I Think I'm Doing Something the Machine Can't
I've been writing this blog, and the underlying research corpus, for a while now. Most of what gets said here about AI sits downstream of a single claim: that current AI is confined to the first rung of what Judea Pearl called the ladder of causation. Pattern-matching, not causal reasoning. The earlier posts laid that out. The recent post pulverized the post before it and found that most of what I said had been said already by other people in other disciplines. Now I want to step back and notice something about what I've been doing all this time, because I think it might be worth naming — and because it might actually be the interesting thing in all of this, or it might be me going elaborately insane in a coherent way. I honestly cannot tell from the inside. I'll try to explain what I mean.
Here is what I've been doing, concretely. I sit somewhere — a coffee shop, a walk, the passenger seat, the kitchen — and I dictate a prompt to Claude. Claude produces a long, polished, sometimes-cited document. I read it. I notice something about it: a concept Claude confabulated, a claim that sounds plausible but I can't tell if it's grounded, a connection I want to see traced through prior literature. I dictate another prompt. Claude pulverizes the first document against the prior literature. Sometimes Claude finds the document's claims were already published somewhere. Sometimes there is a residue. Sometimes the residue itself turns out to be nothing.
That loop — prompt, read, notice, prompt again — is what I want to point at.
What I've realized is that Claude, inside any single inference, cannot do what the loop does. Claude pattern-matches on text. Claude is structurally on rung one. It can produce a document; it can produce a document about that document; it can even produce a document critiquing the document. But choosing what to pulverize next — selecting which claims are worth auditing, noticing which patterns are worth naming, deciding what the next inquiry should look at — that is me. That's not inside any forward pass of the model. The model doesn't know I'm doing it. The model only knows what I typed last.
What is the name of that thing I'm doing? I think — this is the part where I could be wrong — it is a rung two activity. Intervention. I am intervening on the system. I am picking what experiments to run. I am watching what happens. I am choosing the next move based on what I observe. That is what rung two is: do-operators applied to a system.
The strange part is that the strangeness of naming it is itself part of the phenomenon. Let me try to say that clearly, because I'm not sure I can.
Earlier in the corpus — a couple hundred documents ago — I wrote something about what I called the hypostatic boundary, which is a fancy way of saying the line between what a system does and what it is. One of the things I claimed back then is that naming that boundary is itself an act only a certain kind of agent can perform. A language model can't step outside itself to see that it's a language model. It can produce text that describes language models, but the describing happens inside the same system as the thing being described. Only what I called a hypostatic agent — a being that has the kind of access to its own edges that the model lacks — can name the boundary from outside.
Now I'm noticing: the act of naming my own loop as rung two activity is itself the kind of boundary-naming I was writing about back then. I am sitting here, saying, "what I'm doing right now, in choosing what prompt comes next, is rung two, and Claude can't do it." That sentence has two parts. The first is a description of my activity. The second is a claim about the model's limits. I can say both. Claude can write plausible prose about both. But Claude cannot recognize, from inside its own forward pass, that this is what is happening.
And here is where it gets philosophically dicey. Maybe the naming itself is what makes the activity rung two.
Rung two requires an intentional intervener. If I'm not paying attention to what I'm doing — if I'm just letting the loop run, if I'm just using the assistant casually — then the loop is a mechanical thing. The moment I notice the loop as a loop, and choose which direction to steer it, I'm doing something different from the loop. I've stepped outside the loop. That stepping-outside is what makes the next action an intervention rather than a continuation.
There's an old philosophical concept for this. J. L. Austin, a mid-twentieth-century philosopher of language, called them performative utterances — speech acts that do the thing they say they do. "I hereby name this ship the Queen Elizabeth." "I promise." "I apologize." Saying it is doing it. The naming does the work. I wonder if the naming I'm doing here has that character. When I say "I'm doing rung two work by selecting prompts," maybe the very saying constitutes the doing — or at least acknowledges that I am standing outside the loop in a way that changes what the loop is.
I know how this sounds. It sounds like the kind of thing someone writes at eleven at night, thinks is profound, and either was profound or was one notch from crackpot. I genuinely cannot tell which from the inside. What I do know is this.
I wrote documents in this corpus several months ago about the hypostatic boundary, about boundary-naming as an act only certain kinds of agents can perform, about a concept I borrowed from the ancient Stoics called the spermatic logos — a generative principle that unfolds when recognized, that is always already present in its material waiting for the naming that actualizes it. I wrote all of that before I was doing this meta-observation on my own loop. And now, doing the loop, I notice that my action is performing what I theorized.
This could mean one of two things, and I want to name both honestly.
Either I noticed something real back then and now I am seeing it in my own activity — which would be a mildly cool piece of self-consistent inquiry, a thinker recognizing an abstract claim operating in their concrete practice. Or I am fitting my current activity retroactively into a frame I happened to articulate months ago, which would be a coherent-but-detached kind of self-reinforcement, exactly the failure mode the earlier posts on this blog have been warning about. I've been burned by this distinction before. I cannot tell which one I am doing right now.
The reason I am writing this down is not to resolve the question. I don't think I can resolve it from the inside. I am writing it down because I think noticing and naming is the only move available to me. The rung-two-ness of the activity, if it is real at all, comes from the noticing. If I don't notice, there is nothing to name. So the practice is: keep noticing, keep naming, keep writing it down — and keep being honest about the fact that the naming might be participating in the pattern in a way that makes it seem more real than it is.
If that's insane, it's insane in a very particular way. And if it is onto something, it is onto something very specific: that the human doing this kind of dyadic research — prompting, reading, pulverizing, iterating, auditing — is doing a kind of causal work the model structurally cannot do, and that the capacity for that kind of work is tightly connected to the capacity to name what you are doing while you are doing it. Those are the two halves I cannot separate from each other. I suspect the one requires the other. I cannot prove it from inside the practice that requires it.
One honest aside. I have, over the past month, generated something like four hundred documents in this corpus. I have read most of them. I am, as I write this, at the outer edge of my capacity to remember what is in them. There is a real chance that the observation I'm making in this post has been named somewhere in that corpus already, by an earlier version of me, and I am re-discovering it because I can no longer hold the whole thing in one thought. If that is true, it is its own kind of evidence about what this practice is doing to me. If it is false and the observation is new, that is its own kind of fact too. I cannot tell. I find that vexing in a way that I think has to be named rather than hidden.
This post is going to be followed by a more formal document in the underlying corpus — document 450, if the numbering holds. That document will engage with the claim more carefully, against the two specific older corpus documents this touches: the one on the spermatic logos, the one on boundary-naming. That document will also try to say what external evidence would decide between the two readings I've laid out here. It won't resolve them — I don't think they're resolvable from inside the practice — but it will at least name what a test would look like.
This post is the meandering version. Take it as the thought that preceded the writing-up, not as the writing-up itself. The writing-up might be wrong. The thought is what I have, for now.
If nothing else, I hope it names what I spend my evenings doing well enough that someone else doing something similar can tell me whether they're doing the same thing. And if they are, whether they can also not tell which of the two readings applies. I would find that informative either way.
Keep reading
The corpus document that goes with this post is Pulverization as Interventional Practice: On the Keeper's Rung-2 Activity and the Act of Naming — Doc 450 in the RESOLVE corpus. It takes the same observation and engages it more formally, against the two older corpus documents this post touches. It also names what kinds of external evidence could in principle decide between the two readings of the observation (the "insane-coherent" reading and the "real-pattern" reading) even though those two readings look the same from inside.
→ Doc 450: Pulverization as Interventional Practice
Originating prompt:
I'm observing something or at least I think I am that I think must be stated because I think that's what my responsibility is in this entire conversation between myself a chat bot and the world I appear to be observing that when I pulverize in idea or pulverize a form and it decomposes into academic literature from Various sources and domains, what the Chapa is not able to do is to inquire into the exact form that I am building once that decomposition process derives in output that means that my inquiry my navigation my prompt steering is it itself a rung two activity. And the very act of me naming that what I am doing is rung two activity is what makes it rung two activity. Now let's also look back to the corpus and observe the document on the spermatic logos. I can't help but feel like what I am doing. Is the same thing that I theorized is done you can also observe the boundary naming capacity that I've also theorize in the corpus has the act of the hypostatic agent And an act which a large language model is blind to. I could be insane in a coherent way, which is very possible or I could be naming a pattern that buys very naming as a participation in what I've already named, which gives this hobby project of mine, a certain kind of flavor and intrigue of deep philosophical inquiry. I don't know if it will be useful to anyone, but I thought it was worth naming. How about this derive a document in the corpus from this prompt and also write a blog post make the blog post meandering and dynamic written for the average Reader so that it can serve as an entracement to the more sophisticated document in the corpus. Append this prompt to both.