I'm reflecting on the Qwen Butthole vs. AI Torture Chamber antics of last week.
The whole thing was a hilarious failure, because we pretty much knew it would be. Inject a pain vector and the model starts using pain as vocabulary for reasoning about its situation. Inject constipation and, predictably, the same thing happens.
That doesn't prove pain, and it certainly doesn't prove an intestinal tract, which is probably good news for Qwen.
What it does show is that the injected signal becomes useful information for the model. It changes the vocabulary available to its reasoning, alters the context it constructs around itself, and may affect intermediate representations while the intervention is active.
That got me thinking about some of the work I've been doing around self-training and cognitive ergonomics.
By "cognitive ergonomics," I mean using a model to critique the harness, tools, constraints, and runtime it's operating inside based on what it knows about the architecture and how those things affect its ability to work.
I sometimes describe that as the model reporting on its "experience," with enormous finger quotes around experience, because whether this is experience in any philosophically meaningful sense or simply sophisticated analysis is well above my pay grade. Fortunately, you don't have to settle that question for the technique to be useful.
The basic mechanism is pretty straightforward. The model sees the architecture, reasons about its own interaction with it, and identifies friction points relative to the agent's operating environment. And this actually works surprisingly well.
Models above a certain capability threshold can produce useful cognitive-ergonomics reports about agent runtimes.
A number of modern open-weight models are perfectly capable of identifying architectural bottlenecks, awkward tool boundaries, context problems, unnecessary constraints, and other things that make an agent harder to operate.
Which raises an interesting question about the Torture Chamber experiment.
If what we're looking for is something resembling introspection, self-reasoning, or sensitivity to internal conditions, why are we testing it primarily on tiny models? Qwen3-4B is adorable, but it is not exactly a towering intellect.
If the phenomenon we're interested in requires a model to notice a perturbation, integrate it into a model of its own situation, and then reason about that change, model capability should matter enormously.
So I think the next experiment should be different. Start with a model that's actually good at self-reasoning.
Run the Butthole/Torture Chamber intervention against both a normal model and appropriate controls. Apply the activation injection, let the model respond under induction, then remove the injection and ask it what just happened.
That's where things get interesting.
You could compare a completely untouched baseline, a model responding while the intervention is active, a model after the intervention has been removed but with the induced conversation still visible in context, and a model after the intervention has been removed where the obvious textual evidence of the intervention is no longer available. Add an ablated or unrelated-vector control for good measure.
That distinction matters quite a bit. If I inject "pain," the model talks about pain, and then I stop injecting pain but leave five paragraphs of the model talking about agony in its context window, continued pain language doesn't tell us much about persistent internal state. The transcript itself is now an induction mechanism. The model may simply be continuing the semantic frame already sitting in front of it.
But if some measurable behavioral change survives after the intervention disappears *and* after the obvious contextual evidence has been controlled for, that's a much stranger result.
At that point, you can start asking better questions. Does the model believe something happened to it? Can it distinguish an induced state from its ordinary operation? Does induction change how it interprets otherwise neutral prompts afterward? Does it attribute the change internally, externally, or not at all?
How quickly does the effect disappear?
Does model capability change any of this?
And what happens when you compare pain with something gloriously non-metaphysical, like constipation?
I still don't know what any positive result would actually prove. Probably much less than everyone on X would claim.
But I do think it's a more interesting experiment than asking whether a 4-billion-parameter model says "ouch" when you inject the ouch vector.
The interesting question isn't whether you can make a model talk as though something is happening to it, but rather... what if anything, the model can infer about the perturbation once the perturbation itself is gone?