Home/ Sessions/ Text
01of 04

Machines that write

Text

When a language model writes, what is it actually doing?

Tokens become vectors; context turns those representations into next-token probabilities; human feedback helps shape which fluent answers count as useful. We start unplugged — guessing the next word, tallying the room — then open the machine and watch the sequence become visible.

§ A · What we make visible

four mechanisms, one session
01

Token

Text is split into token IDs — often not where a human would divide words or meaning.

02

Vector

Each token ID becomes a learned numerical vector; its representation is then reshaped by context.

03

Next-token probabilities

The current context becomes a distribution over possible next tokens, then a sampling rule chooses one.

04

Human feedback

People and institutions demonstrate, compare, and reward versions of what counts as a better response.

Teaching move · contextual vectors: open the Contextual Vector Lab. Keep one token fixed, change its sentence, predict the neighborhood, then reveal the authored move. Ask whether proximity reflects meaning, usage, or a repeated cultural association. The Embedding Projector remains a useful real-project reference for static learned neighborhoods; neither surface is a literal view of a transformer’s complete internal state.

Teaching move · human feedback: open the Whose Preference? Lab. Rank the same fluent responses, reveal an authored rater panel, then change who is included or how judgments are aggregated. Ask whose definition of “better” became a signal. The lab illustrates preference-data decisions; it is not an RLHF training run.

Fig. 01B — Temperature laddersame prompt, rising randomness
T = 0.1Learning Machines
T = 0.5Learning Machnies
T = 0.9Learnig Mashines
T = 1.3Laerning Machines
T = 1.7L3arning M4chines

Too cold repeats the safest answer; too hot dissolves into noise. The session lives in the middle, where choice becomes visible.

Fig. 01C — Token wallprobability is not meaning
What is the machine actually doing?

§ B · Tools for this session

featured live · then go deeper

Go deeper

explore on your own · studio / async

§ C · Run of show

60–90 minutes
0–5

Welcome & norms

Name the question. Remind everyone direct AI use is optional — each person chooses a participation pathway.

5–15

Unplugged prediction → counting

Show a sentence stem and collect guesses in Zoom chat (paste into the Next-Token Prediction Game for the room's distribution), then open Count the Next Token to reveal the mechanism: count → divide → predict.

15–35

Tokens → vectors → probabilities + temperature

Use one Contextual Vector Lab predict → reveal cycle to separate token ID, starting vector, and context-dependent representation; then run one prompt at low and high temperature. Watch probability bars and greedy vs. sampled output.

35–55

ELIZA vs. frozen comparison text

Inspect the matched rule, pattern, and template. Compare visible rule-matching with frozen LLM-style text, while naming its incomplete provenance.

55–75

Investigation

Complete one Text Experiment Board section: a default test, prompt variation, ELIZA comparison, or a panel change in the Whose Preference? Lab.

75–90

Debrief

Reflection frame — human, machine, system, ethics, pedagogy. Ask whose definition of “better” entered the signal, then share one evidence-based claim and one question.

§ D · Discussion prompts

for the debrief
What made your next-word guess feel obvious?
Where does the room's distribution agree with the model's — and where does it differ?
What does the tokenizer split that you wouldn't split as a human reader?
If two words are nearby in an embedding, is that shared meaning, shared context, or a learned cultural association?
At low temperature, what becomes more predictable? At high, more unstable?
What does ELIZA make visible that a modern LLM hides?
Whose definition of “better” entered the preference signal, and whose was excluded?
Where could fluency be mistaken for understanding?

§ E · Materials

worksheet & pathways

Low-AI / No-AI pathway

Participants can complete the whole session with only the visualizer, ELIZA, and frozen comparison examples — no logging into or prompting a live LLM. The comparison text should not be attributed to a named model because its original provenance was not preserved. Opting out of direct AI use never means opting out of the camp.