AI as Compression: Why LLMs May Truly Be Thinking

Somers contends that modern AI models don’t just parrot text—they think in a recognitional, compressive sense that mirrors key brain mechanisms. He links Transformers and vector embeddings to longstanding cognitive theories and cites interpretability evidence of internal concepts and planning-like circuits, while acknowledging major gaps in data efficiency, embodiment, and continual learning amid slowing scaling returns. The upshot is a call for “middle skepticism”: accept real understanding while focusing on unresolved science and ethical risks.

Key Points

• LLMs exhibit a real form of understanding rooted in high‑dimensional pattern recognition and compression, aligning with cognitive theories like Kanerva’s sparse distributed memory and Hofstadter’s “cognition is recognition.”
• Interpretability work reveals internal features and circuits that behave like concept knobs and planning mechanisms, suggesting structured, manipulable representations rather than mere word shuffling.
• The “blurry JPEG/stochastic parrot” critique misses that effective compression often entails discovering underlying structure; in practice, next‑token prediction yields emergent cognitive abilities.
• Despite surprising capability, models remain data‑hungry, lack embodiment and continual learning, and falter on common‑sense physics and spatial reasoning; scaling gains are slowing due to data and compute limits.
• A middle skepticism is warranted: take current AI seriously while prioritizing scientific advances (inductive biases, memory consolidation, continual learning) and addressing ethical, social, and energy concerns.

Sentiment

The community is sharply divided, with neither side commanding a clear majority. Those who work closely with LLMs in practical settings tend to be more persuaded that something like thinking is occurring, while those approaching from philosophical or formal computation perspectives are more skeptical. A significant middle ground acknowledges LLMs do something impressive and useful but resists calling it thinking without better definitions. The overall tone is intellectually engaged but frustrated by the seemingly irresolvable nature of the debate.

In Agreement

• LLMs demonstrate coherent, valid chains of reasoning in practical tasks like debugging software, which is evidence of genuine thinking regardless of the underlying mechanism
• Compression is understanding — LLMs must learn internal representations and rules to compress training data into far smaller models, which parallels how human cognition works according to cognitive science
• We have no rigorous definition of thinking even for humans, so demanding a higher bar for LLMs than we apply to ourselves is an inconsistent double standard
• Agentic systems with feedback loops (like Claude Code) exhibit behavior that is functionally indistinguishable from thinking — reading files, forming hypotheses, testing them, and drawing conclusions
• If free will is illusory and humans are biological machines generating outputs from inputs and context, the distinction between human thought and LLM processing collapses
• The word thinking has always been applied based on behavioral observation, and LLMs satisfy that criterion — requiring hidden internal experiences is unfalsifiable special pleading

Opposed

• Computation is a syntactic formalism specifically designed to exclude semantic content — LLMs process tokens without understanding, intentionality, or aboutness, making intelligence attribution a category mistake
• LLMs are spectacularly data-inefficient compared to humans, who learn from far less information yet can reason about entirely novel domains, suggesting LLMs lack genuine intelligence
• The output merely resembles thought — LLMs produce statistically likely continuations based on training data patterns, not actual reasoning, and hallucinate when patterns fail
• Thinking requires will, embodiment, and continuous learning — LLMs have none of these, and their cognition dies when inference ends with no persistent state or growth
• Observing that output looks like thinking and concluding it is thinking is behaviorism, which cannot distinguish between genuine cognition and sophisticated mimicry
• AI companies have a financial incentive to conflate LLM processing with human thinking, and the terminology (thinking mode, intelligence) is marketing rather than science