PROMPTSFew-shot examples that don’t backfire
Models copy examples more faithfully than instructions, including their mistakes. How to pick, order and test few-shot examples that help.
PROMPTSThe "red-team yourself" prompt: self-critique that works, and when it backfires
Asking a model to attack its own draft can catch real errors, or talk it out of a right answer. What the research says, and the variants worth using.
AGENTSDesigning tools that AI agents can actually use
Most agent failures start with an ambiguous tool. Naming, schemas, error messages and permissions that make function calling reliable.
PAPERSSpeculative decoding: faster answers without changing them
A small model drafts, a big model checks several tokens at once, and the output distribution stays exactly the same. How and when it speeds things up.
AGENTSTool-use evals are broken. Here's how to fix them.
Exact-match grading, single runs and perfect mock APIs make tool-use scores look better than reality. A better harness design, step by step.
MODELSContext windows are not memory
A bigger window lets a model read more at once. It does not make it remember, prioritise or forget well. How to design for that difference.
TESTSWhat do reasoning tokens actually cost you? How to measure and control them
Hidden thinking tokens are billed like output and vary wildly by task. A practical audit to find where reasoning pays off and where it is burning money.
PAPERSScaling laws, in plain English
Compute, data and parameters: the three-way trade-off behind every model release, explained without equations you have to memorise.
BENCHMARKSYour 2-point improvement might be noise: confidence intervals for LLM evals
A 2-point jump on 200 questions is usually noise. How to compute error bars, run paired comparisons and size your eval set to detect real gains.