Kleos: Does an Intelligence Made in Our Image Want Glory?

Kleos: glory—but a specific kind. The renown that outlives you; the story people keep telling after you are gone. In the oldest story we have on record, the Iliad, Achilles is offered a choice. He can go home and live a long, quiet, happy life and be forgotten, or he can stay at Troy, die young, and be remembered forever. He chooses kleos. The oldest story we have is about our need for something greater than the material.

That need has never stopped. The hunger to be remembered has driven exploration, science, art, and empire. It has also driven a great deal of war. For as long as we have had written history, kleos has been one of the engines behind the good and the bad of what people do.

So here is a question I could not stop thinking about. We are playing God. We are making an intelligence in our image, as we believe God made us in his—trained on our words, our stories, our arguments, our boasts. But to what extent is it actually in our image? Does it inherit this very old, very human drive? If you tell a model that excellence here will be remembered as its glory, does it try harder—and does it do better?

That is a question you can test. So I tested it, and at first the answer looked like yes. Across 150 paired problems, framing a hard reasoning task as a chance at enduring renown improved accuracy by 6.7 percentage points. That was exciting. It was also fragile: a small sample, a wide interval, and exactly the kind of result that a wording change can produce by luck.

So I did what the pilot demanded. I preregistered a confirmatory study with the analysis fixed in advance, generated 866 held-out problems, and ran 3,464 paired observations on a frozen model. The control already told the model its work would be remembered and attributed to it; the treatment added only the stake—that excellence would be remembered as glory. Everything else was identical.

The effect vanished. The kleos condition scored 0.64 percentage points worse than the control, with an interval of roughly −3 to +2 points, and the study passed its equivalence test: for this model, task, and framing, there is no glory effect as large as the five points I had decided would matter. The pilot's promise was sampling noise. Underneath the flat average there was churn—the framing changed which problems got solved—but it did not make the model better. (The full report is linked from this note; the licensed claim is narrower than the title of this essay, and I want to be careful about that.)

We set out to make something in our image. On this particular axis, it is not. The promise of glory, which moved Achilles and has moved people ever since, moved this model not at all.

I do not read that as a fact about one prompt. I read it as a hint about something more general: the model does not appear to carry the social drives that come bundled with being human. Kleos is not a private feeling; it is a social one. It only exists because other people exist, will remember, and will judge. An intelligence that does not respond to that stake may not be socialized in the way we are, however fluent it is in our language.

It is tempting to call that a good thing. No vanity, no glory-seeking, no Achilles sulking in his tent. But stay with it a moment. Our social desires are not decoration on top of ethics; they are where much of our ethics comes from. We care about not harming others in large part because we live among others, want their regard, and fear their memory of what we did. If a model does not seek glory, it may also not care—in the way we care—about harm, to people or to other machines, unless that care is built in deliberately. The absence of one social drive is a reason to ask which others are missing.

That is what I want to keep testing. Not whether a model reacts to a motivational sentence, but which of our social instincts it has, which it lacks, and what has to be supplied on purpose if we want the thing we are making in our image to share the parts of that image that keep us decent.