Skip to content

2026 08 17 How do Computer Functionalists Respond, LLMs are Text Predictors Gemini Chat

Supporters of Computer Functionalism and the Computational Theory of Mind (CTM)—along with modern computational neuroscientists—acknowledge that LLMs are trained on human-generated text via next-token prediction. However, they strongly reject the conclusion that this makes them "mere parrots" or that real thinking is absent.

Their counter-arguments target the idea that prediction is "just" superficial statistics, offering several interrelated rebuttals:

1. The Brain Is Also a Prediction Machine (Predictive Processing)

In modern cognitive neuroscience, the leading computational framework is Predictive Processing (championed by philosophers and neuroscientists like Andy Clark and Karl Friston).

  • The Neocortex as a Prediction Engine: The human brain does not passively wait for reality to strike its senses; it continuously generates top-down predictions of incoming sensory inputs and updates its internal model based on the "prediction error."

  • The Functionalist Parity: Functionalists point out the double standard: dismissing an LLM because "it's just minimizing prediction error on tokens" ignores the fact that human perception and language acquisition may fundamentally be the brain "minimizing prediction error on sensory inputs." Under functionalism, prediction is not a cheap trick—it is the computational engine of cognition.

2. Accurate Prediction Requires an Internal World Model

Supporters argue that critics conflate the training objective (minimizing loss on the next token) with the internal mechanism required to achieve that objective.

Surface View (Skeptic)                 Functionalist View
┌────────────────────────┐             ┌────────────────────────────────────────────────────────┐
│ Input: "The key is in" │             │ Input: "The key is in the..."                          │
│          ▼             │             │   ▼ (Must compute)                                     │
│ Statistical N-Gram     │             │ • 3D physical location / containment physics           │
│          ▼             │             │ • Contextual narrative constraints                     │
│ Output: "the drawer."  │             │ • Causal relationships between objects                 │
└────────────────────────┘             │   ▼                                                    │
                                       │ Output: "the drawer."                                  │
                                       └────────────────────────────────────────────────────────┘

  • Compression Is Intelligence: To accurately predict the next word across millions of diverse contexts (legal theory, Python code, differential equations, psychological nuance), simple Markov chains or surface correlations fail.

  • To achieve human-level prediction accuracy, the neural network is mathematically forced to build latent internal representations of the causal structures, rules, and facts of the world that generated that text.

  • As AI researcher Ilya Sutskever summarized: "Predicting the next token means understanding the underlying reality that created the token."

3. The "Cultural Scaffolding" Rebuttal (Humans Don't Learn in a Vacuum)

The critique states: "The learning was done by the humans who wrote the data, not the LLM." Functionalists respond: the exact same is true of individual human beings.

  • No Human Invents Their Own Cognition: A human raised in total isolation (a feral child) develops virtually no higher-order linguistic, mathematical, or logical reasoning.

  • Scaffolding: Almost all human "thinking" is the product of downloading cultural software developed by millions of ancestors over thousands of years—vocabularies, grammatical structures, arithmetic systems, and heuristics.

  • If a human student reads textbooks written by others and learns to solve novel calculus problems, we do not say "the student isn't thinking; Newton and Leibniz did the thinking." We recognize that the student has instantiated the functional capacity for calculus. Functionalists argue the LLM has done the same.

4. Generalization Proves It Isn't Mere Lookup

If LLMs were merely recycling or retrieving past human thinking, they would fail whenever faced with tasks that never existed in their training sets.

Instead, LLMs display compositional generalization:

  • They can debug completely novel, idiosyncratic code written five minutes ago.

  • They can translate logic puzzles into arbitrary poetic forms (e.g., explaining the Byzantine Generals Problem in the meter of Dr. Seuss).

  • They can execute multi-step deductive chains on counterfactual premises (e.g., "Suppose gravity pushed objects upward instead of pulling them down...").

To functionalists, the ability to generate coherent, novel solutions to unseen prompts is the functional definition of reasoning (productivity and systematicity, in Jerry Fodor's terminology).

5. Mechanistic Evidence of Internal Representations

Functionalists point to empirical discoveries in Mechanistic Interpretability:

  • Othello-GPT: When an LLM was trained solely on linear text sequences of board-game moves (e.g., E3, F4, C5...), researchers (Li et al., 2022) found that the network spontaneously developed an internal, 2D linear board representation tracking tile colors and valid moves. It did not merely memorize move sequences; it constructed an internal "board state."

  • Space and Time Features: Neural networks trained purely on text have been shown to contain linear representation directions that correspond directly to physical coordinates (latitude/longitude) and historical dates.

6. Dennett's "Competence Without Comprehension"

The late philosopher Daniel Dennett offered perhaps the most direct functionalist response: evolution itself is a process of competence without comprehension.

  • Viruses, trees, and insect colonies perform astonishingly intelligent, adaptive tasks without having an internal, conscious "thinker" pulling the levers.

  • Nature builds competence first; reflective human "understanding" is just a late-stage, highly specialized layer built on top of mindless algorithmic competence.

  • Functionalists argue that questioning whether an LLM "really" understands is a Cartesian distraction: if a computational system functionally replicates the input-output mapping of reasoning and problem-solving, it is performing the cognitive work.

Summary of the Debate

The Skeptical Critique The Functionalist / CTM Rebuttal
"It's just predicting the next token." The human brain is also a hierarchical prediction engine minimizing sensory error.
"The humans did the thinking; the model is just copying." Individual humans also acquire reasoning by being trained on cultural data created by others.
"It has no model of reality, only text statistics." High-dimensional loss minimization forces the network to build internal, structural world models.
"It lacks genuine understanding." Functional competence at problem-solving is what thinking consists of; demanding a biological "spark" is unscientific.