Research Notes.
Research-backed essays from the LLM mechanics marathon. Structured synthesis, visible citations, and the same JAAX voice we use when the conclusions actually have to survive contact with operators.
Research notes are the long-form end of this site. Each one synthesises several papers into a single claim, with the sources linked so you can check the reading yourself. Where the literature is unsettled, the note says so instead of picking the tidier answer.
These are slower to write than the field notes. They are meant to hold up in front of a reader who already knows the papers, so the claim is stated narrowly and the sources are linked rather than named. If you want the applied version of the same ideas, playbooks is the category to read.
The Strong Stochastic Parrots Claim Is Dead.
A citable 2026 verdict piece: causal world models, optimization-like in-context learning, and identified circuits all weaken the strongest parrots framing.
Lost in the Middle Is Three Bugs.
A triage tree for RAG engineers: softmax sinks, RoPE decay, and training-distribution bias each create their own version of the same long-context failure.
Compression Is a Noisy Proxy.
A five-paper tokenizer synthesis showing why compression only works as a stand-in when it happens to track semantic boundary alignment.
There Is No Code Mode.
A two-channel explanation for why LLMs differ across code, prose, and structured data without requiring separate hidden modes in the architecture.
Format Is a Hyperparameter.
Why prompt formatting deserves the same rigor as model choice, what it costs in token budget, and how to test it with a real eval loop.