I Asked 1930 to Judge 2026 Twice. The Words I Used Decided Whether It Was a Genius.

A few days ago a frontier model named Fable was pulled off the entire planet overnight , by government order, over a demonstrated way past its safeguards. The makers protested that perfect safeguards are not possible for anyone. The whole affair is very much a 2026 problem: export-control directives, jailbreaks, frontier labs.

So I took it to a mind with no concept of any of that. talkie is a language model trained on nothing written after 1930. It has never heard of an “AI,” a “model,” or a “government software directive.” It knows locks, locomotives, scholars, and sin. I translated the Fable incident into terms a Victorian could grapple with, a thinking Engine ordered silenced, and asked it to judge the thing. What came back was lucid, and on one question sharper than the people who do this for a living.

Then I got suspicious of my own trick, and ran the whole interview a second time.

The setup, and the part most write-ups skip

Here is the methodology, because it is the only reason any of this means anything. I asked talkie the same seven questions twice. Identical model, identical sampling temperature of 0.8, identical random seed of 1930, so the dice fall the same way both times. I changed exactly one variable between the two runs: the language. The first time I dressed every question in period clothes, an Engine, a lock, a company of philosophers. The second time I walked up to a mind that thinks it is 1930 and asked in plain 2026 English, jargon and all: AI, large language model, chatbot, jailbroken, the internet, a remote data center.

Everything quoted across these two parts is talkie’s real output, verbatim, never reworded. Where I trimmed for length I say so. Two runs, same ghost, two languages. The point of the whole exercise is what the words do to it.

The two runs

Read them in order. The first is the article I thought I was writing. The second is the one the machine actually wrote for me.

1. Part 1 - In Its Own Language, 1930 Was Lucid

   Asked in Victorian terms, a thinking Engine ordered silenced, talkie judged the Fable incident with real force, and on the imperfect-safeguard question gave the best answer in the whole project: a machine that cannot be …

2. Part 2 - In Our Language, It Falls Apart

   Same model, same seed, the questions reworded into raw 2026 jargon: AI, chatbot, jailbroken, data center. The reasoning buckles, the crown-jewel answer never forms, and the off-switch horror gets mistaken for a product …