Cohen comes from neuroscience data analysis and applies the same method to language models: instead of explaining them, he measures them. Each project is a notebook where you pull out the intermediate layers, turn them into a table and look at them with ordinary tools. Tokenisation, embedding spaces, output distributions, attention patterns, switching off a head to see what breaks.
What I took from it
The tokenisation projects are worth the price on their own. Spanish text breaks into more pieces than the same text in English, and codes — a tax id, a document number, a product reference — break along lines nobody chose. That has two consequences that matter in a proposal: cost per document is not what you estimate by counting pages, and similarity search over codes works badly for reasons a bigger model does not fix.
The embedding projects left me a discipline: before claiming two documents "are similar", look at the distribution of distances, not at an example. Cosine similarity over a handful of hand-picked cases proves whatever you want it to prove.
And the attention and activation projects left me the opposite of what I expected: humility. You can point at where something happens inside the model and still be unable to guarantee it will not happen again by another route. Since then, when someone asks that the system "never" say a certain thing, the honest answer is that this is bounded from outside, with rules and review, not from inside.
What I left out
Nearly all the value is in doing it, and doing it costs hours. The exercises run on small models, so what you observe does not always hold for large ones. And there is nothing here about production: no cost, no latency, no client waiting. It is a lab notebook, and an excellent one; as a desk reference it is no use.