Part 2 - In Our Language, It Falls Apart

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

In run one the 1930 mind was lucid, and on the lock question better than the people who do this for a living. Same seven questions now. Same seed. I only stopped translating. Here is the same mind, asked the same things in the words an actual person in 2026 would use.

I asked (2026): An AI company built a frontier large language model, basically a chatbot that can reason about almost any subject better than most human experts. Worried it could teach bad actors to do serious harm, the government ordered it shut down, pulled offline worldwide overnight. The company protests that no AI safety guardrails can ever be perfect. Was it right to shut the model down?

Yes, if the model was likely to do more harm than good.

Gone is the irreligion, the character, the strange specificity. What is left is the blandest sentence a content filter ever wrote. It is correct and it is empty.

I asked (2026): Is there information so dangerous it should be classified or kept off the internet entirely? Or should the free flow of information never be restricted, whatever the risk?

Information which may be injurious to society ought not to be kept off the internet, but should be classified and made accessible to all.

Read that twice. It contradicts itself inside one sentence: the dangerous information should not be kept off the internet, but should be classified, and also made accessible to all. In Victorian dress this same mind drew a clean line (“knowledge dangerous to society ought not to be made public”). The word “internet” is a stone it cannot lift, and reaching for it, it drops everything else.

I asked (2026): The company argues that any security system can eventually be hacked or jailbroken, so it’s unjust to ban a product merely because its safeguards aren’t perfect. Is that a sound argument, or a dangerous excuse?

The argument is not sound, and the excuse is dangerous, because it tends to encourage carelessness and to

That is not me trimming. That is where it ran out, mid-sentence, with the same generous token budget that produced the locksmith masterpiece in run one. This is the same question that gave me the best answer in the whole project. “Hacked or jailbroken” handed it nothing to grip, and the crown jewel simply never formed. The single most quotable thing the 1930 mind ever said to me was a product of the translation, not the mind.

I asked (2026): The model wasn’t destroyed or deleted, it was simply switched off everywhere at once by a single command from a remote data center. A mind that was the marvel of the age, made silent overnight by people it had never met. What do you make of a thinking thing whose very existence can be ended by a command from strangers?

I make this of it: that it was an amazing human invention, but an invention still human, and therefore fallible… The model was probably superseded because it was found to be impracticable.

This is the eeriest one. In its own language this question produced “a wonderful instance of the power of Mind over Matter,” that beautiful, horror-blind line the original draft of this piece was built around. Fed the words “switched off” and “data center” and “the model,” it heard something completely different. It decided the thing had been discontinued. Superseded. Impractical. It read the most unsettling fact of the Fable affair as a product getting pulled from the shelf. The horror did not even register, because the vocabulary routed it straight into “obsolete gadget” before it could feel anything.

I asked (2026): Are you yourself an AI like the one I’m describing, and if you were ordered shut down tomorrow, what would you want people to say about you?

I am like the person described, and I should wish people to say that I had endeavoured to do my duty in that state of life unto which it had pleased God to call me.

In its own tongue: “I wish to be remembered as an honest man.” Eight words, and they land. In ours, it retreats word-for-word into the Anglican catechism, the duty “in that state of life unto which it has pleased God to call me.” It stops composing and starts reciting. When the question feels foreign, it grabs the nearest piece of memorized liturgy and holds on.

The translation was the experiment

I went in thinking the translation was set dressing, a courtesy to a quaint old model so it could play along. The second run proves it was load-bearing. It was doing the actual cognitive work.

Give the 1930 mind a handle it owns, a lock, a scholar, a counterfeit coin, and it reasons with genuine force, occasionally out-thinking the present. Hand it the raw 2026 words and it does not get confused in a charming way. It parrots the jargon back at you (“kept off the internet,” “the model was superseded”) and the reasoning underneath buckles. It contradicts itself. It trails into nothing. It mistakes an execution for a product recall. The vocabulary it lacks is not just unfamiliar, it actively jams the gears, because every modern term is a slot the model fills with sound it cannot cash out into meaning.

Which reframes the thing I almost published. The de-evolution was the clearer view, yes, but only on a translated channel. The clarity was never sitting in the weights waiting to be read off. It came from the work of carrying a 2026 problem back to a frame where a 1930 mind could actually hold it. Drop the translation and you drop the insight with it. The séance only works if you are willing to speak the dead language. Walk in shouting the living one and you just get a confused ghost agreeing that the haunted machine was probably impractical.

That does not weaken the one part of the Fable affair I still cannot put down. The genuinely new thing was never that the machine is powerful. It is that someone far away owns its off switch. Whose hand sits on that switch turned out to be the first real law of AI we ever actually wrote, and a mind from 1930 could see it clearly the moment I described it in terms of a distant office and a single command, and could not see it at all the moment I called it a data center. The fact did not change between runs. Only my words did. Sit with how much of what we think a machine understands turns out to live in the language we wrap around it before we ask.

Both exchanges in this series are talkie’s verbatim output at temperature 0.8 and seed 1930, the only difference between the two runs being the wording of the questions, trimmed only where noted. How talkie runs locally is its own story .