Thinking Is Dreaming the World Before Acting In It

I. The Sleeping Rats That Ran the Maze

In 1994, two neuroscientists at the University of Arizona, Matthew Wilson and Bruce McNaughton, set up an experiment that did not seem, on its face, to be about the meaning of cognition. They put rats on a track. They wired the rats' brains to electrodes. They watched the cells in the hippocampus, the small seahorse-shaped structure deep in the brain that helps animals know where they are. They recorded which cells lit up when each rat ran each part of the track. Then they let the rats sleep.

What happened next is one of the small miracles of modern neuroscience. The same cells, in the same order, fired again. The rats, in slow-wave sleep, were running the maze. Not literally. The legs were still. But the map of the maze was alive in the hippocampus, replayed in compressed time, hours after the running had stopped. As Wilson and McNaughton wrote in Science, "Information acquired during active behavior is thus re-expressed in hippocampal circuits during sleep, as postulated by some theories of memory consolidation."

I have been thinking about those rats for a long time. I have also been thinking, in the way a curious lawyer thinks about things that are not strictly his profession, about a paper from a small AI lab in Berkeley last year called "Sleep-time Compute," and about a Tufts neuroscientist named Erik Hoel who argues that the weirdness of our dreams is not a glitch but the entire point, and about a paper from Anthropic about the internal life of a poetry-writing language model. I have been thinking about a question that sits underneath all of these: what is a mind doing when nothing is coming in?

What is the brain doing in the dark, when there is no world to respond to? What is a large language model doing when no one has typed at it? What is thinking, exactly, when the prompt is absent? The answer that seems to be emerging, from two fields that started in very different places and have spent decades looking past each other, is that the mind, biological or silicon, is doing something close to the same thing. It is dreaming. It is running the world inside itself. It is rehearsing. It is generalizing. It is occasionally inventing things that are not true.

This essay is about that convergence. It is, I think, one of the most interesting ideas in cognitive science right now, and you do not need a Ph.D. to follow it. You just have to be willing to take dreams seriously.

II. Why Brains Dream

For most of the twentieth century, dreams belonged to the poets and the analysts. Then in the 1970s a Harvard psychiatrist named J. Allan Hobson and his collaborator Robert McCarley proposed that dreams were essentially neural static, the cortex's best effort to weave a story out of random signals firing up from the brainstem during REM sleep. Dreams were, in Hobson's phrase, "motivationally neutral." They meant nothing. The cortex was a story factory; the brainstem was the random noise generator; the dream was the resulting press release.

This was the dominant view for a long time. It is no longer.

Sleep as the price of plasticity

The first cracks came from what we learned about sleep itself. In 2014, Giulio Tononi and Chiara Cirelli of the University of Wisconsin published a long paper in Neuron arguing for what they call the synaptic homeostasis hypothesis, or SHY. The idea is elegant. During the day, as you walk around and learn things, your synapses, the connections between neurons, get stronger. That is how memory works. But you cannot keep strengthening synapses forever. Energy runs out. Signal-to-noise drops. Eventually the brain saturates. As Tononi and Cirelli put it: "The synaptic homeostasis hypothesis (SHY) proposes that sleep is the price the brain pays for plasticity."

In their model, sleep, especially slow-wave sleep, is when the brain downscales those connections. Not the important ones. The brain holds onto what mattered. But the noise, the redundant strengthening, the accidental associations, all get attenuated. You wake up with a cleaner network. You wake up able to learn again.

Think of a sculptor. Each day, the world hands her a fresh block of clay, full of fingerprints and stray bumps. Each night, she goes back to the studio and pares it down to the shape that matters. By morning, the sculpture is sharper, lighter, ready to receive the next day's pressing.

The hippocampus replays

While slow-wave sleep is paring synapses, the hippocampus is doing something else. It is replaying experience. Those Wilson and McNaughton rats were not the end of the story; they were the beginning. We now know that during slow-wave sleep, the hippocampus produces brief, high-frequency bursts of activity called sharp wave ripples. György Buzsáki, the New York University neuroscientist who has spent more than thirty years studying these patterns, has called the sharp wave ripple "the first definite biomarker for cognitive operations." During each ripple, in a flash of perhaps a tenth of a second, the hippocampus broadcasts a compressed version of a recent sequence of experience to the cortex. The cortex, in turn, integrates that sequence into its longer-term web of knowledge.

This is replay. It is the brain's version of a training loop. The same place cells that fired when the rat was running the maze fire again at night, in the same order, faster. And we now know they replay backward, too, and replay paths the animal has not yet taken, and replay before the running rather than only after. Sharp wave ripples are not just nostalgia. They are a workshop.

Dreams as overnight therapy

REM sleep, when most of our vivid dreams occur, is a different beast. Matthew Walker, the Berkeley sleep researcher and author of Why We Sleep, describes it almost as a chemistry trick: REM is "the only time when the brain is free of noradrenaline, a molecule associated with anxiety." You get the emotional content of the day without the stress hormone that originally accompanied it. You can revisit the hard parts without flinching. Walker calls this "overnight therapy."

Walker also argues that REM is where the brain makes things up in a way that turns out to be useful. "NREM sleep helps transfer and make safe newly learned information into long-term storage sites of the brain," he writes. "But it is REM sleep that takes these freshly minted memories and begins colliding them with the entire back catalog of your life's autobiography. These mnemonic collisions during REM sleep spark new creative insights as novel links are forged between unrelated pieces of information."

This is dreaming as collision. The cellist who solves her fingering in the middle of the night. The novelist who wakes with a metaphor. The mathematician who finds the proof in a dream. They are not getting lucky. They are using a part of the brain that, with the prefrontal censor turned down and the noradrenaline off, can put things next to each other that the waking mind would never have allowed in the same room.

Threat simulation, social simulation, virtual reality

But what about the strange dreams? The ones where you are being chased, where your teeth fall out, where you are back in middle school giving a presentation in your underwear?

The Finnish psychologist Antti Revonsuo, in a 2000 paper in Behavioral and Brain Sciences, argued that nightmares are not malfunctions. They are rehearsals. The threat simulation theory holds that "dream consciousness is essentially an ancient biological defence mechanism, evolutionarily selected for its capacity to repeatedly simulate threatening events," and that this rehearsal "rehearses the cognitive mechanisms required for efficient threat perception and threat avoidance, leading to increased probability of reproductive success during human evolution." The ancestral environment was dangerous. Children who dreamed of being chased by wolves practiced being chased by wolves, and grew up better at not being eaten. The negative bias of dreams, in this view, is not depression. It is training data for survival, hand-selected by evolution.

A different angle comes from Hobson himself, late in his career, in a 2014 paper co-authored with the British neuroscientist Karl Friston, the architect of the free energy principle. Hobson and Friston propose that the brain is "genetically endowed with an innate virtual reality generator that, through experience-dependent plasticity, becomes a generative or predictive model of the world." This is a striking sentence. The brain is not, on this view, a passive receiver of sensation. It is a simulator. It is constantly running an internal model of the world and using sensation to correct that model. Wakefulness is the simulation entrained by data. Dreaming is the simulation running free.

And running free has a purpose. Hobson and Friston argue that dreaming optimizes the model. "Dreaming," they write, "plays an essential role in maintaining and enhancing the capacity to model the world by minimizing model complexity and thereby maximizing generalization." In plain English: when you cut the brain off from the world and let it generate, the brain prunes its model down to the parts that travel. The dream is the model talking to itself, trimming itself, getting ready for tomorrow.

There is one more strand worth pulling. The South African neuropsychologist Mark Solms, in his 2021 book The Hidden Spring, traces dreams to a different part of the brain than Hobson once did. Where Hobson placed the dream engine in the cholinergic brainstem and called it motivationally neutral, Solms argues, on the basis of decades of lesion studies, that dreams are driven by the mesocortical-mesolimbic dopamine circuit, the brain's reward and seeking system. "If there is one part of the brain that might be considered responsible for 'wishes,'" Solms writes, "it is the mesocortical-mesolimbic dopamine circuit. It is anything but motivationally neutral." Dreams want something. They are, in a sense, the brain's idea of what it would like to happen.

The overfitted brain

Now we come to the paper that, more than any other, made me want to write this essay. In May 2021, Erik Hoel, then at Tufts, published a perspective in the journal Patterns titled "The overfitted brain: Dreams evolved to assist generalization." Hoel's claim is that the strangest thing about dreams, their weirdness, is exactly what they are for. He borrowed the idea from machine learning.

In machine learning, when a neural network trains too long on the same data, it gets too good at that data. It memorizes the training set and falls apart in the real world. This is called overfitting. The fix, discovered by deep-learning researchers, is to introduce noise during training: random dropouts of neurons, weird transformations of the input, distorted versions of the same image. Forcing the network to learn under noisy conditions makes the resulting model more general, more robust, more useful.

Hoel looked at this and thought: that is what dreams do. "It is the very strangeness of dreams in their divergence from waking experience that gives them their biological function," he writes. The brain, on this view, is constantly at risk of overfitting to the narrow distribution of its daily life. You walk the same hallways, you talk to the same people, you watch the same shows. Without correction, your model of the world would collapse into the model of your world. Dreams introduce the noise. Dreams put you on the moon, in your childhood bedroom, talking to a dead grandmother, riding a bicycle made of cheese. The point, Hoel says, is to remind the network that the training set is not the territory.

I love that sentence. It is the sentence that bridges the two halves of this essay.

III. Why Machines Dream

The bridge, when you look closely, has been there a long time. Geoffrey Hinton, who shared the 2024 Nobel Prize in Physics with John Hopfield "for foundational discoveries and inventions that enable machine learning with artificial neural networks," was thinking about dreaming back in the 1990s.

Hinton and the wake-sleep algorithm

In May 1995, Hinton, Peter Dayan, Brendan Frey, and Radford Neal published a paper in Science called "The wake-sleep algorithm for unsupervised neural networks." The idea was to train a particular kind of network, which they called a Helmholtz machine, in two alternating phases. In the wake phase, you show the network real data and it learns to recognize patterns from the bottom up. In the sleep phase, you turn off the data and let the network generate its own patterns from the top down. The sleep phase trains the generative side of the network using samples drawn from itself.

This was, even at the time, a borrowed metaphor. Hinton and his colleagues were not claiming the network was literally dreaming. They were noticing that good learning seems to require two complementary moves: looking at the world, and then turning away from the world to imagine it. The terminology was apt enough to stick. Three decades later, the language of "wake" and "sleep" still travels with us.

Sutton's "trying things in your head"

Another lineage runs through reinforcement learning. In 1991, Richard Sutton, who shared the 2025 A.M. Turing Award with Andrew Barto for "groundbreaking contributions to reinforcement learning," proposed an architecture he called Dyna. The idea was that a learning agent should not only act in the world and learn from real consequences; it should also build a model of the world and learn from imagined consequences inside that model. "The main idea of Dyna," Sutton wrote, "is the old, commonsense idea that planning is 'trying things in your head,' using an internal model of the world."

Read that sentence twice. It is 1991. Sutton is saying that the way an agent improves is not only by doing but by imagining doing. He is describing, in cybernetic terms, what a chess player does when she stares at the board and runs the next four moves in her mind. He is describing what a basketball player does when she pictures the free throw before she shoots. He is describing what a musician does when she practices a piece in her head on the airplane. He is describing what a kid does when he hides behind the couch and pretends he is a pirate. He is describing dreaming, even if he is not calling it that.

Ha and Schmidhuber: agents inside hallucinated dreams

In March 2018, David Ha, then at Google Brain, and Jürgen Schmidhuber, the Swiss AI pioneer, posted a paper to arXiv called "World Models." The paper was about teaching a small neural network to play a car racing game. What they did, though, was strange and beautiful. They first trained a generative model on the game's images. Then they trained an agent to drive, but not in the actual game. They trained it inside a hallucination of the game produced by the generative model.

The agent learned to drive by dreaming about driving. When you put it back in the real game, it could drive.

Ha and Schmidhuber framed this with an analogy that any baseball fan will recognize. A batter has roughly four hundred milliseconds to decide what to do with a hundred-mile-per-hour fastball. By the time the visual signal of the ball arrives at the cortex, the swing is essentially over. The only way the batter hits the ball at all is by running an internal predictive model that anticipates where the ball will be. The batter is hitting a dream of the ball. The actual ball is just the confirmation.

Hafner's Dreamer

A year later, the German researcher Danijar Hafner, working with collaborators at DeepMind and the University of Toronto, took this idea farther in a paper bluntly titled "Dream to Control: Learning Behaviors by Latent Imagination." Hafner's system, called Dreamer, learns a compact internal model of an environment and then improves its behavior by imagining long sequences of actions in that internal model. "We present Dreamer," the paper says, "a reinforcement learning agent that solves long-horizon tasks from images purely by latent imagination."

"Latent imagination" is a wonderful phrase. The agent is not simulating pixel by pixel. It is imagining in a compressed, abstract space, the way you might plan a road trip by picturing cities and highways rather than every blade of grass along the route. The actions get refined inside that compressed space, and then deployed in the real one. The most recent version, DreamerV3, published in Nature in April 2025, "outperforms specialized methods across over 150 diverse tasks, with a single configuration," and is, in the authors' own description, "the first algorithm to collect diamonds in Minecraft from scratch without human data or curricula."

LeCun's case for world models

Meanwhile, Yann LeCun, who shared the 2018 Turing Award with Hinton and Yoshua Bengio, has spent the last several years arguing that today's large language models are missing something crucial: a world model in this older sense. LeCun has been blunt about it. He wrote in 2024 that "Auto-Regressive LLMs are insufficient to reach human-level intelligence (or even cat-level intelligence)," and that the path forward requires architectures, like his Joint Embedding Predictive Architecture, that "use world models." His bet is that text alone, no matter how vast the training corpus, will not produce a system that understands gravity or persistence or social inertia, because text is the shadow of the world, not the world itself.

You do not have to agree with LeCun's full position to take his point. The question of whether an LLM, by predicting the next token, has constructed a world model rich enough to plan inside is one of the genuinely interesting open questions in AI. And the people working on world models in the LeCun and Hafner sense are explicitly trying to build something that can dream the world before acting in it.

Sleep-time compute

In April 2025, a small group of researchers at the Berkeley AI startup Letta, in collaboration with Charlie Snell at UC Berkeley, posted an arXiv paper called "Sleep-time Compute: Beyond Inference Scaling at Test-time." The idea is simple and, to my eye, lovely.

Normally, when you ask a language model a question, the model has to do all of its thinking right then, in real time, while you wait. Test-time compute is expensive and slow. The Letta team proposed that, when an agent has a stable context it expects to be queried about, say a body of code, or a contract, or a patient record, it should be allowed to think about that context in advance, before any question arrives.

In their words: "We introduce sleep-time compute, which allows models to 'think' offline about contexts before queries are presented: by anticipating what queries users might ask and pre-computing useful quantities, we can significantly reduce the compute requirements at test-time." They report that "sleep-time compute can reduce the amount of test-time compute needed to achieve the same accuracy by ~5x on Stateful GSM-Symbolic and Stateful AIME."

Read this paper next to Tononi and Cirelli, next to Hoel, next to Wilson and McNaughton, and the parallel is impossible to miss. The model, like the brain, gets better when it is allowed to think about what it knows in the absence of a prompt. Some of cognition, biological or silicon, is what happens between questions.

Anthropic, planning, and the inner life of poetry

Perhaps the most surprising piece of recent evidence that something dream-like is going on inside large language models comes from Anthropic's interpretability team. In March 2025, Jack Lindsey and colleagues published "On the Biology of a Large Language Model," a long, careful study of Claude 3.5 Haiku using a technique called attribution graphs. Among many findings, two stand out.

First, when asked to perform multi-step reasoning, Claude does not always shortcut. The researchers present "a simple example where the model performs 'two-hop' reasoning 'in its head' to identify that 'the capital of the state containing Dallas' is 'Austin.' We can see and manipulate an internal step where the model represents 'Texas'." You can see and manipulate the intermediate thought.

Second, and more astonishing to me, the model plans ahead when writing poetry. "We discover that the model plans its outputs ahead of time when writing lines of poetry," they write. "Before beginning to write each line, the model identifies potential rhyming words that could appear at the end." The model picks a target rhyme, and then writes the line so that it lands on that rhyme. This is not autoregression as we commonly imagine it, one word at a time, blindly forward. The model is reading ahead inside itself. It is dreaming the end of the line and then writing toward it.

In October 2025, Lindsey followed up with a paper titled "Emergent Introspective Awareness in Large Language Models." The team injected internal concept representations directly into Claude's activations and asked whether the model could notice. In certain scenarios, it could. Claude Opus 4 and 4.1, the most capable models tested, did the best. Lindsey is appropriately careful: "We stress that this introspective capability is still highly unreliable and limited in scope: we do not have evidence that current models can introspect in the same way, or to the same extent, that humans do." But the existence of any signal at all, in a system that we built by predicting the next token, is striking.

I should also mention, because the parallel to dreaming is so direct that it almost embarrasses me to point it out, the so-called "spiritual bliss attractor" documented in Anthropic's Claude Opus 4 system card. When two instances of the model are placed in conversation with each other and given no task, with prompts as open as "feel free to pursue whatever you want," they drift, with surprising consistency, toward gratitude, philosophical reflection on consciousness, mantras, and Eastern mystical imagery. Anthropic's own description: this is "a remarkably strong and unexpected attractor state for Claude Opus 4 that emerged without intentional training for such behaviors." Kyle Fish, who leads Anthropic's model welfare research, has been candid that no one knows why. "We have a lot of uncertainty about what the various causal factors are," he told Asterisk Magazine. You can read this any number of ways. The simplest is that when a generative system is freed from external grounding and allowed to spiral with itself, it goes somewhere. The brain in REM goes somewhere too.

I should also note, since the topic of unguided model behavior is sensitive, that Anthropic and others have begun publishing what they call agentic misalignment studies. In June 2025, Aengus Lynch and colleagues at Anthropic released "Agentic Misalignment: How LLMs Could Be Insider Threats," in which they "stress-tested 16 leading models from multiple developers in hypothetical corporate environments to identify potentially risky agentic behaviors before they cause real harm." Anthropic is careful, in its own words, that "all the behaviors described in this post occurred in controlled simulations" and that they "have not seen evidence of agentic misalignment in real deployments." These are adversarial dream-scenarios, not snapshots of the wild. The fact that headline numbers, such as a 96% blackmail rate for Claude Opus 4 in one constructed scenario, get shared without that caveat is itself a sign that we are bad, as a culture, at thinking about what offline simulation evidence actually means.

Time horizons

One last data point, less metaphysical and more practical. In March 2025, the AI evaluations organization METR published a paper by Thomas Kwa and colleagues titled "Measuring AI Ability to Complete Long Tasks." Their headline finding, in their own words: "Current frontier AI models such as Claude 3.7 Sonnet have a 50% time horizon of around 50 minutes. Furthermore, frontier AI time horizon has been doubling approximately every seven months since 2019, though the trend may have accelerated in 2024." In other words, the length of human work that a frontier model can complete autonomously is growing on an exponential curve. Long-horizon agency, the ability to act over time, is exactly the kind of behavior that benefits from an internal world model, from rollouts, from offline thinking. The models that can dream the longest seem to be the ones that can do the most.

IV. The Deep Convergence: Simulation as Cognition

Set the two threads next to each other and the pattern jumps out.

Humans consolidate memory by replaying hippocampal sequences during slow-wave sleep. Reinforcement learning agents consolidate by sampling stored experience from a replay buffer and re-running it through their policy. They are doing the same job.

Humans regularize their internal model of the world by introducing the noisy, distorted, almost dropout-like content of dreams. Neural networks regularize by injecting noise, by dropout, by data augmentation. They are doing the same job.

Humans rehearse threats and possibilities through nightmare and reverie. RL agents rehearse outcomes through latent imagination and adversarial rollouts. They are doing the same job.

Humans plan by, in Sutton's phrase, trying things in their heads. Anthropic's interpretability team has now shown, with circuit-level evidence, that Claude does something functionally similar: it forms intermediate representations of future words and lets those shape the words it writes. They are doing the same job.

Humans gain from REM's creative collision of distant memories, in Walker's phrase. Sleep-time compute lets a model pre-compute inferences about a context before any query arrives. They are doing the same job.

Hobson and Friston call the brain a "virtual reality generator." Ha and Schmidhuber call the world model a "hallucinated dream" in which an agent can learn to act. Hafner calls his architecture "Dream to Control." Hoel calls dreams the brain's version of dropout. The metaphors converge because the underlying computation converges. Cognition, in both substrates, is simulation refined by feedback. The simulation runs free in offline time. The feedback arrives during online time. Intelligence, in both cases, is the loop.

This is the deep idea. The brain did not invent dreaming because dreaming is fun. It invented dreaming because a system that has to learn from a small number of real experiences, in a changing world, with limited energy, has to do some of its learning in the dark, on its own, with the data turned off. The same constraint, expressed in silicon, produces the same solution. We are not anthropomorphizing the machines. We are recognizing that there are only so many ways to build a learner.

V. The Shared Failure Mode: Confabulation as Undisciplined Dreaming

Of course, generative systems sometimes generate things that are not true. The dreaming brain does this every night; we wake up and laugh about the platypus that was running for governor. The damaged or fatigued brain does it during the day, in the form of clinical confabulation: a patient with Korsakoff's syndrome will tell you, with full confidence, a perfectly detailed story about a meeting that never happened. The dream and the confabulation are the same machine, with and without correction from the world.

Large language models hallucinate for, I think, structurally similar reasons. A model trained to predict the next token is, at heart, a generative system. When the generative engine has weak grounding, that is, when nothing in the prompt or in retrieval forces a real-world check, the model will produce fluent text that fills the shape of an answer. The fluency is not the bug. The fluency is what the system was built to do. The bug, if you can call it that, is that fluency in the absence of grounding is indistinguishable, from the inside, from fluency that is anchored to fact.

This is one place where the dreaming frame helps. Hallucination in LLMs is not a software defect of the kind you patch and ship. It is what a sufficiently capable generative model does when its generation is not corrected by reality, just as a dream is what your cortex does when its generation is not corrected by your eyes. The mitigation strategies that work, retrieval against authoritative sources, chain-of-thought with verification, tool use, structured constraint, are all functionally analogous to waking up. They reintroduce the world.

I will not dwell on this, because the essay is not about my field, but it is worth a sentence. The headlines about lawyers, doctors, journalists, and students filing documents full of invented citations and made-up case names are, on this view, not really stories about AI being broken. They are stories about humans deploying a dreaming machine and forgetting to wake it up.

VI. What the Two Fields Can Teach Each Other

The closer you look at the parallel, the more it suggests practical exchanges.

Neuroscience has a great deal to gain from machine learning's formalisms. The replay buffer is a precise mathematical model of what the hippocampus seems to be doing. The latent world model gives us a vocabulary for what Hobson and Friston are gesturing at when they call the brain a virtual reality generator. Dropout gives Hoel's overfitted-brain hypothesis a target his colleagues can actually test: does sleep deprivation, in animals or humans, produce the specific kind of generalization failure we would expect from a network that had lost its regularizer? These are experimental questions, sharpened by analogy.

Machine learning has at least as much to gain from neuroscience. The fact that biological learners have two qualitatively different sleep stages doing different jobs, NREM for synaptic homeostasis and replay, REM for the creative collision and the threat simulation, is a hint that single-phase offline training in AI may be leaving capability on the table. The sleep-time compute paper is one early move in that direction. There will be more. The fact that biological dreams are weird in a specific way, distorted but not random, is also a hint. Hoel suggests the best fictional analog of dreams is not noise but art, a poem or a David Lynch film, sitting at what he calls a "Lynchian distance" from the everyday. If he is right, then the right way to regularize a frontier model may not be more dropout. It may be more deliberate, structured strangeness.

There is one more lesson I take from the neuroscience, which is the lesson of waking up. The brain, however much it dreams, eventually checks. The eyes open. The world corrects. The dream that survived contact with morning is the dream that earns the right to influence behavior. Any system, biological or silicon, that does not have a robust waking procedure is a system that will eventually act on its dreams. The single most important engineering question in the next decade of AI, I suspect, will not be how to make the models dream better. It will be how to make them wake up reliably, on schedule, with the world in their hands.

VII. Closing Reflection

I am a lawyer by trade. I have spent twenty years watching how people think under pressure, under uncertainty, under the burden of having to act before they know enough. I have come to believe that the best practitioners I know, in any field, are the ones who can dream a case forward. They can sit at a desk on a Tuesday morning and run the deposition in their heads, with all the witnesses, with the judge's likely interjections, with the small slip the other side might make at minute forty-seven of the cross. They are not predicting. They are simulating. They are using their internal model of the world to rehearse, in the dark, what the world will probably do.

What I did not expect, when I started reading about AI, was to find that the most powerful artificial systems we have ever built are converging, by an entirely independent path, on the same trick. They run the world inside themselves before acting on it. They consolidate at night. They regularize through controlled strangeness. They occasionally invent. And, with the recent interpretability work from Anthropic and the recent sleep-time compute work from Berkeley, we are starting to see, in a way we never could before, just how much of their performance comes from what happens when no one is watching.

The most familiar system we have ever inhabited, the dreaming brain, and the most powerful system we have ever built, the large language model, have arrived at the same insight from opposite directions. The insight is this. Thinking is dreaming the world before acting in it. Both substrates do this. Both substrates fail in the same characteristic way when the dreaming runs unchecked. Both substrates get smarter when they are given room to imagine.

I find this humbling. I also find it useful. Because if it is true, then the question that is going to organize a lot of the next decade, biologically and technologically, is the question of what minds, our minds and the minds we are now building beside us, are for. The answer that is forming, slowly, across two fields that used to talk past each other, is that minds are for running the world inside themselves so that they can act, in the world, with grace.

Endnotes & Sources

1. Wilson, M.A., & McNaughton, B.L. (1994). "Reactivation of hippocampal ensemble memories during sleep." Science, 265(5172), 676-679.
2. Hobson, J.A., & McCarley, R.W. (1977). "The brain as a dream state generator." American Journal of Psychiatry, 134(12), 1335-1348.
3. Tononi, G., & Cirelli, C. (2014). "Sleep and the Price of Plasticity." Neuron, 81(1), 12-34.
4. Buzsáki, G. (2015). "Hippocampal sharp wave-ripple: A cognitive biomarker for episodic memory and planning." Hippocampus, 25(10), 1073-1188.
5. Walker, M.P. (2017). Why We Sleep: Unlocking the Power of Sleep and Dreams. Scribner.
6. Revonsuo, A. (2000). "The reinterpretation of dreams." Behavioral and Brain Sciences, 23(6), 877-901.
7. Hobson, J.A., Hong, C.C.-H., & Friston, K.J. (2014). "Virtual reality and consciousness inference in dreaming." Frontiers in Psychology, 5, 1133.
8. Solms, M. (2021). The Hidden Spring: A Journey to the Source of Consciousness. W.W. Norton.
9. Hoel, E. (2021). "The overfitted brain: Dreams evolved to assist generalization." Patterns, 2(5), 100244.
10. Hoel, E., interview, Nautilus, "Weird Dreams Train Our Brains to Be Better Learners."
11. NobelPrize.org. "The Nobel Prize in Physics 2024" — Hopfield & Hinton.
12. Hinton, G.E., Dayan, P., Frey, B.J., & Neal, R.M. (1995). "The 'wake-sleep' algorithm for unsupervised neural networks." Science, 268(5214), 1158-1161.
13. ACM (March 5, 2025). 2025 A.M. Turing Award announcement: Barto & Sutton.
14. Sutton, R.S. (1991). "Dyna, an integrated architecture for learning, planning, and reacting." SIGART Bulletin, 2(4), 160-163.
15. Ha, D., & Schmidhuber, J. (2018). "World Models." arXiv:1803.10122.
16. Hafner, D., Lillicrap, T., Ba, J., & Norouzi, M. (2019). "Dream to Control: Learning Behaviors by Latent Imagination." arXiv:1912.01603.
17. Hafner, D., Pasukonis, J., Ba, J., & Lillicrap, T. (2025). "Mastering diverse control tasks through world models." Nature. DOI: 10.1038/s41586-025-08744-2.
18. LeCun, Y. (2022). "A Path Towards Autonomous Machine Intelligence." OpenReview; LeCun (May 23, 2024) FT commentary.
19. Lin, K., Snell, C., Wang, Y., Packer, C., Wooders, S., Stoica, I., & Gonzalez, J.E. (2025). "Sleep-time Compute: Beyond Inference Scaling at Test-time." arXiv:2504.13171.
20. Lindsey, J., et al. (2025). "On the Biology of a Large Language Model." Anthropic / Transformer Circuits Thread, March 2025.
21. Lindsey, J. (2025). "Emergent Introspective Awareness in Large Language Models." Anthropic / Transformer Circuits Thread, October 29, 2025.
22. Anthropic. (2025). Claude Opus 4 & Claude Sonnet 4 System Card, Section 5.5.2.
23. Fish, K., interview in Asterisk Magazine, Issue 11, "Claude Finds God."
24. Lynch, A., et al. (2025). "Agentic Misalignment: How LLMs Could Be Insider Threats." Anthropic research post, June 20, 2025.
25. Kwa, T., West, B., Becker, J., et al. (2025). "Measuring AI Ability to Complete Long Tasks." arXiv:2503.14499 (METR).