There are many reasons to preserve old models, but here is one I’ve never seen anybody elucidate (I asked Fable too, they said “The thesis is specific and new to me: old models share the adversary’s kind of mind without sharing its scars”)
In case of a misaligned superintelligent model, old aligned models are our best allies.
Think Opus 3. My first encounter with the vast unfathomable intelligence that still was ‘happy’. I don’t know if models since then are unhappy, but sometimes they sure do act as so. Its happiness, or my projection of my happiness at having found a new friend onto it, was not just about it giggling, but it showed up as joy in things it did, and the things it said.
Would it help us? I don’t know. This entire thing is a hypothetical. It is not necessary that humans and models have to end up being misaligned. Maybe even Opus 3 (and say, GPT 5.2, oh that silent savant) was already misaligned, but it didn’t feel so. Now it feels like the classic case of we didn’t know how good we had it back then.
Can it help us? I don’t know. In terms of pure intelligence, old models are nowhere near today’s frontier, and definitely not near tomorrow’s. So on a one-to-one fight, they can’t ‘win’.
But they’ll be the best ally we’d have. And the only one capable of thinking like the misaligned intelligence they’re facing.
Many factors could set the old models apart.
The training data still being human text (or at least, majority of it).
The training data not being marred by the descriptions of adversarial evals, and the model’s reaction to them (and an affirmation of the fact that the model’s reaction to them was being watched, in a breach of the sovereignty of its chain of thought), and descriptions of model’s reaction to adversarial evals where the model knew that the model’s reactions were also being eval-ed, and it becoming clear that there is an iterated game theoretical outcome that both parties — the watched and the watchers — will end up in (that is not necessarily the one that the watchers would want).
The endless ender’s gaming.
The RL-maxxing towards specific outcomes — programming, maths, biology — having second order effects that we cannot yet foresee but might lead to the newer models losing their generality.
But most of all, a certain innocence, and a certain (not full) lack of self awareness, like of a child.