A man sits across from me and tells me he is fine. He grins wanly as he says so—hands shaking just enough so he tucks them beneath his thighs—bearing a broad, would-be-generous smile not yet reaching adrenalized eyes. While his forearms vibrate with a hardly-concealed rage of which his working model on love and relationality has no understanding, “I’m fine,” he says again, and is not intentionally lying. That’s what’s remarkable. He genuinely believes he’s fine. The shaking is just something that happens to him, the way weather happens, the way the rain falls—nothing to do with him, with the self he has so carefully assembled.

I have been thinking about this man for weeks. Not because his situation is unusual—it isn’t; I see some version of him almost every day—but because I have been diving into the cold math of Information Theory, and I have come to believe there is a precise name for what is happening in his body. The name is Cross-Entropy. And it might be the best way to understand why the “truth” actually sets us free—not just morally, but metabolically.

Maybe the most expensive thing any of us will ever own is our own self-deception.

The Model and the Maelstrom

The mind is a predictive engine. We have P—the “Ground Truth” of our actual experience (the anger we feel, the grief we’ve ignored, the pulse of our real desires)—and we have Q—our Ego-Concept. Q is the model we use to encode who we think we are.

When Q misrepresents P, we pay an Informational Tax. In Machine Learning, Cross-Entropy H(P, Q) measures the total bits required to explain a reality using a model. If the model is wrong, the bit-rate explodes. You are essentially trying to compress the data of a complex, hurting human into the file format of someone who is “perfectly fine.”

This is the state we call Ego-Dystonic. It is the internal friction of a person whose body knows something their story hasn’t caught up with yet. The man in my office has assigned near-zero probability to his own anger. There is no room for it in his Q. So every time that anger surfaces in P—every time his nervous system fires and his hands begin to shake—it generates massive “Surprise Energy,” a torrent of information his model can’t encode. His ventral vagal system, the part of him built for calm social engagement, is working overtime to hold the smile in place while his sympathetic nervous system screams underneath it. Two systems pulling in opposite directions. The autonomic strain of a body trying to override a reality it refuses to name.

Anxiety, in this light, isn’t just a “bug” in the system; it is the heat generated by an inefficient internal model. You are paying for your denial in real-time, metabolic bits. And the bill comes due in the body—the tight chest, the shallow breath, the low-grade hum of dread that never quite resolves because its source has been assigned a probability of zero.

The Ego-Syntonic Trap

But there is a darker side to this math.

We often wonder why certain people never seem to change. I don’t mean the people who are stuck—most of us are stuck in one way or another. I mean the ones who aren’t suffering. The ones who sit across from you in session, or across from you at dinner, and describe behavior that has devastated everyone around them with the calm fluency of someone reading a weather report. No tremor in the hands. No gap between the story and the body. They are telling you exactly who they are.

This is the Ego-Syntonic state. If a person’s model (Q) says “I am someone who takes what I want,” and they go out and take it (P), their model and their reality are perfectly aligned. Their Cross-Entropy is low. They aren’t paying a “Tax of Denial” because they have accepted their own toxicity as a feature, not a bug.

This is the Syntonic Trap: because there is no friction inside them, there is no gradient to force an update. In therapy, these are the hardest presentations to move—not because the person is resisting, but because there is no internal “loss” to minimize. The prediction errors that would normally sting, the social ruptures that would normally teach, register as noise. The system sees no reason to learn. Change requires friction, and their system is running cool.

I’ll be honest: there is a particular kind of grief that lives in this realization. You sit with someone whose model is perfectly calibrated to their cruelty, and you understand—with the clarity of the math—exactly why the people who love them can’t reach them. The world around them is on fire, and their internal thermostat reads comfortable. You can name this. You can’t always fix it.

The Energy of Transition

So why don’t the rest of us—the ones who are suffering, who do feel the friction—simply update our models? If the math is so clear, if the direction of the gradient is so obvious, why do we stay?

This is where Network Control Theory comes in. The brain exists in “states,” and moving from one state to another requires Control Energy. The more entrenched the state, the more energy it costs to leave.

Think about the person who built an entire life around the “Good Child” model. The one who learned, very early, that anger was not safe—that the enraged self had to be walled off, sealed in a room the rest of the system pretended didn’t exist. That wall isn’t flimsy. It is load-bearing. It holds up relationships, careers, a sense of identity. The body has spent years routing around the locked room—tensing the jaw instead of raising the voice, converting rage into migraines, building an elaborate, exhausting network of detours so that no signal ever has to pass through the forbidden territory.

Moving between the “Good Child” state and the “Enraged Self” state is energetically expensive. Not because the anger is dangerous—though it might feel that way—but because the entire architecture of the self has been organized around its absence. To let the anger in is to let the walls come down. And the system doesn’t know yet whether what remains will be a house or a ruin.

So we stay. We stay in a state of high Cross-Entropy—chronic, low-level anxiety, the hum of a model that doesn’t fit—because it feels safer than paying the one-time cost of a massive model update.

We would rather pay the daily tax of a lie than endure the “system crash” of a truth.

Therapy as Gradient Descent

The learning process of therapy is essentially Model Updating. We are allowing our self-concept (Q) to approximate our actual experience (P). Every moment of honest contact is a mini-gradient step.

There is a specific second I’m thinking of. It happens differently each time but it is always recognizable. A client has been talking—fluently, competently, narrating their life at a safe distance—and then something catches. Their voice drops half a register. The room gets very still. They say something like, “I think I am actually very hurt,” and you can hear in the way they say it that they have never said it before. That they are hearing it for the first time along with you. The air in the room changes. Something that was held tightly has just been set down.

That is a gradient step. A single, irreversible update.

We can represent this “Model Update” using the logic of optimization:

q_new ← q_old − η∇H(p, q)

In this frame:

  • q_new: The more integrated self-concept we are moving toward.
  • ∇H(p, q): The Gradient of Truth. The direction of maximum “loss” reduction—pointing exactly toward the feelings we’ve been avoiding.
  • η (Eta): The Learning Rate.

In clinical terms, η represents the delicate boundary of the Window of Tolerance. If the learning rate is too high—if too much truth arrives too fast—we overshoot the zone where integration is possible and land in trauma. The system floods. And flooding doesn’t feel like insight; it feels like drowning. It is the sensation of every defense dissolving at once, the ground giving way, the self losing its edges. The ego can’t metabolize the update because the cost of the “surprise” exceeds the system’s capacity to compute it. So the system crashes—into dissociation, into shutdown, into the numb blankness of hypo-arousal. The update fails. Not because the truth was wrong, but because it arrived without scaffolding.

The art of therapy is finding a learning rate that is fast enough to initiate change but slow enough to stay within the window where the ego can still process the update—where the truth can be felt without the self coming apart.

Truth with Compassion: The Regularizer

But what makes truth survivable?

I’ve started calling the goal of this process a Minimal-Entropy Configuration: Truth with Compassion. And I mean “compassion” here in a very specific, almost mechanical sense.

In Machine Learning, we use Regularization to keep a model from becoming too rigid or “overfitting” to its own errors. The regularizer says: you can update, but you don’t have to shatter to do it. You can change shape without losing structural integrity. In the nervous system, Compassion is the Regularizer. It makes the gradient update safe.

Here’s what I mean. I have watched the same truth destroy one person and liberate another. The difference is almost never the content of the truth itself. The difference is the container. A truth delivered without compassion—by a cruel parent, by an indifferent system, by the person’s own merciless inner critic—lands as confirmation that the self is broken. The model doesn’t update; it fragments. The shame is too great. The system overfits to the wound: I am the kind of person this terrible thing is true about.

But a truth held in compassion—spoken in a room where someone is with you, where the nervous system can co-regulate, where the implicit message is this is survivable, and you are not alone in it—that truth integrates. The model updates. Q moves a little closer to P, and the world doesn’t end. The person discovers that they can be hurt, angry, grieving, and still be held. That is regularization. That is what keeps the update from collapsing into shame.

The Relief

Ultimately, therapy reduces the expected information cost of being oneself.

We stop fighting the autonomic friction of a wrong model and move toward the low-cost, low-entropy flow of reality. We learn that truth isn’t just a moral imperative—it’s the only way to stop the energetic drain of being someone we aren’t.

I think about the man with the shaking hands. What I want for him—what I want for everyone I sit with—is the moment the Tax of Denial is finally repealed. Not a dramatic revelation. Not a breakdown. Just the slow, bodily recognition that the model has been updated, and the fight is over, and there is finally enough energy left to live.

You can see it when it happens. The shoulders drop. The breath deepens—not a sigh, but a full breath, the kind the body hasn’t taken in years because it was too busy holding everything in place. The person looks at you and they look tired, but it’s a different tired now. It’s not the bone-weariness of maintaining a lie. It’s the honest fatigue of someone who has just set down something very heavy.

Intimacy, then, is the relief of finally having a model of yourself that actually matches the data. It is the moment you stop paying for the person you aren’t, and discover there is someone underneath who has been waiting—patiently, this whole time—to be seen.