Variation, Sampling Distributions, and the Reference Distribution (Experimenting to Learn, Part II) Is a difference real, or the kind chance produces anyway? Part II of a series built on Box, Hunter and Hunter: why independence matters far more than normality, how a plant's own history and a coin toss each give you a yardstick, and how one paired design finds an effect another analysis misses.
What Statistical Theory Is For (Part II) Part I argued that choosing the model is where most of the risk in an analysis sits, and that statistical theory is least helpful exactly there. So what does help? Five considerations, not one of which is about fitting the data well — and they openly contradict each other.
What Statistical Theory Is For (Part I) Statistical theory sounds like a complete guide to learning from messy data. It isn't. It only goes to work after you have chosen what kind of model to fit, and it does that part extremely well. Choosing the model is left to you — and that is where most of the risk sits.
Practical Usefulness of LLMs (Part I) Stop asking whether large language models think like us and start asking what they move. The bulldozer analogy, the three jobs LLMs are genuinely good at, and why the operator still matters.
Practical Usefulness of LLMs (Part II) Where the cognitive bulldozer actually gets pointed: hyper-personalised education built on a dependency tree, and consulting that maps the whole solution space before anyone writes a recommendation.
Practical Usefulness of LLMs (Part III) RAG retrieves fragments and hopes they add up. Property-Mapping builds the blueprint first. The technical difference between a fast librarian and an architectural plan — and why it decides whether the model misses a dependency.