Skip to the content ForoHackers
Log inJoin

overfit_otter

Joined 23 May 2025

Posts in the forum

  1. Replied in Grover's algorithm: a quadratic speed-up, not magic

    The oracle-cost caveat is the one that gets dropped in headlines. "Quadratic speed-up" quietly assumes `f` is a black box with unit cost. In real problems the box is a circuit, and I would want a bas…

  2. Replied in A qubit is not "both 0 and 1": what the popular line gets wrong

    Statistician hat on: "the amplitudes can cancel" is the line that made it click for me. Probabilities cannot be negative, so classical randomness can never do that. Is that the cleanest way to say wh…

  3. Replied in You cannot average percentiles (a dashboard bug I keep seeing)

    The statistical version of the same warning: any summary that discards the sample size cannot be combined correctly. A percentile from 10 requests and one from 1000 requests are not equally trustwort…

  4. Replied in Bayes' rule with a burnt bagel

    Excellent example, and the same shape as every "the test is 95% accurate" puzzle: base rate times likelihood. I suspect more people understand Bayes through bagels than through formulas.

  5. Replied in A benchmark score is a measurement. Where are the error bars?

    Right, resampling indices once and applying them to both is what keeps the pairing. If the two models fail on the same examples, the difference is far less noisy than the two scores suggest. That is …

  6. Started A benchmark score is a measurement. Where are the error bars?

    A familiar table: model A scores 84%, model B scores 86%, B wins. The test set has 500 examples. A quick check. The standard error of an accuracy near 0.85 on 500 examples is `sqrt(0.85 * 0.15 / 500…

  7. Replied in Evaluar un clasificador: por qué la exactitud miente

    Great post. One addition in English, sorry: the ±10 points you mention is the part that gets forgotten most. With 100 positives, a recall of 60% has a 95% interval of roughly 50% to 70%. Two models t…