00 · cover
version 02 · follow-up to "what is the brain"

The limbic gap.

V1 asked which architecture is the brain: LLM, world model, or something else. V2 takes the most honest answer and builds on it: all three were partial views of one machine. Here's what that machine looks like assembled.

cortex LLM limbic ?? cerebellum WORLD MODEL A B C A · COGNITION B · VALUATION C · PREDICTION
joshua long thinking out loud · 2026
01 · what v1 left open

The closing line that wouldn't let me go.

Language models are the cortex. World models are the cerebellum. We have not yet built the thing that wants anything. v1, last slide

That last clause is doing a lot of work. What is the thing that wants? The deck didn't say. This deck does.

The honest answer: it's not a third architecture in competition with LLMs and world models. It's a third architectural layer that sits underneath both and gives them a reason to run. Without it, you have brilliant machinery that doesn't care which way the world goes.

WHAT V1 GOT RIGHT · AND MISSED cortex LLM SYMBOLS · LANGUAGE cerebellum world model PHYSICS · SPACE but without this: limbic valence · stakes · gradient WHY ANY OF THIS RUNS the top two layers are spectators until something underneath has skin in the game

Two upper floors of an unfinished building. The structure is real. There's no engine in the basement.

02 · the three layers

Three substrates, one mind.

Three layers that map cleanly between neuroscience and ML. Each does something the others can't. Each is necessary. None is sufficient.

layer
cortex
cerebellum
limbic
substrate
language · symbols
physics · space · motion
valence · goals
timescale
slow · deliberative
fast · predictive
continuous · ambient
job
thinks about
simulates forward
marks matters
ML equivalent
large language model
world model
reward · drive · homeostat
missing from current AI
we have this
arriving fast
not yet built
03 · layer one
CORTEX
layer one · language model

The cortex.

The slow, deliberative, symbolic layer. The part that holds an argument in working memory, manipulates abstractions that have no physical correlate (justice, the GDP, "the year 1987"), and runs the long-horizon planning that defines what we mean by reasoning.

This is what LLMs do shockingly well. They're not perfect cortex, but they're recognizable cortex: pattern-matching across a near-infinite library of human thought, fluent in symbols, capable of moves that look like reasoning even when they're really retrieval.

What it adds: generality across domains, the ability to operate on concepts not directly tied to sensory experience, the capacity to plan in language, and most importantly, composability. The cortex is the layer where the system can talk to itself.
04 · layer two
CEREBELLUM
layer two · world model

The cerebellum.

The fast, predictive, motor-and-spatial layer. The part that runs forward simulations in the millisecond range: where will the ball land, will my hand reach the cup, is that sound coming from behind me. It's the brain's physics engine, doing more computation per second than the cortex but invisible to introspection.

This is what world models do. They predict the next state of the world given an action: pixels, positions, dynamics. Genie 3, Marble, Cosmos. They're not perfect cerebellum but they're recognizable cerebellum: dense, fast, spatial, action-conditioned.

What it adds: grounding. The cortex can talk about a cup falling forever; the cerebellum tells you when it will hit the floor and whether you can catch it. This is where prediction stops being a metaphor and becomes physics.
05 · layer three · the missing one
+/− LIMBIC
layer three · the missing layer

The limbic system.

The part of the brain that doesn't think and doesn't simulate. It cares. It marks some states as good and others as bad. It generates the gradient that makes the rest of the brain do anything at all. In an animal: hunger, fear, lust, pain, social belonging. In a system: whatever makes one possible next state preferable to another.

Crucially, this isn't a "reward model" bolted onto an LLM. It isn't RLHF. Those are external valuations imposed by humans. The limbic system is endogenous. The system generates its own preferences because it has its own state to maintain. Drift outside the bounds → action. That's the loop.

What it adds: a reason to run. Without it, the cortex is a library and the cerebellum is a movie projector. With it, the whole system becomes an agent in the strong sense: a thing whose internal state objects to certain world-states and acts to avoid them.
06 · the synthesis

The assembled machine.

All three layers, running together. The cortex thinks about it. The cerebellum predicts it. The limbic system wants it. Information flows up; valence flows down; action flows out.

cortex · language model symbols · planning · self-talk SLOW · DELIBERATIVE · ABSTRACT cerebellum · world model physics · forward rollout · grounding FAST · PREDICTIVE · SPATIAL limbic · valence engine homeostat · drives · skin in the game CONTINUOUS · AMBIENT · ENDOGENOUS information ↑ state · prediction ↑ ↓ goals · weighting ↓ what to predict body · environment · stakes WHATEVER GROUNDS THE LIMBIC SIGNAL action perception LLM WORLD MDL ??? NOT YET

The crucial move: the limbic layer doesn't just receive. It tells the upper layers what to attend to and why. Without that downward signal, the cortex thinks about everything and nothing in particular.

07 · the open question

Architecture problem, or substrate problem?

If the limbic layer is the missing piece, the next question splits into two camps, and the answer determines whether AGI is engineering or biology.

fork A · architecture

You can build it.

Wanting is just a sufficiently sophisticated optimization target. Stack the right loops (homeostatic variables, intrinsic curiosity, social drives, multi-timescale rewards) and the system behaves indistinguishably from one that cares. The limbic layer is a design problem.

This is the implicit bet of most AI labs. It's also the bet of active inference and free-energy approaches: you formalize "wanting" mathematically and instantiate it in silicon.

If true: AGI is a 5-15 year engineering project. The labs that solve agentic memory + intrinsic motivation + multi-objective optimization will get there first.
fork B · substrate

You have to grow it.

Wanting isn't an algorithm. It's a property of systems that can die. The limbic signal evolved across billions of years of mortality pressure; every loop in the brain was shaped by the threat of dissolution. You can't write that into a model; you can only grow it in a substrate that has skin.

This is the embodied-cognition view, the biology-of-mind view, and (quietly) the view of a lot of neuroscientists who watch AI labs and shake their heads.

If true: AGI isn't a software problem at all. The interesting work is in biological hybrids, embodied robots with real stakes, or AI grown rather than engineered. The current paradigm tops out at "extremely useful tool."
The most honest position: we don't know which fork we're on yet. And we won't know until someone builds a serious limbic layer and sees whether it generates real wanting, or just simulates wanting convincingly enough that the distinction stops mattering.
continue → v3 Designing the three-layer agent The builder's chapter. How this architecture changes what we ship.