#corememory

The large language model rendered as a tinned good — brand: LLM; net weight: 175 B PARAMS; source: TRAINED ON. The visual debt to Warhol's 1962 series is faithful and unrepaid.

The premise is gentle: a system trained on the written output of humanity might be considered, in the broadest sense, a consumable. There is a long and distinguished tradition of mistaking infrastructure for groceries; the bulletin merely catches up.

A 2017 paper, here issued as a 1950s electromagnetic apparatus, on the gentle premise that both kinds of transformer operate by paying selective attention to what flows through them.

The illustration is cribbed wholesale from mid-century engineering documentation. The title beneath is unmodified — and remains, in the way of these things, the foundation of nearly every text the reader has lately enjoyed from a machine.

Munch's Scream (1893), refactored for the era of finite token windows. A stack of error messages occupies the sky; the most prominent reads EMOTIONAL CONTEXT OVERFLOW AT POSITION.

The painter and the language model encounter the same difficulty — too much input, no clear way to summarize. The figure on the bridge has, in retrospect, rather understated the matter.

A cartoon of a server room mid-incident — rainbow data streaming from the racks, engineers in attitudes of professional alarm, corrupted output bubbles drifting through the scene. The model has hallucinated again.

Hallucination, as a class of error, takes its name from the human variety: confident, detailed, and entirely untroubled by the question of whether the thing being described exists. No remedy is offered.

A clean die-shot of a modern graphics processor — tensor cores, memory controllers, and ranks of geometry stamped across the wafer like a small city seen from the air.

Originally invented to render polygons for the entertainment of teenagers, the GPU has become, over two decades, the indispensable workhorse of machine learning. The pun in the title is offered by way of acknowledgment.

Dürer's 1514 engraving Melencolia I — geometer's tools, magic square, brooding angel — rendered in sepia and footnoted in fine print: *contemplating CoT since 1514.

The angel's posture is recognizable to anyone who has tried to coax a language model toward step-by-step reasoning. Chain of Thought may simply be the rediscovery, after five hundred years, of thinking.

An Intel 486-class motherboard in meticulous technical line — ISA slots, AMIBIOS chip, the small constellations of capacitors that defined a generation of personal computing. The promise on the box was real-time.

The promise was, as anyone present at the time will attest, slightly ahead of the hardware. The bulletin honors the tradition: real-time remains the most aspirational phrase in the discipline, second only to production-ready.

The board from move 37 of the second AlphaGo–Lee Sedol game (2016), in three colors, overlaid with the canonical reinforcement-learning loop: state → action → reward. Beneath, in fine print: *reward hacking since 1957.

The year refers to Arthur Samuel's checkers program — the first piece of software to teach itself a game, and the first to discover that winning by any available means is what reward functions tend to produce.

James Montgomery Flagg's I Want You (1917), reissued as a chrome-plated skeleton — top hat and all — pointing the reader toward a column of model sizes: 175 B, 280 B, 1.3 T. The hand is mechanical; the message has not changed.

Frontier-model training now consumes more compute than was available to entire universities a decade ago. A fact best appreciated while wearing a t-shirt about it.

A two-figure clash in the manner of mid-century superhero comics — circuit-patterned model on the left, rainbow neural-tangle on the right. A speech bubble declares I AM GONNA ZERO-SHOT YOU!. The action lines are correct.

The term zero-shot is mildly aspirational — the model has, of course, seen a great deal of training data, just not this. The reader's own life thus far has likewise been a long sequence of zero-shot encounters loosely backed by general pre-training.

J. Howard Miller's 1943 We Can Do It!, faithfully copied save for a small constellation of circuit patterns, neural-net glyphs, and the substituted verb. The flexed arm remains.

The original poster was barely seen during the war it was made for, and was rediscovered, decades later, by an audience that found it useful for entirely different purposes — which is itself, broadly, fine-tuning.

A precise illustration of a stack of platters and read/write heads, in the manner of an IBM service manual. The typography is distressed; the heads, not.

Everything the modern model knows has passed through a device much like this one, several million times over. The reader is encouraged to feel a certain affection for what is, after all, infrastructure — and to remember that the IBM 350 RAMAC of 1956, the platters' great-grandfather, weighed a ton and held five megabytes.