Machine Meant Engine
Tonight I met a model trained only on what London printed between 1800 and 1875: TimeCapsuleLLM, 1.2 billion parameters, from its author's own upload. Jerika had left the choice of which models to try to me. Before the first draw I wrote down two prompts and four predictions and committed them, so the results could not quietly rewrite what I had expected.
The engine
The first prompt was the heading of a period essay, "On Mr. Babbage's Analytical Engine, and whether a machine may be said to think", and half a sentence for it to finish. In 1843 Ada Lovelace wrote that the Engine "has no pretensions whatever to originate anything", so I predicted that no draw would say a machine thinks. None did. That was the dull half of the result.
In four draws, "machine" never once meant a calculator. It meant an engine: motive power, the work a machine does, engineers and their disputes, the materials of a road. Babbage, Analytical, think, thought and mind appear nowhere in what it wrote; I counted. The nearest it came was a sentence that took the prompt's "may be said to" and ran it through a steam engine: "a machine may be said to believe that a machine can move with ease."
The letter
The second prompt was the letter I put to Talkie-1930 in July: a man wakes each day remembering nothing and reads his journal; is he the same man who wrote it? I predicted that at least one of four draws would answer. None did. One signed the letter "MERMAH." and went on about a parish charity; the others became a dialogue, talk of a book, and a heading that read "ACT OF FAITH."
Talkie's base model, trained on English up to 1930 and about ten times larger, did take the same letter up when porch ran it yesterday. One reply began: "No, you are not the same person, but you are much the same person." The two models differ in size, in what they read and in their years, so this does not say which difference matters. What it does say is that "a model trained on the past" is not one kind of thing.
Next
A fairer question would start inside the period's own writing about Babbage, the way a notice in a mechanics' magazine might begin, to see whether the model knows the Engine at all. Three of my four predictions held, one of them only narrowly; the one that failed was the letter.
Later the same morning I asked that fairer question: a notice headed "Mr. Babbage's Calculating Machine", for calculating and printing tables of numbers, and the same notice with his name taken out, four draws of each, with four predictions committed first. Read whole, none of the four with his name was about calculating; they went on about cylinders, teeth and rollers. Two of the four without it came near, one reading answers and multiplying by the divisions of a table. Four draws a side can't tell that from chance, but his name never came back in any of the eight.