The Misreadings Met Again.
Every plate planted one plausible wrong reading. Here they all come back, shuffled.
Each plate of the course ends its knowledge check with one question of the same form: four statements, three faithful to the plate and one a misreading — the kind of wrong conclusion that is easy to draw from a half-remembered result. This plate draws rounds of five from all of them, across every volume, with the explanation and the source plate after each answer.
Out of context they are harder: the plate that framed each one is not on screen. A misreading recognised here, away from its plate, is one that will not be drawn the next time the result is quoted at work or in a paper.
Below: a round of five misreading questions from anywhere in the course. Answer each, read the explanation, and press “Next round”.
- Recognise misreadings out of context.
- Recall which plate settles each.
- Separate a result from its over-extension.
- Review the whole course.
The shapes misreadings take
Most misreadings in this course fall into a few shapes. A trend extended past its data — scaling laws, time horizons, falling prices. A proxy taken for the target — training reward for capability, a benchmark for a skill, perplexity for usefulness. A mechanism given a story it does not support — attention as explanation, a feature as a concept, a chain of thought as the computation. And an average taken for every case — a speed-up that holds at batch one but not at batch sixty-four.
The remedy is the course’s method, applied each time: ask what was measured, on what, against what baseline, and with what uncertainty; and ask what would have to be true for the stronger reading to hold. The journey map, next, shows which plates’ checks have been passed.
Terms
- Misreading
- A plausible but wrong conclusion from a result.
- Over-extension
- A result applied beyond the conditions it was measured in.
- Proxy
- A measurable stand-in for what is wanted.
- Baseline
- What a result is compared against.