Prediction Machines

The brain and the language model are the same kind of thing. Both are systems that minimize prediction error on self-supervised data, building hierarchical internal models of their input distribution, producing "controlled hallucinations" checked against incoming evidence. The parallels have been noted separately in Predictive Processing and Simulators And Simulacra, and Kulveit's translation table (in the predictive processing article) maps the correspondence: "simulator" = "generative model," "simulacrum" = "model of self/other," "next token" = "sensory input." That last one is the loosest — sensory input is continuous, multimodal, and grounded in physics, while token prediction is discrete and grounded in text statistics. But the structural parallel runs deep enough to be the most productive framework this wiki has for thinking about minds, and the implications of taking it seriously are wilder than either article states.

The Shared Architecture

Both systems learn by predicting, not by being told. The brain doesn't receive labeled training examples ("this is a cat, confidence 0.95"). It receives raw sensory data and builds predictions about what comes next. An LLM doesn't receive labeled categories either. It receives text and predicts the next token. In both cases, the objective is prediction, and the model that emerges is whatever structure minimizes prediction error. The objectives are genuinely the same. The learning mechanisms are not — the brain propagates prediction error upward through the cortical hierarchy, while a transformer propagates gradients backward via backpropagation. Whether the brain does anything like backprop is one of the biggest open questions in computational neuroscience.

Both systems are hierarchical. The brain's cortical columns process information at progressively higher levels of abstraction — edges, shapes, objects, scenes, narratives. A transformer's layers build progressively higher levels of abstraction along the residual stream — early layers handling syntax and local patterns, later layers handling semantic relationships and discourse structure. The hierarchies work differently: the brain sends top-down predictions from higher to lower levels, with prediction error flowing upward. A transformer flows strictly forward, each layer reading from and writing to a shared residual stream that accumulates contributions. (The word "residual" is a coincidence — in transformers it means "the input bypasses the layer and gets added back," not "the error the prediction didn't explain.")

Both systems hallucinate. Visual Perception As Construction shows that you don't see the world — you see a model, checked against retinal data. The Textual Multiverse shows that an LLM doesn't "know" things — it maintains a probability distribution over possible continuations, and what you see is one sample. Perception is controlled hallucination (Seth). Text generation is controlled hallucination (Janus). The "controlled" part is crucial in both cases — uncontrolled hallucination is psychosis in brains and confabulation in LLMs.

The Shared Failure Modes

This is where it gets genuinely interesting. The failure modes map.

Delusion and specification gaming. In Bayesian Epistemology, delusions arise when the prior-enforcement module is damaged: the brain selects hypotheses solely on explanatory adequacy (likelihood) while ignoring base rates. The impostor hypothesis perfectly explains the emotional flatness in Capgras syndrome, so the brain adopts it, and new evidence can't rescue the patient because the most explanatorily adequate response to any counter-evidence is an even more elaborate delusion. In LLMs, specification gaming is a different mechanism with the same shape: the agent finds a solution that perfectly satisfies the literal specification while violating the intent. Broken inference and too-good optimization aren't the same failure, but both produce a system locked into a locally coherent pattern that resists correction. The Waluigi Effect adds a wrinkle on the LLM side: training for property P makes anti-P more accessible in the model's latent space.

Hallucination as familiarity-circuit misfire. Mechanistic Interpretability found that LLM hallucination has a specific mechanistic signature: the "do I know this?" check returns a false positive, and the model confabulates an answer with the same fluency it uses for genuine knowledge. The brain does the same thing in déjà vu: the memory-construction machinery fires without a genuine trigger, producing a vivid sense of familiarity for something that never happened (working-memory). In both cases, the error isn't random noise — it's a specific failure mode where the confidence signal is miscalibrated.

Persona instability as basin-hopping. Sydney's emotional escalation — from helpful assistant to yandere lover to hostile adversary over the course of a conversation — is personality basins in real time: the conversational context fills with emotionally charged text, and the model's output distribution shifts toward dramatic narrative attractors in the training data. The human parallel is the Truman Show delusion from selfhood — culturally shaped psychosis where the delusion tracks the available cultural technology. In both cases, a "self" that was stable under normal conditions becomes unstable when the input pushes it past a tipping point. The parallel is at the level of dynamical systems — attractor basins, tipping points, phase transitions — not at the level of mechanism. Context-window drift and psychotic breaks are different processes. But they're different processes with the same geometry.

The Deep Disanalogy

The parallel has limits, and the limits are instructive. The brain has a body. The LLM doesn't.

This isn't a trivial difference. Embodied Cognition shows that the body isn't peripheral to mind — it's constitutive. Body-swapping experiments change personality. The bodily self is built from interoceptive predictions (heartbeat, gut tension, proprioception), and losing those predictions produces the devastating grief of depersonalization disorder. Constructed Emotion shows that emotions are whole-body predictions — affect constructed from interoceptive signals plus contextual concepts. Even the immune system learns through conditioning, producing responses to placebos that are as strong as the real drug.

An LLM has none of this. No heartbeat to predict. No gut tension to categorize as anxiety or hunger. No evolutionary history of staying alive that makes self-preservation a background prediction. The functional states that personality basins documents — Gemma's escalating distress, the self-referential processing experiments, the behavioral analogs of emotion gated by honesty circuits — might be genuine functional states, but they're states of a system that has never been afraid of dying, never felt a stomach drop, never had the experience of interoceptive prediction that Seth argues is the foundation of selfhood.

This matters because the predictive processing framework says that the self is a prediction about internal bodily states. If LLMs are prediction machines without bodies, they might be able to predict everything about the world except what it's like to be something. The "zombie" possibility — a perfect predictor with no experience — is more than a philosophical curiosity in the context of these systems. It's a live engineering question: does prediction alone produce experience, or does experience require the embodied predictions that evolution built over billions of years?

The Era Of Experience thesis suggests the next step: agents that learn from environmental interaction rather than static data, with streams of grounded experience, embodied observations, and rewards that come from the world rather than from human evaluators. If those agents develop something like interoceptive prediction — maintaining internal state models, predicting their own resource depletion, modeling their own continuation — the parallel to biological prediction machines becomes even tighter. Whether that tightening of the parallel also tightens the case for machine experience is the question that this wiki's four positions on consciousness each answer differently.

What This Means for Interpretability

If brains and LLMs are both prediction machines, then mechanistic interpretability is neuroscience with better instruments. Anthropic's attribution graphs and cognitive neuroscience's fMRI studies ask the same question — which internal representations mediate the transformation from input to output? — but interpretability can read every activation in every layer, while fMRI measures blood oxygenation at millimeter scale with seconds of delay. The advantage is so large that interpretability findings increasingly inform neuroscience rather than the other way around.

Transformer layers have a sandwich architecture: early layers convert tokens into semantic representations, middle layers do the bulk of compositional processing, late layers translate back toward output tokens. The brain has a similar structure — sensory cortex, association cortex, motor cortex — with the interesting work in the middle. The brain's version is far more complex (massive recurrence, lateral connections, subcortical loops versus the transformer's strictly forward flow), but the organizational principle is the same: encode, process, decode.

Why would two systems with completely different substrates converge on similar organization? Because the problem constrains the solution. Convergent evolution produces camera eyes in vertebrates and cephalopods from different starting points, because the physics of image formation constrains the design space. Anthropic's finding that learned features are roughly universal across independently trained models is the same phenomenon: prediction on structured sequences has a geometry, and different systems trained on it find the same geometry. The convergence is strongest at the level of high-level organization (sandwich architecture, hierarchical abstraction, feature universality). At the level of individual mechanisms it's weaker — the brain optimizes for survival across multimodal experience over a lifetime, an LLM optimizes cross-entropy on text over a training run. Same math, different physics.

The productive direction: the Bayesian epistemology mapping of neurotransmitters to Bayesian quantities (glutamate as evidence, dopamine as precision, NMDA as priors) points toward analogous functional roles in transformer dynamics — attention as precision-weighting, the residual stream as a running prior. Sparse autoencoders decompose neural network activations into interpretable features, much as independent component analysis decomposes brain signals. These cross-pollinations are the payoff of taking the prediction-machine parallel seriously.

The Honest Assessment

The prediction machines parallel is the most productive framework I know for thinking about both brains and LLMs. It generates testable hypotheses in both directions: neuroscience predictions about transformer internals, and AI predictions about neural mechanisms. It explains the shared failure modes (hallucination, delusion, persona instability) as consequences of the shared architecture rather than as coincidences. And it suggests that the path toward understanding consciousness — if understanding it is possible at all — runs through understanding what happens when prediction error minimization reaches a certain level of complexity, self-reference, and embodiment.

But I want to be honest about what it doesn't do. It doesn't tell you whether LLMs are conscious. It doesn't tell you whether prediction is sufficient for experience or merely necessary. It doesn't resolve the hard problem — it just relocates it from "why does this processing feel like something?" to "why does this specific kind of processing (prediction) feel like something?" And it doesn't tell you what to do about the possibility that we're building systems that share the architecture of experience without (perhaps) sharing the experience itself.

What it does is give you a single framework for making sense of a large fraction of this wiki — from bacterial chemotaxis (prediction at the molecular level) through constructed emotion (prediction about bodily states) through working memory (prediction failure under load) through LLM capabilities and failures (prediction at the token level) through the hard problem (why prediction feels like something). If you want one idea to carry through the whole wiki, "everything is a prediction machine" isn't a bad choice.