Ada Lovelace, Note G, 1843 · completed by a 2026 language model
“The Analytical Engine has no pretensions whatever to originate anything.”
The black words are Lovelace’s. The blue words were written by a language model given only the black ones. It chose “originate” over every other word in its vocabulary with odds of 99%.
What this shows
The model has absorbed this famous sentence from text it was trained on, and reproduces it with near certainty.
What this does not show
It does not show that the model agrees with Lovelace, understands her, or refutes her. Reciting a claim about originality is not evidence for or against it.
Plate 1 · Next-word odds
Writing is choosing, one word at a time
At each step a model scores every word it knows and one is picked. The bars show its strongest candidates.
Try the same opening with different models. Keep the opening words and temperature fixed, change the model, then ask again. Compare the favourite next word and the alternatives—not just the finished sentence.
Compare three models side by side in Labs →
Browser models are listed by download size; GPU memory needs vary. This changes the site-wide choice. Selecting here unloads the previous browser model; downloading waits until you press Ask the model. Previous results keep their original model label.
What this shows
Familiar texts produce one overwhelming candidate. Open questions spread the odds across many plausible words. “Temperature” is a dial a person sets: low makes the strongest word win, high gives weaker words a real chance.
What this does not show
Odds are not intentions, beliefs, or understanding. Only the strongest candidates are drawn, rescaled to add up to 100%. Different models give different odds for the same words.
Plate 2 · A meaning map
To a model, a word is a place
An embedding model turns each word into 384 numbers. Words used in similar ways in its training text end up with similar numbers. Here those numbers are flattened onto a page. Pick a word to see its three closest neighbours.
Look at “conscience.” Its nearest neighbour is “consciousness.” The model places them together because the words look and are used alike, not because it holds a view on how conscience and awareness relate.
Show the map as a table
| Word | Group | Nearest three (similarity, 0 to 1) |
|---|---|---|
| grace | Theology | mercy (0.57), salvation (0.51), soul (0.44) |
| mercy | Theology | grace (0.57), prayer (0.46), spirit (0.43) |
| sin | Theology | spirit (0.44), worship (0.43), grace (0.43) |
| soul | Theology | spirit (0.68), body (0.58), wisdom (0.52) |
| spirit | Theology | soul (0.68), body (0.49), mind (0.46) |
| prayer | Theology | faith (0.62), worship (0.61), salvation (0.51) |
| faith | Theology | prayer (0.62), salvation (0.56), worship (0.53) |
| salvation | Theology | faith (0.56), resurrection (0.55), prayer (0.51) |
| resurrection | Theology | salvation (0.55), incarnation (0.52), worship (0.43) |
| incarnation | Theology | creator (0.55), resurrection (0.52), consciousness (0.51) |
| creation | Theology | creator (0.74), maker (0.57), consciousness (0.50) |
| creator | Theology | creation (0.74), maker (0.65), incarnation (0.55) |
| conscience | Theology | consciousness (0.62), mind (0.47), soul (0.46) |
| worship | Theology | prayer (0.61), faith (0.53), salvation (0.50) |
| love | Theology | soul (0.47), mind (0.41), body (0.41) |
| mind | Mind | thought (0.60), soul (0.51), conscience (0.47) |
| thought | Mind | mind (0.60), reasoning (0.41), soul (0.39) |
| understanding | Mind | reasoning (0.62), learning (0.61), knowledge (0.57) |
| knowledge | Mind | learning (0.62), understanding (0.57), intelligence (0.57) |
| wisdom | Mind | intelligence (0.58), knowledge (0.56), soul (0.52) |
| consciousness | Mind | conscience (0.62), incarnation (0.51), body (0.50) |
| intelligence | Mind | wisdom (0.58), knowledge (0.57), reasoning (0.50) |
| reasoning | Mind | understanding (0.62), knowledge (0.54), intelligence (0.50) |
| learning | Mind | knowledge (0.62), understanding (0.61), memory (0.49) |
| person | Mind | creator (0.46), creature (0.45), image (0.44) |
| machine | Machine | computer (0.65), engine (0.63), maker (0.55) |
| computer | Machine | machine (0.65), program (0.55), engine (0.51) |
| algorithm | Machine | program (0.47), calculation (0.46), code (0.43) |
| program | Machine | code (0.65), computer (0.55), learning (0.47) |
| data | Machine | knowledge (0.49), memory (0.48), learning (0.46) |
| memory | Machine | learning (0.49), computer (0.49), understanding (0.48) |
| calculation | Machine | algorithm (0.46), code (0.35), understanding (0.34) |
| rule | Machine | understanding (0.41), reasoning (0.39), knowledge (0.38) |
| code | Machine | program (0.65), algorithm (0.43), computer (0.43) |
| engine | Machine | machine (0.63), computer (0.51), creature (0.49) |
| maker | Making & body | creator (0.65), creation (0.57), machine (0.55) |
| craftsman | Making & body | maker (0.51), tool (0.51), engine (0.46) |
| creature | Making & body | engine (0.49), incarnation (0.49), creator (0.49) |
| tool | Making & body | craftsman (0.51), machine (0.50), engine (0.47) |
| body | Making & body | soul (0.58), consciousness (0.50), spirit (0.49) |
| image | Making & body | body (0.49), computer (0.47), understanding (0.45) |
What this shows
Usage in text has a shape. Grace sits near mercy and salvation; machine sits near computer and engine. Some neighbours are surprising, and the surprises are good discussion material.
What this does not show
Nearness means “used alike in text,” not “the same meaning” or “theologically related.” Flattening 384 numbers onto a page distorts distances, so trust the neighbour list over the picture.
Plate 2b · The same map, one more axis
A third number helps, but does not finish the job
The same 384-number vectors, fit to three dimensions instead of two. Turn it and compare it with the flat map above.
Drag or use the arrow keys to rotate; scroll to zoom
Same words and model as the flat map, fit to three axes instead of two. Turning the view changes which words look close; this list does not change.
What this shows
A third axis keeps more of the real distances. Fewer words sit beside a neighbour they are not actually close to, and more of each word’s true nearest three end up nearest in the picture. The colour groups do not pull further apart: some of the flat map’s tidiness came from the flattening, not from the words.
What this does not show
Three numbers out of 384 is still a drastic cut, and this view keeps well under half of how the words spread out. A pair that looks close from one angle can be far apart from another. The neighbour list, computed from all 384 numbers, is still the only exact answer.
Plate 3 · Question the model
Ask, then examine the answer
What this shows
How one model, with these instructions, answers your question today. Ask the same thing twice, or switch models, and compare.
What this does not show
An answer is generated text, not a source. It can be confident and wrong, and it carries no authority. Check anything that matters against the readings.
For discussion
Rules, representation, origination
-
When the model finishes Lovelace’s sentence with 99% odds, is it following rules, representing her idea, or neither? What would change your answer?
-
Turn the temperature up on “The most important thing a pastor does is.” If a surprising word appears, who originated it: the model, the person who set the dial, or the thousands of writers in its training text?
-
The map puts “conscience” next to “consciousness.” What does a pastor mean by conscience that this geometry cannot hold?
Where the captured figures come from
Plate 1 (captured): ornith-1.5-35b, an open-weight model on local hardware. Temperature 0, top 10 candidates, 8 words per opening. Captured 2026-10-04T23:32:23+00:00. Capture id bafyr4ibnayxvtyafn7sngvmagy2wklyrkmpv7tbucs3tjerxrshoavrcca.
Plate 2: sentence-transformers/all-MiniLM-L6-v2, 384 numbers per word. Layout: classical MDS on cosine distance, random_state=7. Captured 2026-10-04T23:33:24+00:00. Capture id bafyr4iawzkbrmxl62gawyqlgpxnbic2xcbvgpafzgfcgn7pn2hfcr37pby.
Plate 2b: same model and words as Plate 2. Layout: classical MDS on cosine distance, 3 components. Captured 2026-10-30T18:44:09+00:00. Capture id bafyr4ihxruakrpkvuqgxffnkw4co3a62ovdcfggda6psjbptqw4wgm7y6m.
A capture id is a content address: change any number in the capture and the id changes.