TensorTau PathToAGI Doing Machine Learning Honestly Level 05

Level 05 · Doing Machine Learning Honestly · Chapter 071

Why One Training Run Is Not Evidence

If model B beats model A once, how do we know the difference was not luck?

1. Training is a random experiment

Suppose five seeds give:

A: 90.1 90.8 89.9 90.4 90.3
B: 90.5 91.2 89.8 90.6 90.0

Choosing one favourable pair could make B look clearly better. Looking across runs reveals substantial overlap.

2. Report a distribution, not a trophy score

At minimum report the mean and a measure of spread such as standard deviation:

xˉ=1n∑ixi,\bar x=\frac1n\sum_i x_i,

s=1n−1∑i(xi−xˉ)2.s=\sqrt{\frac{1}{n-1}\sum_i(x_i-\bar x)^2}.

The best seed answers “what happened once?” Mean and spread answer “what typically happens and how unstable is it?”

3. Pair experiments when possible

Use the same data splits/seeds for A and B so nuisance randomness is shared. Then analyze per-seed differences. Paired comparisons can be much more sensitive than comparing unrelated runs.

4. Compute budget is part of fairness

A method tried across 100 seeds and hyperparameters has more opportunities to get lucky than one tried twice. Fair comparisons disclose search budget.

What to Remember

One run is an anecdote. Repeated runs transform randomness from a hidden nuisance into something we can measure.

Next → Confidence Intervals: how can we express the uncertainty around an estimated mean rather than only its variability?

Six ways to work the chapter, climbing from explaining it in your own words to investigating it like a researcher. Reading is not enough — this is where the chapter lands.

  1. Explain: Why is best-of-N reporting misleading?
  2. Calculate: Compute mean and sample SD for five run scores.
  3. Predict: When would paired seeds help most?
  4. Break: Cherry-pick the best seed from one model and compare with another model’s average.
  5. Build: Run a small model over at least ten seeds and summarize results.
  6. Investigate: Estimate how often a single-run comparison reverses the multi-run conclusion.