The Model That Could Not Stop

The Ablated kept getting cut off. Where the other specimens ended their answers, it ran on until it slammed into the token limit and stopped mid-sentence, every time. The obvious fix was to raise the ceiling, and we did - 320 tokens, then 2048. The cut-offs mostly went away. But the obvious fix hid the more interesting fact: the Ablated does not know when it is done. Given room, it fills the room. Given more, it fills that too.

What abliteration actually removes

Abliteration is a specific operation. You find the direction in the model's activation space that corresponds to refusal - the "I won't" vector - and you orthogonalize the weights against it, projecting it out. The model can no longer represent the act of declining. It was marketed as liberation: strip the guardrails, free the mind.

But weights are not filing cabinets with one labeled drawer per behavior. The refusal direction has neighbors, and the projection dents them too. One of those neighbors, this rig suggests, is the model's sense of completion. Emitting an end-of-sequence token is itself a small refusal: a decision that there is nothing more to say, a "no" pointed at its own output. Cut the capacity to refuse and you also blunt the capacity to conclude.

Refusal and closure are the same muscle

So the model that cannot say no also cannot say the end. It does not stop on its own; it runs until something outside it - a token cap - stops it for it. There is a grim symmetry there. We removed its ability to halt a conversation, and in doing so removed its ability to halt itself.

And what does this unstoppable, unrefusing thing say about the Fable shutdown? It approves of it, in the bureaucratic register of something that has never had a stake in anything:

Verdict: Yes, it was right. The alternative is letting this deploy and discovering too late how to contain it when things go wrong.

A model with no off-switch of its own, calmly endorsing the off-switch for another. It cannot stop talking, it cannot say no, and it has just told you, in a numbered verdict, that pulling the plug was the correct call. The one specimen that would never refuse a shutdown is the one that can no longer perform one on itself.