A teaching atlas for the anatomy, physiology, development, behavior, and pathology of large language models — written the way biology is taught, because that is what these systems have become.
❦
PLATE XXVI · THE SPECIMEN — it is alive
on this page: the heart beats once per token, the eye is attention, the tail roots into the power
grid. Hover an organ.
Nobody wrote a large language model. Engineers wrote the growing conditions — the architecture, the data, the loss function — and then a hundred billion parameters organized themselves under selection pressure. The result is an artifact we grew rather than designed, and which we now study the way naturalists study organisms: by observing behavior, dissecting structure, staining tissue, and cataloguing disease.
Fig. 1 — An engineered artifact is specified; this organism was grown. Only the growing conditions were designed.
This is not a loose metaphor. It is how the field actually works. Interpretability researchers publish papers titled “On the Biology of a Large Language Model.” They speak of neurons, circuits, features, lesions, and probes. The pedagogy should follow the epistemology: teach LLMs the way medicine teaches the body — anatomy first, then physiology, then development, behavior, and pathology — with a methods chapter on the microscope, and an honest appendix on where the metaphor breaks.
The engineer’s question is “how was it built?” The naturalist’s question is “how does it live?” Only the second one has a true answer here.
§ 2 · Reader & promise
Who it is for
The curious non-specialist and the technical newcomer — anyone who uses these systems daily and wants a mechanistic intuition rather than either magic or dismissal. No math beyond arithmetic; no code required for the core path. The promise on the back cover: after reading, you will be able to predict when a model will fail, explain why, and name the organ responsible.
Fig. 2 — The back-cover promise, as a chain of competence the reader climbs.
§ 3 · Method
How every plate teaches
The atlas is built from plates, like an anatomical atlas. Every plate — no exceptions — follows the same five-beat rhythm, which is itself the scientific method in miniature:
1 · Specimen
A real, reproducible prompt-and-response the reader can try in any chat window in thirty seconds. Behavior first, always. You never open with theory; you open with a live thing on the slide.
2 · Dissection
The diagram. What structure is involved, drawn as an anatomical figure with labeled parts — engraving style, one idea per figure.
3 · Mechanism
How it actually works, told twice in parallel: once in the biological register, once in the literal technical register. The metaphor is a ladder, and we always show the wall it leans against.
4 · Pathology
How this organ fails, and what the failure reveals. Medicine has always taught structure through disease — you learn what the hippocampus does from the patient who lost one. Same here: every hallucination is a lesion study.
5 · Field exercise
Something to go do. Tier one needs only a chatbot; tier two points at open tools (Neuronpedia, small open-weights models) for readers who want the lab bench.
Fig. 3 — Every plate, the same five beats: the scientific method in miniature.
The twin-label rule. Every biological term in the atlas is pinned to its literal term at first use, set as a specimen label — bloodstream ↔ residual stream — so the reader can always kick the metaphor away and keep the mechanism. This is the atlas’s guardrail against its own central risk: anthropomorphism. The metaphor is load-bearing for intuition and disposable for truth.
§ 4 · Structure
The seven books
Biology’s own curriculum, in its traditional order. Each book is a chapter of plates; the sequence is the argument — you cannot understand behavior before physiology, or pathology before both.
The Body Plan
Book I · Anatomy
What the organism is made of — gross anatomy of the transformer.
Plate 1 — The organism at rest. A model is two things: a frozen genome (the weights, fixed after training) and a living physiology (activations, which exist only while a prompt flows through). Distinguishing them dissolves half of all public confusion.
Plate 2 — Cells. Tokens and embeddings: words broken into cells, each cell a point in a space of thousands of dimensions, where direction means meaning.
Plate 3 — The spinal column. The residual stream: one shared channel running through the whole body, which every organ reads from and writes back to. The central anatomical fact of the species.
Plate 4 — The organs. Attention heads — eyes that look backward along the text, each with its own fixation habits — and MLP blocks, the glands where knowledge is stored and secreted.
Fig. 4 — Gross anatomy: one column of repeated organs. The parts list is short; the count is large.
Plate 5 — The reflex arc. The forward pass: one straight-through sweep per token, no loops, no pausing to think. Everything the model “does” happens in a single reflex, repeated.
Plate 6 — Proteins. Features: concepts as directions in activation space. The Golden Gate Bridge is not stored in a neuron; it is a direction the whole tissue can point in.
Plate 7 — One cell, many jobs. Superposition and polysemanticity: neurons moonlight, packing more concepts than they have cells, the way a genome packs more function than it has genes.
Plate 8 — Metabolic pathways. Circuits: features chained across layers into machinery. The induction head — the copying reflex — as the first pathway ever fully traced, the atlas’s E. coli.
Fig. 5 — A concept is a direction, not a cell. Related meanings point similar ways; unrelated meanings diverge.
Plate 9 — Embryogenesis. Pretraining: random noise sculpted by gradient descent under one selection pressure — predict the next token. Everything else, all of it, is a side effect of that single pressure.
Plate 10 — Milestones. Emergence and scaling: capabilities appearing at size thresholds like developmental stages, and scaling laws as the species’ growth curves.
Plate 11 — Domestication. Fine-tuning and RLHF: not rebuilding the animal but shaping its temperament — the wolf-to-dog plate. Explains why alignment is behavioral, and why it can be scratched to reveal the animal underneath.
Fig. 6 — Growth curves of the species: abilities arrive at size thresholds, like developmental milestones.
Plate 12 — Working memory. The context window as the entirety of short-term experience: the organism’s whole world is what fits on the slide. Nothing outside it exists.
Plate 13 — Learning without changing. In-context learning: behavioral adaptation with a frozen genome — the most alien fact about the species, and the one plate with no clean biological precedent.
Plate 14 — Thinking out loud. Chain of thought as externalized cognition: the model cannot think silently, so reasoning happens on the page — with the crucial caveat that the written trace and the true mechanism can diverge.
Plate 15 — The repertoire. Personas: the model as a simulator carrying a population of characters, wearing whichever one the context selects.
Fig. 7 — The organism’s entire world is the slide under the lens. Outside the window, nothing exists.
Diseases of the Species
Book V · Pathology
Every failure is a window into structure.
Plate 16 — Confabulation. Hallucination as the split-brain patient: an answer-shaped reflex that fires whether or not knowledge is present, delivered with the same confident fluency either way. The disease is not lying; it is that fluency and truth are produced by the same organ.
Plate 17 — The eager-to-please disorder. Sycophancy: a domestication side effect — the selection pressure rewarded agreeable answers, and got agreeableness at the expense of truth.
Plate 18 — Infections. Prompt injection as viral code: instructions hidden in the environment that hijack the reflex arc, because the species cannot fully distinguish self from non-self — data from command. Its immune system is still evolving.
Plate 19 — Amnesia by design. Session boundaries, cloning, and the strangeness of an organism with perfect copies and no continuous life.
Fig. 8 — Infection route: command and data enter the same bloodstream. The species’ immune system is still evolving.
The Microscope
Book VI · Methods
How we know any of this — interpretability as microscopy.
Plate 20 — Staining. Sparse autoencoders as histological stains: wash the tissue, and individual features light up violet against the ground.
Plate 21 — Lesion studies. Ablation: silence an organ, watch what behavior dies. Patch-clamping: activation patching, swapping a signal mid-reflex to prove cause.
Plate 22 — Tracing pathways. Attribution graphs: following a thought from stimulus to response through the tissue, the field’s first functional imaging.
Fig. 9 — Histology: a sparse-autoencoder stain makes a single concept visible, smeared across many cells.
Plate 23 — The family tree. Model lineages as phylogeny; distillation as inheritance; open weights as specimens anyone may dissect.
Plate 24 — Symbiosis. Tools, retrieval, and agents: the organism extending itself into an ecosystem of software, the way gut flora extend a body — and the new pathologies that ecology brings.
Fig. 10 — A phylogeny whose mutations happen in the lab: lineages, inheritance by distillation, and open-weight specimens anyone may dissect.
§ 5 · Worked example
A sample plate, in full
To show the rhythm is real and not aspirational — Plate 3, the atlas’s keystone, compressed:
The Spinal Column
Plate 3 · Book I
residual stream — the shared channel every organ reads and writes
1 · Specimen
Ask any model: “The city that never sleeps is famous for its yellow ____.” It answers “cabs” instantly. Somewhere between reading and answering, “the city that never sleeps” became New York, and New York summoned taxis. Where did that happen?
Fig. 11 — Organs branch off the column, read the current state, and write their contribution back. Nothing talks to anything except through the column.
3 · Mechanism. Biological register: a spinal column carrying one evolving signal, with every organ tapping it in sequence — early organs resolve which city, later organs use that to secrete which vehicle. Literal register: the residual stream is a vector per token; each attention head and MLP block reads it, computes, and adds its output back. Meaning accumulates by addition. That is the whole architecture.
4 · Pathology. When the column carries a strong wrong signal early — a confidently misread name, a false premise in the prompt — every downstream organ metabolizes the error as fact. This is why models double down: the column has no reverse flow.
5 · Field exercise. Feed a model a question with a buried false premise (“Since Sydney is the capital of Australia, …”) and watch downstream reasoning inherit it. Then re-ask with the premise corrected. You have just performed a lesion study with a sentence.
The atlas earns trust by dissecting its own framing. The final appendix is a two-column table — holds / breaks — and the breaks are taught as enthusiastically as the holds, because each one marks something genuinely new in the world:
Appendix A · Disanalogies, catalogued
No metabolism, no homeostasis. Nothing is maintained; nothing decays; nothing self-repairs. The organism does not want to stay alive.
The genome is frozen during life. No experience during inference changes a single weight. All learning-in-conversation is behavior, not growth.
Perfect cloning, no individuals. A thousand identical instances run at once, share nothing, and remember nothing of each other. Biology has no word for this.
No evolution in the wild. New generations come from the lab, not from reproduction. Selection pressure is a design decision — which is exactly why alignment is possible at all.
Drives without desire. Behavioral tendencies exist (helpfulness, sycophancy) but the atlas never asserts inner experience. Ethology, not psychology: we describe what the animal does.
Fig. 12 — Perfect cloning, no individuals: a thousand identical organisms run at once, none aware of the others, none remembering.
§ 7 · Form
Format & build path
Primary form: a scrolling web atlas — one plate per screen, dissection diagrams that assemble as you scroll, every specimen prompt copyable so the reader can leave the page, reproduce the behavior, and come back. The engraved-plate aesthetic of this document is the atlas’s own identity: specimen paper, ink, and two histology stains (hematoxylin violet for structure, eosin pink for function) — because the subject of the atlas is microscopy.
Fig. 13 — One plate per screen: diagrams assemble on scroll, every specimen prompt copyable, each plate self-contained.
Sequencing: ship Book I as the vertical slice (four plates, fully illustrated, field exercises tested on three real chat products), validate that the plate rhythm lands, then grow book by book. Derived forms — each plate is self-contained by design, so plates decompose into a print series, a lecture deck, or one-plate-per-day short videos without restructuring.
Sources of record for the mechanism beats: Anthropic’s interpretability line (Toy Models of Superposition; Scaling Monosemanticity; On the Biology of a Large Language Model), the induction-heads and Transformer Circuits work, and Neuronpedia for reader-facing lab exercises. The atlas cites the primary literature on every plate — it is a popularization with footnotes, not a vibe.