Implications
What changes if cognition is autoregressive generation
If the brain is an autoregressive generator, then every major cognitive phenomenon looks different than traditional accounts assume. What follows is not speculative extension. These are direct consequences of the architecture: a system that generates each next experiential state as the optimal continuation of its own prior output, conditioned on current sensation. Each section identifies what changes and why.
Perception
Traditional view: The brain receives sensory input and constructs a representation of the world.
Autoregressive reframing: Perception is output, not input. Sensation arrives; what the system experiences is what the generator produces in response. These are not the same thing. The system “sees” what it generates, not what it receives. Sensation constrains generation but does not constitute it.
This is why illusions are not errors. Filling-in, amodal completion, size constancy, the Gestalt grouping principles: these are the generator doing exactly what it always does, producing the most probable continuation of the perceptual trajectory given available constraints. Change blindness follows naturally. If the change does not disrupt the generator’s continuation, it is not generated as part of the experiential state, and so it is not perceived. There is no “failure to detect” because there is nothing that detects. There is only generation.
Perceptual experience is always richer than the sensory signal warrants because the system cannot deliver raw sensation to experience. It can only generate a state that is consistent with sensation, shaped by everything the generator has learned.
Attention
Traditional view: Attention is a spotlight, a filter, or a limited resource that selects information for further processing.
Autoregressive reframing: Attention is the autoregressive process itself. The brain performs massive parallel processing: sensory transduction, feature extraction, autonomic regulation, reflexive loops. All run concurrently. Cognition is the regime in which this enormous parallel state is funneled into a single serial stream: one generated state at a time.
What we call “attending” to something is that thing dominating the current generation step. The system can only generate one perceptual-cognitive state per cycle, and whatever content shapes that state is what the system is attending to. Inattentional blindness is not a failure to notice something that was there. It is a failure to generate it. If it was not part of the generated state, it was not part of experience.
The longstanding puzzle of how attention relates to consciousness dissolves. They are the same thing. The serial nature of conscious experience is not a limitation imposed by scarce resources. It is a structural feature of the architecture.
Memory
Traditional view: The brain stores experiences and retrieves them later. Short-term and long-term memory are separate systems with a transfer mechanism between them.
Autoregressive reframing: Memory is not storage and retrieval. It is the continued influence of past generation on present generation, operating at multiple timescales.
What we call “remembering” is reconstructive generation. When you recall your tenth birthday, you do not play back a recording. You generate a plausible trajectory through experiential state-space, constrained by available cues and by the distributional structure learned over a lifetime. This is why memories are context-sensitive (different cues seed different trajectories), malleable (each act of remembering is a new act of generation), and seamlessly continuous with imagination (both are generation, differing only in constraints).
The short-term/long-term distinction dissolves. There are not two memory systems. There is influence at different timescales. Fast state carries influence that decays over seconds. Slow structure carries influence that persists across a lifetime. The difference is the time constant of the substrate, not the nature of the operation.
False memories are not corrupted files. They are trajectories generated coherently but diverging from the historical trajectory, often because the available cues were compatible with multiple continuations. The system has no way to distinguish a “true” memory from a “false” one, because both are generated by the same process. There is no original to compare against.
The forgetting curve is not the rate at which the brain loses what it stored. It is the rate at which successive autoregressive compression reduces the specificity needed for regeneration. This is developed formally in the information geometry of memory.
Consciousness
Traditional view: Consciousness is a mysterious property that somehow arises from neural computation. The “hard problem” asks how physical processes produce subjective experience.
Autoregressive reframing: The serial, contextual sequence of generated states is not correlated with conscious experience. It is conscious experience. What enters the generated sequence is conscious; what remains in the parallel sensory state without being integrated into that sequence is not.
This gives unity of consciousness for free: there is one generator, producing one state at a time. It gives the stream quality of experience: each state conditions the next, producing the continuous temporal flow that characterizes subjective life. And it dissolves the binding problem: the question “how does the brain combine color, shape, and motion into a unified percept?” presupposes that these are computed separately and then combined. In the autoregressive architecture, they are generated together as aspects of a single output. There is nothing to bind.
Why anesthesia abolishes consciousness: it disrupts the autoregressive loop. Why dreams are conscious: the generator runs with sensation largely disconnected, but generation continues. Why flow states feel different: the generator runs with reduced self-modeling in the generated state, still conscious but less conscious of itself.
The deeper philosophical implications of this account, including the dissolution of the hard problem and the reframing of subjectivity as recursive navigation, are developed on the Philosophical page.
Reasoning and Decision Making
Traditional view: Reasoning is performed by a logic engine or separate reasoning module. System 1 and System 2 are distinct cognitive systems.
Autoregressive reframing: Reasoning is extended autoregressive generation with the generated trajectory itself as the dominant conditioning content. The system generates a state, that state conditions the next, and the chain extends through a sequence of inferential steps. There is no separate reasoning module. There is depth of generation.
The System 1/System 2 distinction is reframed as the difference between low-depth and high-depth generation. Fast automatic responses are shallow trajectories: the generator produces a continuation in one or a few steps, driven by well-learned distributional structure. Deliberate reasoning is deep generation: the system extends the trajectory through many steps, producing the serial, effortful character of conscious thought.
Decision making is generation under outcome constraint. The system generates possible trajectories forward from the current state, and the trajectory that dominates is the one most consistent with the generator’s learned structure. The system does not compute expected utility. It generates continuations and is shaped to favor those whose downstream sensation has historically been favorable. This is why decisions feel like something: they are experienced as generated trajectories, not as computed outputs. Indecision is the generator failing to settle on a single continuation. Regret is generating the alternative trajectory after the fact.
Imagination and Creativity
Traditional view: Imagination is a separate faculty. Creativity involves novel combination of stored concepts.
Autoregressive reframing: Imagination is generation with sensation playing a reduced or absent role. The same process that produces perception produces imagery when sensation is attenuated. This is why mental imagery shares neural substrates with perception, why imagined actions activate motor-associated areas, and why vivid imagination can be confused with memory. The generator does not distinguish between generating a percept from sensation and generating an image from a cue. The process is identical; the constraints differ.
Creativity is generation at the edge of learned distributional structure. The generator produces trajectories that are improbable but not impossible given its learned constraints: novel enough to be interesting, coherent enough to be intelligible. Highly creative states often involve the relaxation of the most constraining distributional priors, allowing the generator to explore regions of state-space it would not normally visit. This is consistent with the association between creativity and states of reduced cognitive control, including mind-wandering, dreaming, and certain pharmacological states.
The Self and Agency
Traditional view: There is a self that perceives, decides, and acts. Agency is a basic property of the organism.
Autoregressive reframing: The “I” is a generated construct, not a foundational entity. It is a token, perhaps the most powerful token in the system’s repertoire, that compresses an enormous amount of trajectory regularity into a form the generator can efficiently condition on. “I” organizes the relationship between generated states and returning sensation into a coherent narrative. But it is authored by the generator, not observed by it.
The system has no direct access to its own behavior. Motor commands, hormonal signals, autonomic activation are consequences of the neural activity that generates experience, not outputs that the system observes. What the system feels is not the motor command but proprioception. Not sympathetic activation but its heart racing. Every channel through which the system might “know” what it is doing is mediated by sensation and processed into perception.
Attribution of agency is always a generated inference, never a direct observation. The system cannot distinguish self-caused sensation from world-caused sensation. Both arrive as ordinary input. Over time, temporal contiguity gets compressed into the construct of agency: “I did that.” This is an efficient generative token, not a report from a privileged self-monitoring channel. Attribution errors are not failures of an otherwise reliable system. They are the baseline. Attribution is always confabulation that happens to align, more or less, with external causal structure.
LLMs exhibit the same phenomenon: they use “I” with apparent coherence, not because they have a self-model, but because “I” is a statistically powerful token that constrains continuation.
Learning and Development
Traditional view: Learning is a separate phase of operation. The brain stores experiences during learning and uses them later during performance.
Autoregressive reframing: Learning is not a separate phase. Every cycle of the generator simultaneously produces the current experiential state and encodes it into the system’s structure. Fast state is updated with the current trajectory; slow structure is reshaped by the current activity. There is no distinct write operation. Generation is encoding.
A single optimization process shapes both behavior and representation. The system does not just learn what to generate; it learns how to generate in a form optimized for its own subsequent ingestion. Development is partly the progressive refinement of output format. The babbling infant is not practicing motor commands. It is calibrating the generator’s output format against returning sensation.
Inner speech may be specifically useful because linguistic tokens are a format co-optimized with the machinery that runs on them. Language is not just a communication tool but an internal representational technology that makes the generator more effective at conditioning on its own output.
Behavior
Traditional view: The brain computes motor commands and sends them to the body. Behavior is an output of cognition.
Autoregressive reframing: Behavior occupies a unique position in the framework: it is the entire reason the system exists, but it is computationally invisible to the system that produces it. The neural activity that generates each perceptual-cognitive state also drives muscles, glands, and autonomic pathways, by anatomy, not by computation. The generator does not compute motor commands as outputs. It generates experience. Behavior happens because the tissue that generates experience is connected to effectors.
The consequences of behavior return to the system only as ordinary sensation, indistinguishable from any other input. The system’s own speech arrives as audition. The visual consequences of its movements arrive as vision. There is no action-feedback channel. The system never observes what it does. It only experiences what comes back.
This indirection is the architecture. The system optimizes over what it can see (its own generated experiential trajectory) for outcomes it cannot see (the behavioral consequences that exit the body and change the world). This is why credit assignment is hard, why delayed consequences are difficult to learn from, why immediate reinforcement is more effective than delayed reinforcement, and why superstitious behavior exists. The system is learning from temporal contiguity in a stream where self-caused and world-caused sensation were never distinguished.
74 Phenomena, One Principle
The autoregressive framework generates specific, testable accounts across the full range of cognitive phenomena: behavioral, psychological, neurophysiological, and neuroanatomical.
In language, the framework explains garden-path sentences, incremental parsing, structural priming, and the N400 as natural consequences of serial commitment. In memory, it accounts for false memories, context-dependent recall, reconsolidation, serial position effects, and the dissolution of the short-term/long-term distinction. In perception and consciousness, it explains change blindness, perceptual completion, binocular rivalry, and the stream of consciousness as properties of a system that generates rather than records experience. The framework extends to deliberation and reasoning, social cognition, aesthetics, and clinical phenomena including schizophrenia, confabulation, PTSD, and aphasia.
74 phenomena across 12 domains, unified under a single computational principle.