Evaluating Claude Sonnet and Opus
Starting guidance is useful; choosing a configuration requires evidence from your workload.
Latest writing
15 pieces · RSSWriting on computing, education, and evidence — for people who teach, learn, and decide.
Published articles
How Transformer models changed
Architecture, training and inference answer different questions.
Inside an AI model: from tokens to reasoning
Follow one generation step, then distinguish architecture, training and reasoning.
Difficulty learning does not, by itself, show inability
Cognitive effort and uncertainty offer possible explanations, not a diagnosis of an individual learner.
Compact controls in Edmetry Focus
Reviewing one settings group before adding more local styling.
When the scores are the same
Similar results can hide very different demands on the people who use them.
Computer science has two parents.
Why studying the subject means learning both what can be computed and how to make computation work.
Debugging as a teachable habit.
Why errors are information, not embarrassment.
Fluent answers are not understanding.
When correct-looking work teaches nothing.
LLM fluency is not reasoning.
Why fluent language from a model is weak evidence of understanding.
Measure before you optimise.
Why the first question in analytics is what counts as evidence.
Productive struggle in learning.
Why some difficulty earns learning — and some shortcuts remove it.
Recursion without mystery.
Base case, recursive step, and when to trust the smaller problem.
Teaching computing beyond syntax.
Why correct programs are not always enough.
Bellman–Ford, explained clearly.
Finding shortest paths when edge weights can be negative.