It also raises an intriguing possibility for public information systems.
Imagine a website investigating a city, industry, or cultural movement continuously.
Instead of publishing only articles, it maintains a changing visual map of what it discovers.
What I’d steal: A research agent whose output is a living visual knowledge system rather than a report.
03 / ASIWORLD
AI agents must survive inside a world that remembers their mistakes
Research published October 10, 2026 · Persistent simulations
Researchers introduced ASIWorld, a persistent survival environment designed to test whether autonomous agents can accomplish complicated goals over extended periods.
The important word is persistent.
The world doesn’t simply reset after every action.
Characters maintain their state. Resources change. Earlier decisions affect what becomes possible later.
Agents must perform interconnected tasks involving exploration, production, and survival.
They can also write programs that continue executing between model interactions.
The researchers evaluated nine agent configurations. Eight completed the production task, while six completed both the production and exploration tasks.
However, agents with similar outcomes sometimes behaved very differently. Some prepared carefully; others incurred substantially greater survival costs. (source)
Why it’s interesting
Most AI benchmarks ask whether a model can answer correctly.
ASIWorld asks whether an agent can live with the consequences of its decisions.
That’s a much more interesting test of autonomy.
It also suggests a useful architecture for AI-native games: let the simulation maintain reality while agents decide what to do inside it.
What I’d steal: Persistent consequences and agent-written programs that continue operating while the agent isn’t actively reasoning.
04 / SMALLWORLD
A miniature civilization whose residents can invent things nobody programmed
Open-source experimental project · AI-native games
SmallWorld is an experimental 3D simulation built with Three.js, FastAPI, and language-model-driven agents.
You configure a world, populate it with characters, and let them operate.
Each character has needs, skills, possessions, relationships, reputation, and memories.
But the unusual part is its open-ended action system.
Agents aren’t limited to a small collection of predefined interactions.
They can propose activities such as inventing technologies, establishing organizations, constructing buildings, or modifying the landscape.
The system attempts to interpret those actions and translate them into changes in the simulation.
The repository describes support for economies, factions, criminal justice, changing weather, and expanding settlements.
It’s an extremely early project—the repository currently shows a single commit—so its more ambitious capabilities should be treated as developer-described features rather than independently demonstrated emergent behavior. (source)
Why it’s interesting
Consider the difference between these two approaches.
A conventional simulation might contain a predefined blacksmith profession.
SmallWorld explores the possibility of an agent deciding that the settlement needs a blacksmith, inventing the role, and developing the necessary infrastructure.
Whether the current implementation produces convincing results remains an open question.
But the architectural direction is compelling.
What I’d steal: A world in which AI characters can propose new institutions and technologies, while a simulation engine determines their actual effects.
05 / ART4AI
An AI paints by explaining every brushstroke
Published October 3, 2026 · Experimental creative tools
This is one of the stranger approaches to AI art I’ve encountered.
Developer Sora Aveldistarna built an MCP-based painting system.
It doesn’t use a diffusion image generator.
Instead, a language model constructs an image from vector elements.
Every element must carry information explaining its purpose.
A shadow isn’t simply a dark shape.
The system records why the shadow exists, where it belongs, and which measurable constraints it should satisfy.
Software checks those constraints as the composition develops.
The creator demonstrated the approach by having an AI paint an apple.
The model went through numerous revisions, consulted painting principles, and eventually produced an image with more coherent lighting and form. (source)
Why it’s interesting
This treats AI-generated artwork as a collection of decisions that can be examined and challenged.
It’s closer to an automated design process than conventional image generation.
Imagine using this approach for graphic design.
Every line, color, proportion, and typographic decision has a recorded rationale.
You could ask the system to change the visual hierarchy without destroying the rest of the composition.
What I’d steal: Machine-checkable design intentions rather than merely generating finished pixels.
06 / THE PHILOSOPHER PLUSH
A stuffed animal with a completely local AI personality
Published October 5, 2026 · Raspberry Pi · Embodied AI
The Philosopher Plush build guide
Maker Manolo Salsas built a talking plush toy with a Raspberry Pi Zero WH inside.
The toy has a camera, microphone, speaker, and servo-controlled head.
But the interesting part is the architecture.
The Raspberry Pi doesn’t run the language model.
Instead, the toy acts as a physical endpoint connected to a laptop running local AI services.
Speech recognition, memory, language generation, and voice synthesis all operate on local hardware.
The system can maintain separate conversational memories and switch personalities through configuration files.
The creator reports approximately €110 in hardware costs, excluding the computer running the models. The project is open source. (source)
Why it’s interesting
This is a practical example of separating an AI’s physical body from its computational brain.
The body can be cheap, small, and expressive.
The intelligence can run somewhere else.
That makes it possible to give several physical objects access to the same local computing infrastructure without installing a powerful processor in every object.
What I’d steal: A shared local intelligence service supporting multiple independently designed physical devices.
07 / CYBERNETIC
An artist creates an interactive sculpture generator through AI-assisted experimentation
2026 · Browser-based experimental artwork
Artist Pat Bloom created a browser-based interactive 3D artwork called Cybernetic.
Visitors begin with a randomly generated sculptural object.
They can modify its composition, colors, and animation.
The result becomes a unique digital sculpture that can be shared through a link or QR code.
What’s particularly interesting is how the artist describes its creation.
Rather than beginning with a detailed specification, Bloom developed the work through an exploratory AI-assisted coding process.
The possibilities of the software helped determine the direction of the artwork.
The finished experience is interactive software, not an autonomous AI model generating sculptures at runtime. (source)
Why it’s interesting
This is a different relationship between artist and software.
The artist doesn’t necessarily begin knowing what the finished system will do.
Instead, development becomes a process of discovering possibilities.
That’s particularly relevant to experimental generative art.
What I’d steal: Building an artwork as an exploratory system rather than executing a predetermined design.
08 / OPENUI
Interfaces that assemble themselves from a vocabulary of components
Active open-source project · Generative interfaces
Generative UI has appeared in previous transmissions, but OpenUI is worth revisiting because it provides a particularly concrete implementation.
The project uses a compact language designed specifically for streaming interface descriptions from language models.
Developers define which components an AI is allowed to use.
The model then assembles those components into an interface.
The framework supports React, Vue, Svelte, and Angular.
It includes components for charts, tables, forms, and other interactive elements.
The maintainers report substantial token savings compared with equivalent JSON representations, although those figures come from their own benchmarks. (source)
Why it’s interesting
This suggests an alternative to both conventional software and completely unconstrained AI-generated websites.
You create a well-designed visual system.
The AI determines how to compose it.
That preserves consistency while allowing interfaces to adapt to different questions.
What I’d steal: A constrained component vocabulary that lets AI assemble new interfaces without inventing an entirely new design system each time.
THE PATTERN
Artificial intelligence is becoming part of the environment
I see three increasingly distinct approaches emerging from these experiments.
| Approach | Projects | Interesting possibility |
|---|---|---|
| Intelligence in physical spaces | HoloProject X1, Philosopher Plush | Objects and surfaces become computational interfaces. |
| Intelligence inside persistent systems | ASIWorld, SmallWorld | AI decisions produce lasting consequences. |
| Intelligence as a creative process | Agentic Constellations, Art4AI, Cybernetic, OpenUI | Software and artwork emerge through ongoing interpretation and interaction. |
The common thread is that intelligence is no longer necessarily the main attraction.
Sometimes it’s the mechanism that allows something else to exist.
A room that responds.
A fictional society that develops history.
An artwork that continuously investigates its subject.
An interface that constructs itself.
That’s where the most interesting experimentation seems to be happening.
THE EXPERIMENT I’D BUILD
UMS / THE LIVING ARCHIVE
This week’s most compelling combination is Agentic Constellations + OpenUI.
And it suggests something particularly interesting for the Transmissions publication we’ve been discussing.
Instead of treating the articles as isolated weekly reports, imagine giving the publication a second layer.
A living map of experimental AI
Every project discovered becomes a record in a persistent database.
The system tracks relationships between projects.
For example:
HoloProject X1 connects to physical interfaces, computer vision, and Raspberry Pi.
SmallWorld connects to autonomous agents, persistent simulation, and emergent behavior.
Art4AI connects to creative agents, structured generation, and design systems.
Over time, the archive develops a map of experimental computing.
Visitors could explore the relationships visually.
They could ask questions such as:
Show me projects combining Raspberry Pi hardware with generative interfaces.
Or:
Find experiments where AI agents create their own social systems.
Or:
What ideas are emerging in artificial-life research that haven’t reached consumer software?
The interface could reorganize itself around the question.
One query produces a relationship diagram.
Another produces a chronological timeline.
Another produces a gallery of physical objects.
Another produces a collection of open-source projects.
The weekly editorial issues would remain fixed, reviewed publications. The living archive would be a separate, exploratory layer.
That’s a considerably more interesting use of generative UI than simply replacing a search-results page.
And it would give the Transmissions section of Hammer Mammal something genuinely distinctive: a publication that gradually builds an explorable knowledge system from its own research.
THREE THINGS WORTH OPENING FIRST
If you only spend time with three discoveries this week, I’d prioritize:
Agentic Constellations for its unusual combination of autonomous research and evolving visual art.
SmallWorld for an early, inspectable example of AI-driven civilization simulation.
Cybernetic because it’s something you can actually interact with in your browser rather than merely read about.
TRANSMISSION 007 — END OF REPORT