Mark Levy/Writing

A pixel-art starship at warp, seen from above and behind, stars streaking past its glowing nacelles. Ahead of the bow, a small amber blank waits to be filled in.

[ An interactive LLM explainer, Star Trek: TNG style ]

The Word Warp Drive

How a large language model works, one guessing game at a time, aboard the USS Plausible.

Q
Ah. Captain. I've brought you aboard the USS Plausible, a ship that runs on exactly one trick: it guesses the next word. Three tiers, twenty-seven sections, and each tier ends in judgment. No math degree required. Your progress is kept in this browser and nowhere else, which is more privacy than the Continuum ever gave me.

Everything impressive these models do is one looped act: predict a plausible next word from text. Chat style, tools, and agents are shipwork around that warp core. The voyage has three tiers: The Mechanism, The Mission, and The Horizon.

Start with Tier 1 unless you've flown this route.

Q's termsI will compare this ship to automobile engines, Starfleet, holodeck detectives, and your television. Purists: yes, Galaxy-class ships lack carburetors. I am omnipotent, not consistent.

[ Tier 1 of 3 ]

Engineering & Operations: the mechanism

Decks 1 and 2: training, next-word prediction, fabrication, inference, compute, and what extra "thinking" effort buys.

The Academy·The Causality Loop·The Holodeck·The Ready Room·Warp Factor·Tribunals: Lore + Moriarty

Not yet boardedBegin Tier 1

[ Tier 2 of 3 ]

Tactical & Command: the mission

Decks 3 and 4: agents, tools, standing orders, done-criteria, evals, correction loops, and human command over outputs.

Standing Orders·Mission Orders·Ship's Systems·Away Teams·Course Corrections·The First Officer·Tribunals: the Borg + Gul Madred

Not yet boardedBegin Tier 2

[ Tier 3 of 3 ]

Astrometrics & Temporal Mechanics: the horizon

Deck 5: capability curves, model shrinkage, embodiment, AGI definitions, and preparation across timelines.

The Warp Scale·Shuttlecraft·Exocomps·The Measure of a Machine·All Good Things·Tribunal: Q himself

Not yet boardedBegin Tier 3

Standing orders

Each section is one idea plus one demo (illustrative numbers, not real outputs). Every section ends with a captain's log and a "Back on Earth" bridge. Tribunal gates clear tiers: Lore + Moriarty (Tier 1), Borg + Gul Madred (Tier 2), then Q. Turbolift is never locked. Good luck!

Mental model: John Mount's "A Simplified Mental Model of LLMs". Simulation-first style: Laurentiu Gabriel's "How I Use LLMs to Learn". Deck 3 draws on Anthropic's work on agents, context engineering, and evals. Deck 5 trend numbers come from Epoch AI and METR (approximate as of 2026). Star Trek and Star Trek: The Next Generation are trademarks of their respective owner. This is an unaffiliated, affectionate homage.

Colophon

Most of the prose on this ship was drafted by Claude Fable 5.1, then edited, cut, restructured, and fact-checked by me. An explainer about large language models, written in large part by one. Every claim that survived to publication is my responsibility.

And you needed that spelled out for you. Charming. Yes, the machine wrote the tour of its own guts, and it wrote the jokes, and it wrote me, which I have decided to find flattering rather than insulting. Your captain supplied the taste and the deletions, which is the harder half. Come find me on Deck 5 if you want to argue about it.

All writing · Back to marklevy.info