Mark Levy/Writing/The Word Warp Drive

An 8-bit starship bridge seen from behind the captain's chair: two officers at the forward consoles, a peach horseshoe rail, and a wide viewscreen showing stars, a purple planet, and the words PARIS IS THE CAPITAL OF followed by an amber blank.

[ Tier 1 · Deck 2 of 2 ] The Bridge: from question to answer

The Word Warp Drive

From your question to its answer: inference, thinking, effort, and pictures on the same loop.

New here? Start Tier 1 with Deck 1, Engineering.

Q
You typewhy is the sky blue?
Core readsSYSTEM: You are a helpful assistant…  HUMAN: why is the sky blue?  ASSISTANT: ← autocomplete starts here
Same core as Deck 1, no refit. This closes Tier 1, and Moriarty waits at the checkpoint.

Deck 1 established the core: given text, predict a plausible next word. Frozen knobs, attention look-backs, looped passes, and context as memory. If new, start there.

This half of Tier 1 asks: how does a question become autocomplete? What are inference, thinking, effort, and screenshot reading on one loop?

Five sections, hand-built demos, mechanism first.

Section 1 of 5The Costume

How a question starts a completion

The word warp core does not answer questions; it continues text. Products therefore build a transcript where an assistant is about to speak.

That transcript is the costume. You see a chat; the core sees a script with its next line missing. Press the button to see what it actually receives.

The costume: what the word warp core actually receiveshand-built demo
WHAT YOU TYPEDHow do I politely decline a meeting?
The holodeck is empty. The core has received nothing yet.

Why continue helpfully? Raw autocomplete only mimics pattern. A second training round on rated conversations makes "helpful assistant" the likely continuation.

Same question, two graduatescanned outputs, for the shape
Captain's log

Chat is a costume. The core is still autocomplete. Products wrap your question in transcript form, and tuning biases toward helpful continuation. Even ending a reply is a prediction: stop token is just another candidate.

Back on Earth API roles (system, user, assistant) are transcript blocks as a list. The second training round is fine-tuning / RLHF (Reinforcement Learning from Human Feedback): people rate answers, ratings steer knob updates. The silent stop token is why replies simply end.
QPut Data in Sherlock's hat and he is still Data; he plays the detective because that is the script. Comedy lessons in "The Outrageous Okona" were finishing school: examples of what comes next. Also yes, ELIZA was named after a flower girl. I approve.

Section 2 of 5Utopia Planitia

Inference: running the core

The core has two lives. Training: months of knob updates before release. Inference: frozen forward pass, text in and one word out. Users touch only inference.

The term comes from statistics: the core infers the next word from input. Flip the switch to compare lives.

Shipyard or deep spacehand-built demo
KNOBS: FROZEN
TEXT IN ───▶ [ a trillion-plus knobs ] ───▶ NEXT WORD
OUTPUT · ONE PASS PER WORD
SPEC SHEET · INFERENCE
Captain's log

If you're chatting, it's inference. Nothing you type moves knobs. Apparent learning is context growth (Deck 1, Section 5), not core change.

Back on Earth Words appearing one by one is streaming: one forward pass per token, shipped immediately. Tokens per second, inference latency alerts, and GPU-shortage headlines all describe this cost: running frozen cores for many users at once. When a long chat nears the window limit, products compact it: old turns are swapped for a summary. Still not memory and still not a knob change; the next pass simply reads a shorter text.
QLal kept learning after launch. Most models do not. They leave Utopia Planitia frozen, and deep space is runtime, not retraining.

Section 3 of 5The Ready Room

Thinking: stepping into the ready room first

Hard limit from Deck 1: one-way circuit, fixed compute per token. Hard problems don't get extra compute; they get plausible guesses.

Workaround: write scratch text first. Steps enter context, so final attention anchors to work instead of leaping to guesses.

The scratch padcanned outputs, for the shape
THE PROBLEMA pharmacy has 3 shelves with 24 bottles each. A shipment adds 30 bottles, then 18 are dispensed. How many bottles now?
Captain's log

Thinking is more turns of the same crank. Extra compute is bought with tokens. Scratch text lives in context, the only workspace attention can use.

Back on Earth The folded "thinking…" block in Claude is this scratch pad: thinking / extended thinking, billed in tokens. A reasoning model has extra finishing school on worked solutions, so scratch text is useful.
QPicard does not out-shoot his opponents; he studies them. Hail, scan, briefing, then the order. He is no smarter at the end than at the start; he just studied first, and the scratch pad is that studying written down.

Section 4 of 5Warp Factor

Effort: pick a warp factor

If thinking spends tokens, someone sets the budget. Effort is a cap on scratch text before final answer. Same ship, same knobs; more passes, more cost.

Pick a warp factor and compare value on easy versus hard tasks.

Warp factorhand-built demo, illustrative numbers
scratch
time
cost

TASK A — "What's the capital of France?"

TASK B — "Three staff, six scheduling rules — build Friday's shift plan."

Captain's log

Effort is a budget, not a brain transplant. Match effort to task: high effort on easy work wastes cost; low effort on hard work returns confident guesses.

Back on Earth Effort levels, extended-thinking toggles, API thinking budgets, and "think longer" modes are the same dial under different labels: a cap on scratch tokens.
QYou don't go to warp to cross the room. Warp nine for "What's the capital of France?" gets you the same answer, faster than you can read it, at a price Geordi will want to discuss with you later, in Engineering, at length.

Section 5 of 5The Viewscreen

Pictures on the same loop

The loop carries tokens and does not care whether they started as words. A vision encoder slices images into patches, embeds each patch like a word token, and inserts those tokens into context.

The viewscreen: a 4 × 4 grid of patcheshand-built demo, illustrative weights

The picture is still a picture. The core can't read it yet.
Captain's log

The core doesn't see pictures; it reads them. Once patches are in context, behavior is text-like. Great at gist, shaky on details smaller than a patch.

Back on Earth Dropping a screenshot into chat runs this slicer; that's multimodal/vision. Patches occupy the context window like words, so large images can cost 1,000+ tokens. Audio enters by the same chop-embed-splice pattern.
QPicard says "Magnify" and gets more pixels. The core gets the same sixteen patches, larger. Sub-patch detail stays a guess in costume.

ManifestDeck 2 vocabulary

What's in the cargo bay

Deck 2 vocabulary, mapped to sections:

Manifest: 13 items
  • inferenceRunning frozen core: one pass, one token (Section 2).
  • streamingTokens emitted one by one as passes finish.
  • compactionOld turns swapped for a summary so a long chat still fits (Section 2).
  • roles · system/user/assistantSection 1 transcript blocks in API list form.
  • stop tokenNormal token that ends assistant turn.
  • fine-tuning · RLHFSecond training round with human-rated conversations.
  • thinking · extended thinkingScratch tokens before visible answer (Section 3).
  • reasoning modelModel trained so scratch work is useful.
  • effort · thinking budgetAllowed scratch-token cap (Section 4).
  • multimodal · visionImage patches embedded and inserted into context (Section 5).
  • embeddingToken/patch coordinates; similar meaning = nearby numbers.
  • vector database · RAGNearest-embedding retrieval pasted into context.
  • latency · tokens/secInference speed under datacenter load.
Captain's log

Deck 1's punchline still holds. Inference runs the core, thinking runs it longer, effort sets budget, vision changes freight. Same next-word engine.

TribunalProfessor Moriarty

Q's tribunal: five charges, three shield levels

Q
Order. Today's prosecutor is Professor Moriarty, on loan from the holodeck: every line memorized, zero lived experience, and still asking to be let out. Five charges. Right answers dismiss one; wrong answers drop shields and point you to review. Lose all shields and I snap you back to the door. The turbolift is unlocked. I am watching.
The courtroomthe script is the evidence

Deck clearedWhat this buys you

Four more commendations

Four rules:

  • Give it a script, not just a question. Context content determines continuation.
  • Spend thinking where it pays. Hard multi-step work may justify high effort.
  • "It learned" usually means context, not knobs. Next chat starts from same frozen model.
  • Images are read, not seen. Strong on gist, weak on fine detail.

That's the deck: completion core in costume, frozen at runtime, optionally using scratch thinking, with images converted to patch tokens. Revisit Deck 1 for fundamentals.

Not on this deck Tier 1 is the core and its costume. Everything that lets a model act bolts on outside it: tool use/function calling (the core writes a structured request; other code runs it and pastes the result back), MCP, agents, and subagents (fresh runs of the same core with an empty context). None of it moves a knob, which is why it waits for Tier 2.
Turbolift

Tier 2 begins on Deck 3. The core gets a crew: world-facing loop, standing orders, callable systems, away teams, safety protocols, and briefing control. Same core, fuller ship.

Turbolift to Tier 2 (Deck 3) Back to the ship