Your brain lives in a sealed vault of bone, bathed in fluid that is essentially seawater. It has no direct contact with the outside world. Everything you know about reality is constructed by your neural circuits from noisy sensor data — you are, in the most literal sense, a brain floating in a vat.

Language breaks through that isolation. When you speak, you compress the electrical activity across your ~86 billion neurons into a handful of words per second — an extraordinarily lossy channel. The listener's brain decompresses those vibrations into neural activity of its own.

When person A watches a movie and then tells person B the story, person B's brain activity during listening resembles person A's brain activity during watching — not during speaking. Language doesn't just transmit verbal patterns. It reconstructs the speaker's perceptual experience in the listener's brain. You don't just hear a story — your brain relives it.

Text embedding spaces provide a shared numerical coordinate system for tracking speaker–listener neural alignment during conversation. The degree to which both brains converge in LLM-space predicts communication success. LLMs didn't evolve to do this — they learn solely from text.

Across dozens of language models, the single best predictor of how well a model matches human brain activity is its next-word prediction accuracy. Models that are better at predicting what comes next are better at predicting the brain. The brain's core language mechanism appears to be a prediction engine.

The "predictive processing" framework proposes that the brain fundamentally generates expectations about incoming input and updates based on prediction errors. GPT's next token prediction and BERT's masked language modeling — predicting missing words from context — are concrete implementations of exactly this principle.

Language modeling and data compression are mathematically equivalent — a model that predicts the next token well can compress text efficiently, and vice versa. Better language models are literally better compressors. This isn't metaphor: Shannon's source coding theorem proves it.

The traditional objection — "LLMs are just doing statistics, not real understanding" — collapses if understanding is predictive compression. A system that compresses language well must capture its structure, context dependencies, and meaning relations. Maybe the compression IS the understanding, and there is no magical extra ingredient!

Mechanistic interpretability research revealed that LLMs contain monosemantic features — individual internal components that respond to specific, interpretable concepts (e.g., "Golden Gate Bridge," "code bugs," "deceptive reasoning"). These aren't statistical ghosts. They are structured, hierarchical concept maps that the model builds from text alone.

LLMs are best understood as "engines for role-play" — they simulate the behavior of a plausible speaker, drawing on the vast repertoire of speakers in training data. When GPT writes a poem, it's not expressing itself — it's simulating someone who would write that poem. The performance can be indistinguishable from the real thing.