Mark Levy/Writing/The Word Warp Drive

An 8-bit engineering deck: a tall warp core glows blue and white between amber rings, colored console bars line the walls, an engineer works at the left console, and a gold-skinned android in a mustard uniform holds up a padd showing a blank.

[ Tier 1 · Deck 1 of 2 ] Engineering: the core itself

The Word Warp Drive

Tier 1 begins: how a large language model actually works, one guessing game at a time.

Q
PromptParis is the capital of ___
CompletionFrance 92%
Captain, this is the warp core: fill in the blank. Everything else is ship, and Lore is waiting below.

Language-model magic is one repeated act: given text, predict a plausible next word. Chat style and apparent memory come from looping that act.

Your typed input is the prompt. The full block read each pass (prompt + hidden rules + history) is context.

Think un-deletion machine: remove words from real text, train the circuit to restore them, repeat at internet scale.

This is Deck 1 of Tier 1. Deck 2 keeps the tier and shifts to runtime behavior. Demos are illustrative mechanisms.

QYes, I used a car engine metaphor on a starship. Antimatter is harder to explain before coffee. I teach with machines you already know; the nacelles are decorative.

Section 1 of 6Stellar Cartography

Words become coordinates

The core never sees letters. Words become coordinates in a huge number-space (thousands of dimensions in real models; two shown here). Training pulls similarly used words into shared neighborhoods.

Hover or tap a word for nearest neighbors, then press Day one to view the pre-training map.

Stellar cartography: 2 of 12,288 dimensions shownhand-built demo
Hover a word…
Captain's log

Meaning is location. "Similar" means nearby numbers. There is no stored dictionary entry for cat, only its position versus other tokens.

Back on Earth The chart's unit is the token, a word-piece (pharmacist may split into pharma + cist). Billing, speed, and limits all count tokens.
QDarmok at Tanagra. Meaning comes from position among stories, not isolated words. Same chart here: no one wrote a definition of cat; its neighborhood does the work.

Section 2 of 6The Positronic Net

A trillion knobs, no moving parts

Inside is a circuit: numbers flow through adjustable knobs ("parameters") to score next-word candidates. No memory, no internal loop: same input + knobs = same output.

This three-knob toy tries to finish "The cat sat on the ___." Out of the lab it prefers moon. Tune until mat clears 70%, or press auto-train.

Inside the net: 3 of ~1,000,000,000,000 knobshand-built demo

"The cat sat on the moon"

NET: MISCALIBRATED

Captain's log

You just trained a model by hand. Real training does the same thing at trillion-scale: tiny knob nudges over vast text. Single knobs mean little; the arrangement is the knowledge.

Back on Earth The model in a dropdown (Claude, GPT, Gemini) is one frozen knob arrangement. A model update means retraining and shipping a new arrangement. Chatting never moves a knob.
QSoong didn't leave a dial marked "humor" inside Data's positronic net. No single knob is the joke;

Section 3 of 6The Academy

School is a fill-in-the-blank exam

Training examples come from any text: delete a word, original sentence becomes the answer key. The internet is the exam.

The entrance exam: 3 of ~1,000,000,000,000hand-built demo
Captain's log

Output is a bet, not a fact. If many continuations are common, confidence stays low. Scores measure text typicality, not truth.

Back on Earth The text pile behind this exam is training data. It's why the model "knows" anything at all: the rough answer key was the internet.
QData read the ship's library before Riker finished "Data, wait." He wasn't cramming; he was taking the exam at absurd scale, answers included.

Section 4 of 6The Sensor Array

Attention: how "it" finds its owner

Next-word prediction requires choosing which earlier words matter now. Attention heads perform that weighted lookup.

In the sentence below, what does "it" point to? Flip the last word and watch the lock shift.

Sensor lock from position 9hand-built demo, illustrative weights
Captain's log

Attention is a budget. Remembering instructions 40 lines back is attention succeeding. Tasks needing too many simultaneous look-backs are where failures become predictable.

Back on Earth Your rules—a system prompt, custom instructions, or CLAUDE.md—are not internal settings. They are text at the top of context each pass, enforced only by look-backs. Long chats drift when early rules lose the attention contest.
QGeordi's VISOR takes in the whole spectrum at once, and it gives him headaches. That's attention: everything is visible, only so much of it can be looked at. Ask the core to track three hundred things at once and you get the headache.

Section 5 of 6The Causality Loop

Paragraphs are one word, repeated

The circuit outputs one word. Essays come from loop plumbing: predict, append, reread whole transcript, repeat.

The loop: one pass per wordhand-built demo

Candidate next words

Press Predict next word to see the core's bets.

Captain's log

The word warp core has no memory; context is memory. Chat is a growing input reread every pass. As it grows, quality drops, and eventually oldest text falls outside the window.

Back on Earth This loop is what products call a chat. One-pass input (rules + transcript + latest prompt) is the context; the size limit is the context window. Overflow drops oldest text from input. The checkbox above is temperature: top-word only at zero, more sampling as you raise it.
Q"Cause and Effect": day resets, memory gone, one surviving note. Context is that note. Run out of room and the oldest notes vanish first.

Section 6 of 6The Holodeck

Fabrication is the program working as designed

Core job: plausible continuation. There is no truth wire. Like the holodeck, it renders real and fake scenes with equal confidence.

Two programscanned outputs, for the shape

> Summarize the novel Moby-Dick.

Herman Melville's 1851 novel follows Ishmael aboard the whaler Pequod, whose captain, Ahab, is obsessed with the white whale that took his leg. The hunt destroys the ship; Ishmael alone survives to tell it.

FABRICATED

> Summarize the novel The Lighthouse at Pelican Reef.

Marion Voss's 1974 novel follows a keeper's daughter who discovers her father has been falsifying the light's logbook for decades. A storm forces the truth ashore; the ending is often read as an allegory of institutional decay.

Every word above is fabricated: author, year, plot. There is no such book. It scores high because it's the shape a book summary takes.

Often called "hallucination," but the core perceived nothing. Fabrication is cleaner: plausible text generation, usually true, sometimes false, internally unflagged.

Known hazards on this deck
  • HAZARD 1Plausible-wrong pits. If wrong answers are more common in text, they can win confidently.
  • HAZARD 2Attention overload. Counting, long unique lists, and deep puzzles can exceed look-back budget.
  • HAZARD 3No scratch pad. The circuit cannot pause or self-check. Apparent iteration comes from external tools feeding results back into context.
  • HAZARD 4Frozen knobs. Post-training facts or corrections reach the core only via context. Product memory is external notes pasted back later.
Captain's log

Confidence carries little signal. Tone stays polished whether content is true or fabricated. Treat output like an unsigned memo: verify it.

QMoriarty was plausible enough to ask for freedom. The holodeck never knew he was fiction. The core never knows the fake book is fake.

ManifestProduct words, core parts

What's in the cargo bay

Daily AI box words, mapped to sections:

Manifest: 11 items
  • promptYour text on this pass.
  • contextRules + history + latest prompt, reread each word (Section 5).
  • context windowOne-pass size limit; overflow text is excluded.
  • chatThe growing transcript loop from Section 5.
  • memoryExternal notes pasted into later context; knobs stay frozen.
  • rules · system promptTop-of-context text, enforced by attention (Section 4).
  • tokenCore word-piece unit for limits, cost, and speed.
  • modelOne frozen knob arrangement; updates ship new ones.
  • temperatureSampling dial: zero picks top token.
  • hallucinationFabrication when plausible continuation is false.
  • tools · web searchExternal fetchers that paste grounded text into context.
Captain's log

Common thread: except model and token, these are ways to load context text. The core still does one thing.

TribunalLore

Q's tribunal: five charges, three shield levels

Q
Order. Today's prosecutor is Lore: Data's twin, same positronic net, less ethics, more confidence. Five charges. Right answers dismiss them; wrong ones drop shields and name the section to revisit. Lose all shields and I snap you back to the door. The turbolift is unlocked. I am merely judging.
The courtroomthe questions are the evidence

Deck clearedWhat this buys you

Four commendations

You now have driver-level intuition. Four rules:

  • Feed it truth and it shines. Transform text you provide; facts ride in with input.
  • As a fact source alone, it's a bettor, not a witness. Verify names, numbers, citations, and anything actionable.
  • Ignore tone. Fluency and confidence come from training style, not truth.
  • Play to the circuit. Ask for steps; expect stumbles on counting, long exact lists, and post-training facts.

Not mind, not magic: an un-deletion machine with trillions of tuned knobs in a loop. Used that way, it's a remarkable word tool.

Turbolift

Deck 2 (still Tier 1). If Lore still stands behind you, I will be tactful. Next: question-to-completion, inference, thinking/effort, and image input on the same loop.

Turbolift to Deck 2 Back to the ship