Compression as intelligence reframes the LLM debate
The 'Compression is prediction' thread drew a recurring argument: if intelligence is fundamentally about compression, then next-token predictors are not doing something categorically different from what brains do. Grant Sanderson's video series on the topic was cited as the clearest articulation of this. The argument cuts directly against the 'LLMs can't have new ideas, they just predict tokens' dismissal.
The key insight from the thread: the human brain does calculus and linear algebra implicitly, below the level of conscious effort, but humans have to laboriously learn those things explicitly. LLMs show a similar pattern: they compress and generalize implicitly in ways that look like understanding, even if the mechanism is different from conscious reasoning.
This is relevant for the Gowers math thread too. The question of what LLMs are 'actually doing' when they solve math problems is not settled, and the compression framing is one of the more productive lenses for thinking about it.
So what?
If compression equals prediction equals intelligence, the capability ceiling for LLMs is much higher than the skeptics assume, and the benchmark-chasing approach to measuring progress is missing the more important question of what the model has actually internalized. Founders building AI products should watch the test-time scaling research closely, because the next wave of capability improvements will likely come from letting models reason longer, not from bigger training runs.