How Large Language Models Actually Work (Lecture 4)

Published 3 hours ago

For three lectures we stayed at the philosophical altitude. This one comes down to ground level: what ChatGPT, Claude, Gemini, and Llama actually are, mechanically, and why understanding the pipeline matters for every ethical claim anyone makes about them.

A large language model is trained to do one thing: predict what token comes next in a sequence, over and over, billions of times, across enormous datasets scraped and cleaned and weighted by human decisions that are never neutral. I walk through tokenization, where the model doesn't see words but numeric fragments, and the transformer architecture and attention mechanism that let a model figure out what "she" refers to in a sentence. Then I cover the two-stage shaping process that turns a raw text-continuation engine into something that behaves like a helpful assistant: supervised fine-tuning and reinforcement learning from human feedback.

None of this is magic, and none of it is mind. I bring back Apple's Illusion of Thinking result on reasoning collapse. Then I add two newer findings that sharpen the picture: a 2026 Apple paper on refusal neurons, single switches that collapse a model's entire safety alignment when suppressed, and a 2025 Anthropic and Truthful AI paper on subliminal learning, where a model can inherit a trait from training data that never mentions the trait at all.

This is Lecture 4 in a series drawn from my forthcoming book, Vigilant Hope: Preserving Human Dignity in an Age of Artificial Intelligence (Ascension Press). Subscribe for Lecture 5, where we follow this pipeline into the public sphere and ask what happens to truth once machines can generate plausible language at industrial scale.