Making a visual

FYI: I had Claude create this page for me based on the session history from a long conversation on decision models. The lead/summary for each section is generated. This page basically just exists to illustrate what the process can look like.

This page shows every iteration behind the two interactive visuals in Understanding Jev and System 1 Models, with each round moving from my feedback to AI-built candidates and another review.

1Ten takes on a galaxy

I spoke a starting brief that combined an ambiguous use of "free", backward attention arrows, a galaxy of words that lights up by meaning, a model filling with context and a generated continuation, then AI produced ten interpretations.

v01 canvas paperv01 canvas paper
v02 three.js galaxyv02 three.js galaxy
v03 svg minimalv03 svg minimal
v04 dark terminalv04 dark terminal
v05 editorialv05 editorial
v06 parallax depthv06 parallax depth
v07 light webglv07 light webgl
v08 constellationv08 constellation
v09 dense fieldv09 dense field
v10 step playerv10 step player

2The spin

I asked for boxed input words, a closer and denser galaxy that spins related words above the sentence, and a model box containing a blue weights matrix and wrapping context row.

w01 canvas heirw01 canvas heir
w02 ringw02 ring
w03 sphere swingw03 sphere swing
w04 orbit lanesw04 orbit lanes
w05 cloudw05 cloud
w06 cylinder drumw06 cylinder drum
w07 vortex pullw07 vortex pull
w08 weights herow08 weights hero
w09 hover herow09 hover hero
w10 three.js maxw10 three.js max

3Three experiments on the winner

Each experiment changed one part of the winner: a dense galactic core, active words flying in front of the model, or fixed blue weights lighting up as green word vectors pass through them, with the discussion clarifying what each colour represented.

x01 core galaxyx01 core galaxy
x02 words forwardx02 words forward
x03 weights transformx03 weights transform

4All the notes at once

The next round combined a spherical galaxy behind the diagram, exact step replay on hover, a readable sequence from the selected word through the weights and word space, and a slide-deck scrubber added after the first feedback.

y01 faithful heiry01 faithful heir
y02 causal theatrey02 causal theatre
y03 deep spherey03 deep sphere

5Fix the hole

An annotated screenshot showed the galaxy masked into a ring around the diagram, so this round placed its densest region behind the diagram, tried both semi-opaque and solid boxes, and removed the idle shimmer from the weights.

z01 through boxz01 through box
z02 solid boxz02 solid box

6Transport controls

This round made every button solid, extended hover replay to every word, replaced waves with sparkles in the weights, and exchanged the word-labelled buttons for a transport bar, with the bolder version becoming the first finished visual.

za01 restrainedza01 restrained
za02 bold, the finalza02 bold · the first finished visual

7A second visual: side by side

A voice note proposed putting the same question to a coaxed LLM and a decision model: the LLM types JSON one character per pass through the weights, while the decision model fires once and stamps the complete answer.

sb01 restrainedsb01 restrained
sb02 bolder, the finalsb02 bolder · the second finished visual

8Five polish turns

Focused edits made the inputs directly comparable, paired corresponding data on hover, kept each row to one line with probabilities inside the JSON, let the LLM drift into a red conversational coda, and revised the transport and probability controls. A last pass swapped in plain titles, removed the explanatory text and tightened the spacing, giving the version now embedded in the article.

The current visual, as embedded in the articlecurrent · embedded in the article

The prose in my articles is handwritten, while the visuals come from a collaboration in which I describe the idea, AI builds candidates, and I review and correct them. I explain the wider process in How I Use AI in Writing.