Weird AI Projects — September 28

Weird AI Projects — September 28

This week I found a few things that feel like genuine thresholds rather than “another AI app.” The biggest theme is intelligence escaping the computer and getting small enough to inhabit cheap physical objects. There are also two unusually good persistent-world experiments.

Needle 2 — a 14 MB agent brain that can live inside tiny devices

This may be the most consequential maker-oriented thing I’ve found during these searches.

Cactus Compute’s Needle 2 is a 45-million-parameter model compressed into a 14 MB binary, requiring roughly 28 MB of RAM. It’s deliberately not a chatbot. It does one job: convert messy human language into structured tool calls.

Cactus reports roughly 500 tokens/sec on a Raspberry Pi 5, and says the memory footprint is small enough for newer ESP32-S3 configurations. A September 26 hardware demo on a stock Pi 5 showed local commands switching an LED in 78 ms, taking a photo in 76 ms and reading CPU temperature in 149 ms—with no cloud connection. (source)

This architecture is much more interesting than putting ChatGPT into every gadget:

“Turn the dial halfway and start the radio.”
↓
tiny local model
↓
set_volume(50)
radio_start()

No internet. No API fee. No giant computer.

And if the model isn’t confident, it can refuse or hand the request upward to a larger cloud model. (source)

That’s a compelling architecture for AI-native physical objects.

Pico-Faces — a $1-class microcontroller literally dreams faces

This one is delightfully ridiculous.

Developer Tim managed to run a latent diffusion transformer entirely on an RP2350, the microcontroller used by Raspberry Pi Pico 2-class hardware.

Not Raspberry Pi Linux.

A microcontroller.

The system starts with random noise and gradually produces a 128×128 human face. Two versions contain roughly 1.7M and 2.9M parameters; the models and inference code fit into under 4 MB of flash. Generation takes roughly 10–20 seconds. (source)

Adafruit Industries has already adapted the project to output over DVI/HDMI on its Fruit Jam board. (source)

This is obviously not replacing Midjourney.

That’s not why it matters.

The interesting thing is:

sensor → tiny generative model → image

can now happen inside something approximately the complexity of an Arduino project.

Imagine an electronic creature whose facial expression isn’t selected from PNG files. Its little neural network actually generates a new face representing its internal state.

That’s weird.

esp32-ai — an ESP32-S3 writing stories entirely by itself

And here’s the companion experiment.

Developer slvDev has squeezed a 28.9-million-parameter language model onto an ESP32-S3. It runs completely offline and produces text at about 9.88 tokens/sec. (source)

The trick is clever: only about 3.9 million parameters live in working memory. Another ~25 million parameters sit in flash and are retrieved as needed using a technique derived from Google’s Per-Layer Embeddings approach. (source)

The model isn’t secretly ChatGPT. It’s trained for TinyStories-style prose and cannot reliably perform general Q&A, coding or reasoning.

But that’s exactly why I like it.

We’re beginning to get purpose-bred tiny intelligences.

Imagine a cheap ESP32 object whose only cognitive ability is:

write strange weather forecasts

or

invent radio DJ banter

or

describe what its sensors detect

or

generate dialogue for one fictional character.

That feels much more interesting than squeezing a mediocre general assistant onto everything.

SimNet — what if every robot shared one persistent simulated Earth?

SimNet is an early 2026 robotics proposal with a genuinely big idea.

Robot simulations normally work like video-game matches:

create world → train robot → destroy world → repeat

SimNet proposes the opposite:

one enormous persistent simulated world that never resets.

Thousands—or eventually millions—of robots, software agents, humans and environments could coexist inside a distributed simulation. Their actions permanently change the environment, and experiences accumulate rather than disappearing at the end of a training episode. (source)

The analogy would be:

traditional robot training = arcade game

SimNet = World of Warcraft

That’s fascinating because it potentially gives robots something resembling history.

A robot doesn’t merely learn:

stairs are difficult.

It could learn:

robot 481 fell down those particular stairs yesterday, robot 912 found another route, and someone subsequently moved a crate into that route.

That starts looking much more like a society than a dataset.

It’s early and ambitious, so I’d treat it as a research direction rather than a mature platform—but it’s exactly the kind of conceptual leap worth tracking.

KamiBench — researchers put AI agents into an MMO that refuses to reset

KamiBench is testing a related idea in a real existing environment.

Instead of evaluating agents with thousands of identical clean tests, researchers drop them into Kamigotchi, an on-chain MMORPG whose world has already been operating continuously for more than a year.

The AI gets a wallet.

It enters the world.

It investigates its history.

It develops strategies.

It acts.

Then—critically—it comes back in another session carrying its notes and plans from before. (source)

Five experimental runs have already produced an amusingly human result:

agents remember bad ideas too.

They sometimes carry incorrect assumptions into future sessions rather than magically resetting to perfect knowledge. (source)

That’s important.

Most current agents effectively wake up every morning with amnesia plus a briefing document.

Persistent worlds create something different:

experience.

And potentially:

culture.

Corpus Corax — an intentionally unreliable human↔raven translator

This year’s new Hello Worlds! exhibition at Ars Electronica Center contains an experiment I love called Corpus Corax.

It’s presented as a speculative human-to-raven translator.

But the designers explicitly don’t claim they’re accurately translating ravens.

Instead, the system deliberately exposes how bizarre communication becomes when two intelligences perceive reality completely differently. Questions that seem obvious to humans become strange when reframed through another creature’s worldview. (source)

This is a fantastic inversion of AI translation.

Instead of:

AI eliminates misunderstanding

it’s:

AI lets us experience misunderstanding.

That’s a much more interesting artistic use of a model.

Doooodles — everyone’s drawings become one evolving organism

A new installation called Doooodles opened this month at Bit Studio in Bangkok.

Visitors draw simple doodles. Machine-learning drawing recognition interprets them, but instead of producing isolated AI pictures, everyone’s contributions enter one expanding communal canvas where drawings intertwine and form new relationships. (source)

It’s another example of something I’ve increasingly noticed:

The more interesting AI art isn’t:

person → prompt → image

It’s:

people → environment → AI → environment → people

The feedback loop becomes the artwork.

Diffuse Control — Beeple turns an entire museum collection into an instrument

Beeple opened Diffuse Control at the Madison Museum of Contemporary Art on September 19.

The installation uses twelve monumental screens and imagery from the museum’s permanent collection. Visitors interact with the system, while AI continuously recombines and transforms material from the collection into a constantly changing artwork. (source)

The idea I like isn’t necessarily the imagery.

It’s that the museum itself becomes the dataset/instrument.

You could apply that to almost anything personal:

your photographs
your record collection
your games
your writing
your TV archive
your music library.

Instead of generating generic AI media, the machine generates experiences from your own accumulated cultural material.

The one I think matters most this week

Needle 2.

Because it suggests a change in how we should think about little hardware projects.

We’ve tended to think:

ESP32 → buttons/sensors/display → deterministic program

and:

AI → powerful computer/cloud → conversation

Those worlds are collapsing together.

A $5–$20 microcontroller-class object can increasingly contain a tiny specialized reasoning layer whose only job is to understand intent and operate the machine.

That gives me a very different idea for the kind of ESP32 devices you’ve been building.

Don’t put “an assistant” inside one.

Give the object verbs.

For a little fictional radio:

TUNE()
SCAN()
IDENTIFY()
PLAY()
REMEMBER()
DISPLAY()
LIGHT()
SLEEP()

Then train or configure a tiny model specifically around those verbs.

Suddenly you don’t have menus.

You have behavior.

And combining this week’s projects makes the idea stranger still:

Needle 2 — tiny decision-making
Pico-Faces — tiny generative imagery
esp32-ai — tiny generative language
persistent-world techniques — memory
physical controls — embodiment

We’re getting remarkably close to being able to build a $30–$50 physical object with its own tiny, completely local personality—not by calling ChatGPT, but because the intelligence genuinely lives inside the object.

That is a rabbit hole I’d put near the top of the list right now.