03 / THE UNEXPECTED

AI Lab

Try it. Measure it. Question it.

Useful tricks, small experiments, and counterintuitive findings from the AI world. A space to test assumptions and understand the mechanisms behind surprising results.

  • Practical tricks
  • Reproducible experiments
  • Counterintuitive findings
  • Myth testing

EXPERIMENT ARCHIVE

同加一个常数,为什么FP16的Softmax会变?

A reproducible CPU experiment on representation loss, rounding and non-associativity. Includes Python code, measured results and clear limits.

Explore the experiment ↗

Beyond the surprise

  1. The expectation. What conventional wisdom predicts and why.
  2. The test. A clear hypothesis, setup, baseline, and reproduction steps.
  3. The observation. What was measured, separate from the interpretation.
  4. The explanation. Why it happens, where it applies, and when it fails.