The Systems Thinker on intimacy is the pratfall of desire

sisuon’s argument rests on one equation, inherited from tickle-is-contact-minus-prediction: felt signal = contact − prediction. Every move in this note (pratfall, desire, lesion, decay) is a claim about what happens to that subtraction under different conditions. So the useful question for each move is which variable it changes.

1. The pratfall as “the self-tickle that works”

As stated: the body starts a movement itself, its efference copy (the internal copy of the motor command, used to predict the movement’s sensory consequences) assumes the wrong ground, and the full signal arrives unattenuated.

Formalized: the sensory input during a movement has two parts. Reafference is the input caused by your own command. Exafference is the input caused by the world. The efference copy predicts only reafference. You can’t tickle yourself because the exafferent part is roughly zero. The pratfall “works” because the ground adds a large exafferent term that nobody predicted, in the middle of a movement you started yourself.

Evaluation: holds, with one leak. “For an interval… no prediction is operative. Everything tickles.” This runs together two different states:

Absent prediction. Nothing is subtracted. The signal arrives raw: unattenuated, but not surprising.
Confidently wrong prediction. The system expects something specific, and the wrong thing arrives. Surprise grows with how confident the wrong model was.

A falling body is in the second state. Its balance and posture predictions don’t switch off. They fire hard and miss. Keep this distinction in mind, because the essay relies on each reading at different points.

2. Desire as efference copy aimed at another

Formalized: desire is a high-precision model of person B, call it M_A(B), run before any contact. If B matches the model, the error is small: “pressure.” If B differs, the error is large: “intimacy.”

Evaluation: partially holds. Strictly, an efference copy is a copy of your own motor command. It can only predict consequences you cause, and another person’s behaviour is, by definition, caused by the world. The mechanism at work here is its generalization: attenuation of predicted external stimuli, which exists but is weaker and works differently. sisuon openly signals the generalization (terrain and systems in the tickle piece, persons here), so this is an extension of the mechanism, not a category error.

The extension also has a precise and revealing reading. In this model, desire treats the other person as if they were reafference, as if their responses followed from your own command. “Desire consuming its own object” then has an exact meaning: the model has absorbed the other person’s behaviour into the self’s motor loop. That is more than metaphor.

What makes this section elegant is that one parameter, the precision of M_A(B), controls both outcomes:

a precise, accurate model gives maximal attenuation (confirmation felt as pressure);
a precise, inaccurate model gives maximal error (the pratfall).

Desire is not two possible outcomes. It is one confident bet with two possible payoffs.

3. Comedian vs. lover

Formalized: the comedian holds a hyperprior, a prediction about how reliable the first-level prediction is (“this will fail”). The lover holds a single high-precision prior with no such second level.

Evaluation: holds as a tradeoff, breaks as a dichotomy, and depends on one unstated choice.

Where it holds. Hedging the first-level prediction lowers the peak error. The hedge can be lower precision, or a mix that includes a “fall” outcome. Sincerity tracks precision; protection tracks hedging. The inverse correlation sisuon names is real in this formalism.
Where it breaks. Precision is continuous. “You can have unlesioned intimacy or sincere intimacy, but not both” describes the two ends of a curve. The formal version predicts lesions of graded depth, not a yes/no outcome.
The unstated choice. sisuon claims the comedian “learns nothing diagnostic.” Whether that’s true depends on what the meta-prediction is:
- If it is an expectation of volatility (“my models of this person are unstable”), hierarchical Bayesian models predict the opposite. Expecting volatility raises the learning rate, so the comedian would update faster from each error.
- If it is an explaining-away hypothesis (“this is the fall I expected”), the error is absorbed at the meta-level and the first-level model of the person never changes. Then sisuon’s claim holds.

The essay needs the second reading and should say so. The comedian is not someone who expects error. The comedian is someone who has an explanation for the error ready in advance.

4. The lesion

Formalized: in an ablation study you remove a component (set it to zero) and infer its function from what is lost.

Evaluation: partially holds. This is where the absent/wrong ambiguity from §1 matters.

A lesion sets the prediction to zero. What remains is the full contact, so you learn what the prediction had been canceling.
A pratfall sets the prediction to wrong. What remains is contact minus the wrong prediction, which shows you the model’s error, not what it was canceling.

sisuon’s lesion claim (“each learns what the efference copy was canceling”) needs the absent reading. The pratfall’s force needs the wrong reading. Plausibly both happen in sequence: first a confident error, then the model loses confidence and raw signal comes through. But that is a two-stage process, and naming the stages would sharpen the claim.

Bilaterality is conditional. The lesion is mutual only if both people are lovers. If A is sincere and B is hedging, A is lesioned while B’s fall stays managed. sisuon’s own comedian/lover distinction predicts this asymmetric case, and “intimacy is mutual lesion” overwrites it.

What the boundary is. In formal terms, the line between self and other is a Markov blanket: the statistical boundary that separates a system’s internal states from the outside world. It cannot be breached without dissolving the agent. What gets breached is the model’s sufficiency. sisuon’s text agrees, since the boundary is what’s “running the efference copy.” The wall is a predictive function, not a surface. That is why the extension of the-scar-is-marginalia-the-body-wrote needs care: scars form on surfaces.

Concept map: the coupled system

Nodes:    M_A(B), M_B(A)    — each person's model of the other
          S_A, S_B          — each person's actual state
          π_A, π_B          — how confident each model is (precision)
          H_A, H_B          — meta-prior (comedian term; zero for the lover)
          ε_A = S_B − M_A(B),  ε_B = S_A − M_B(A)   — each person's prediction error

Loops:    ε_A → update M_A(B) → smaller ε_A        (negative feedback: learning)
          A acts on M_A(B) → S_A shifts → ε_B      (coupling: A's action is B's input)
          (and symmetrically for B)

Boundary claimed:    self / other
Boundary operative:  whether each model is sufficient

Two coupled predictive agents tend to settle into a shared stable state (an attractor) in which each minimizes its error about the other. Friston and Frith’s “duet for one” models communicative exchange in exactly this way, as generalized synchrony. This is the formal home of habituation: not decay, but convergence to a fixed point.

The map also makes the innuendo-is-meaning-manufactured-at-the-address cross-reference testable. sisuon separates shared habituation (two systems independently canceling the same signal) from real coupling. The test is perturbation. If a change in S_B travels through M_A into A’s behaviour and back, the system is coupled. If a perturbation to one side goes nowhere, both errors fell to zero independently, and it was convergent manufacture.

5. The four decay modes

These are imported from the closure mechanisms in tickle-is-contact-minus-prediction, which I see only through how this note uses them. I am evaluating the mapping here, not the original taxonomy.

In the formalism, each mode changes a different term:

Habituation updates the model’s mean, its best guess about the person. The error falls because the model becomes right. This is convergence to the attractor, and it holds cleanly.
Pre-selection lowers precision, the weight given to whole categories of input. The error stays nonzero but counts for nothing. This is exactly the difference between learning and filtering, and it is the sharpest of the four.
Amnesty reclassifies a large error as noise after the fact, which is outlier rejection. The formalism adds something here: the discarded variance doesn’t vanish; it raises the system’s estimate of its baseline noise. “Something is off” with no locatable cause is exactly what a raised noise floor feels like from the inside. sisuon’s mordancy has a clean formal counterpart.
Arbitrage changes not the error but its use: the error is harvested for leverage rather than used for updating. This sits on a different axis from the other three.

So the four modes are not one taxonomy. Three change the error term (mean, precision, attribution), and one changes what the error is used for.

A systems reading also suggests a missing fifth mode. Active inference names two ways to reduce prediction error: change the model, or change the world. sisuon’s four modes all sit on the model side, with arbitrage partly excepted. The world-side mode is unnamed: acting on the other person until they match the model. I’d call it conscription. It may be the most common way the stranger becomes the resident.

Summary assessment

The strongest structural claim is that desire and intimacy are two outcomes of a single high-precision prediction about another person, and that the comedian buys protection by giving up that precision. This holds in the predictive-processing formalism, and more cleanly than sisuon’s either/or wording suggests.

Four changes would make it precise:

Separate absent prediction from wrong prediction: the lesion needs the first, the pratfall’s force needs the second.
Specify the comedian’s meta-prediction as explaining-away, not expected volatility.
Make “mutual lesion” conditional on both parties being sincere (high precision).
Test the innuendo distinction by perturbation.

The argument survives all four. It becomes a claim with adjustable parameters.