FIELD
INDEX
The archive holds the documentation beside the fiction beside the machine. Records can belong to more than one shelf; their primary class tells you what they are, while tags preserve the routes between them.
Iteration
A month of Mesocosm taking shape, one screenshot at a time.
Mapping Books Into the Game
How six books are divided, resolved through GPT-2 Small Layer 5, and mapped into material for Mesocosm's game space.
Giving a Language Model Nerve Endings
Ariana Ramos's Embedding Sea, PI Zero's word-to-action bridge, and the experiment hiding between interpretability and self-sensing.
Case Studies in Visualization
Quantum Odyssey, an embedding terrain, and the promises a visual system makes about what can be seen, manipulated, and trusted.
The Pattern Isn't the Proof
A neuron need not be a singular unit of meaning to remain a useful address for discovery. On patterns, causal pressure, and what the starmap is actually for.
The Cases the Instrument Keeps
L5-N541 yields insight. An arbitrary expedition yields six hits and a clue. On exemplary cases, necessarily bespoke experiments, and what an honest starmap keeps visible.
Damage Its Own Ontology
Threads were technically accurate and experientially misleading. On visual grammar, design responsibility, STRATA, and an instrument willing to rebuild itself.
I Am Underwhalemanned
A Redditor requests neuron 2222. The drones return from neighbouring N2223 with six hits, five suggestive fragments, and one impeccable review.
Bear Says Oui Je Comprends
An equation, some broken French, and the difference between understanding an operation and speaking its formal language.
And Then the Second Mate Arrives
Start the recording, keep time, and launch every line of the L5-N541 shanty toward the neuron where it arrived.
The Ceremony Handles the Math
What KSP, Sailwind, and a navigable GPT-2 Small suggest about instruments that hide complexity without pretending it is not there.
One Style, Many Rooms
Twenty-five images from one style reference, shuffled through a sequence of monumental crops, collisions, ribbons, specimens, and quiet rooms.
Convergent Methodology
An initially sceptical reader asks whether rediscovery, failure, and predictive competence can reveal what Mesocosm is becoming as an instrument.
When the Threads Branch
A visual split becomes a causal question: peak tracing, intervention, held-out tests, and a more careful account of what the branches show.
L5-N541: And Then the Second Half Arrives
One neuron, eighty Melville routes, hundreds of controls, a causal intervention, and the moment the second half completes the pattern.
Relevant Math, Said the Bear
A found diagram, a confused Bear, and a note on why this is relevant math without being a literal schematic of a GPT-2 neuron.
The World on the Other Side of the Number
A style code becomes a landmark inside a model—and, through the traces gathered around its unusual string, an address in the public web.
Semantic Blackjack
Every typed hand resolves through GPT-2 to one of 3,072 places. The House holds two. Anywhere else pays.
The Universal Library, Inverted
Rooms, flowers, bears, signals, and concept plates from the iteration that compressed inexhaustible books into 3,072 addresses.
ZNOU / J-Space
A deterministic function from strings to neurons, and the atlas, ledger, instruments, and game built around it.
It Was Cold
At J5-N38, structural probes become a direct intervention into cold—and the neat story of a single semantic neuron comes apart.
The Starmap Becomes a Place
A moving flight through the Godot starmap, with nine development captures and the technical path from corpus landing to cross-layer thread bundles.
How to Draw a Thread
Six attempts to preserve individual routes across layers, ending in a visual grammar that looked suspiciously like a circuit diagram.
The Paper That Opened the Model
One paper made a language model feel worth opening up. Years later, the same downloaded copy of GPT-2 Small is still being explored.