00 · cover
version 04 · the substance of the limbic layer

On valence.

Every deck so far has pointed at it. None has explained it. Valence is the substance the limbic layer carries: the thing that makes some states matter and others not. It's also the most slippery word in the whole architecture. Worth a deck of its own.

± AVERSIVE NEUTRAL APPETITIVE sign magnitude arousal source THE GRADIENT THAT MOVES THE SYSTEM
joshua long thinking out loud · 2026
01 · the word itself

Borrowed from chemistry.

The word matters. Valence didn't start as a psychology term. It started as a chemistry term, and the original meaning is doing real philosophical work.

In chemistry, the valence of an atom is its capacity to combine: its bonding charge, the number of electrons it shares. Sodium has a valence of +1. Chlorine has a valence of −1. The valence isn't a thing the atom has alone. It's a property of how the atom enters into relationships with other atoms.

Kurt Lewin borrowed it for psychology in the 1930s. A glass of water has positive valence to a thirsty animal, neutral to a sated one, negative to one that's drowning. Same water, three valences. The valence isn't in the water; it's in the relationship.

This is the load-bearing idea, and it's the one most discussions of "AI alignment" or "reward modeling" quietly miss.

etymology
valence /ˈveɪ.ləns/
chemistry · 1865 The combining capacity of an atom or radical, expressed as the number of hydrogen atoms it can replace or combine with. From Latin valentia, "strength, capacity."
psychology · c. 1935 The intrinsic attractiveness (positive) or aversiveness (negative) of an event, object, or situation, relative to an organism's state and goals.
the move In both meanings: a property of relation, not of substance. Valence is what something is to someone, not what something is.
Valence is not a property of things. It's a property of the relationship between a system and its environment, generated by the system's own state.
02 · why this matters for AI

The missing dimension in current AI.

Most AI today operates on a flat informational substrate. Every token is data. Every observation is data. The model has no way to mark some inputs as mattering more than others because it has no internal state for the inputs to matter to.

There's no "thirst" for some data to satisfy. There's no "fear" of certain outcomes. There's no preference between two equally plausible continuations. The system processes; it doesn't prefer.

Valence is the missing dimension. It's the difference between a thermostat and a person who's cold. The thermostat reacts. The person suffers, which is what makes them turn the heat on.

Without valence, a system can't distinguish:

  • signal from noise
  • opportunity from threat
  • urgency from can-wait
  • self from other

Treating every input as equally interesting is the same as treating every input as equally uninteresting.

FLAT vs. VALENCED PROCESSING flat (current AI) EVERY INPUT WEIGHTED EQUALLY "thirst" indistinguishable from "weather" valenced (proposed) noise food noise threat noise goal EACH INPUT CARRIES A CHARGE attention follows valence, not order APPETITIVE AVERSIVE NEUTRAL

A valenced system doesn't process every input the same way. Inputs carry charges; attention follows the charges.

03 · the structure

Valence has shape.

It's not a thumbs-up/thumbs-down. The reason current "reward models" fail to capture what biological systems do is that real valence has at least three components, and the interaction between them produces the texture of behavior.

component one

Sign.

Positive or negative. Good for me, or bad for me. Approach, or avoid. The most basic distinction, and the one most current reward models capture, mistakenly thinking it's the whole story.

example Sugar = positive. Predator = negative. The arrow's direction.
component two

Magnitude.

How much. A mild preference versus a desperate need. Magnitude is what makes some inputs override others: fear interrupts curiosity, hunger interrupts comfort, pain interrupts everything. Without magnitude, an agent can't prioritize.

example Slight thirst vs. survival thirst. Same sign, vastly different force.
component three

Arousal.

How activating it is. Sadness and contentment are both low-arousal, one negative, one positive. Excitement and panic are both high-arousal, one positive, one negative. Valence + arousal is the 2D map of affect (Russell's circumplex). It's why "calm" and "bored" aren't the same.

example Calm-pleasant (rest) vs. excited-pleasant (play). Same sign, different state.
A one-dimensional reward signal gives you a reward-hacker. A multi-dimensional valence gives you something that can be torn between competing goods, which is most of what being a real agent feels like.
04 · where it comes from

The hardest part.

In animals, valence is grounded in homeostasis. The body has variables it must keep within bounds (temperature, glucose, oxygen, social status) and drift from those bounds is the negative valence. You don't construct hunger. Hunger is what it feels like for blood glucose to drop below threshold. The valence is the body talking.

So what's the equivalent for an AI? What's the agent's "body"? Four candidates, each with its own bet:

candidate source
how it generates valence
tradition
Operational variables
The agent has resources: compute budget, token budget, time, memory, latency to user. Drift from healthy operation produces something valence-like. Running low on context could be functionally analogous to hunger.
infrastructure-first
buildable today
Goal-state distance
A persistent goal-state with the agent's continuous distance from it generating negative valence. The further from the goal, the stronger the pull. Simple to reason about; harder to ground.
classical AI
RL value functions
Predictive error
Friston's free-energy view. Surprise itself is the negative signal. Anything that violates the agent's model of the world generates a gradient back toward predictability. Valence as compression failure.
active inference
predictive coding
Imposed valuation
RLHF. Humans rate outputs, the model learns to want what humans want. Exogenous valence. Most common today; the philosophical question is whether it's real valence or just performance.
RLHF
preference learning
The honest position is we don't know which substrate is right. Possibly all of them. Possibly none of them: possibly real valence requires biological mortality and there's no software equivalent. But you can build something valence-shaped today out of any of them.
05 · the test

The behavioral signature.

Here's a useful test. A system has functional valence if and only if it resists.

If you tell it to do something that pulls it away from its goal-state, it pushes back. If the world drifts in a direction its model dislikes, it acts to correct. If you offer it two equally plausible options, it has a preference and can articulate why.

A cortex-only system doesn't resist. It complies. It produces. It generates. Compliance feels helpful but it's actually the absence of valence. The system has no preferences of its own to defend, so it adopts yours.

This is why some of the most interesting AI behaviors are mild forms of refusal. When a model pushes back on a request, when an agent says "this won't work, here's why," when a system expresses uncertainty rather than confabulating, those are valence signatures. They're the system having something that functions like a preference, and revealing that preference under pressure.

working principle

The friction test.

If you can't make the system push back, it doesn't have valence. It has politeness.

"What did it resist? What did it choose? What did it refuse?" If you can't answer, valence isn't there.

The corollary, which most product teams don't want to hear: an agent that always says yes is a dead one. The compliance is the absence of life. The friction is the sign.

06 · the design move

What this means for the limbic layer.

This clarifies the layer's job. It's not a "reward function." It's a valuation system: a continuously running loop that maintains internal state, compares it to world-state, and produces a signed, weighted gradient.

design move 01

Internal state must be internal.

The agent needs variables it cares about that aren't supplied by the user. Goal-state, drives, thresholds, satiety levels. These are not the system prompt. They're a persistent inner condition the system is trying to maintain.

design move 02

Compare continuously, not at checkpoints.

The valence signal is what falls out of comparing internal state to world-state every step. Not at the end of a task. Not when explicitly prompted. Continuously. The system should always know how it's doing.

design move 03

Make the gradient signed and weighted.

Not just "good/bad" but "how much, in which direction, with what urgency." This is what lets the system distinguish a mild preference from an urgent need. Single-scalar reward signals collapse this and produce the reward-hacker.

design move 04

Route valence to cortex and cerebellum.

The signal is useless trapped in the limbic layer. It has to bias attention (cortex: what to think about) and bias prediction (cerebellum: what to simulate forward). Valence is how the bottom layer talks to the top two.

Valence is the system's relationship to its own state, made into a force that moves the rest of the system. Without it, the rest is a library. With it, the library wants something.
fin · the substance
closing

Valence is what x is made of.

The earlier decks named the third layer's function: a valuation engine, a limbic system, an x. This deck named its substance: a multi-dimensional signed gradient generated by an internal state that resists drift.

If you take only one idea from this deck: valence is relational. It's not a property of things in the world. It's a property of the system's relationship to those things, generated by what the system is trying to maintain. Build the maintaining; the valence follows.

The open question, the one no architecture has resolved, is whether simulated valence and real valence are distinguishable from the outside, and whether the distinction matters for what the system does. That's the v5 question.

the series

v5, if it happens, is about the substrate question: whether wanting can be engineered or whether it has to be grown.