How we grade them

How do you grade something that can talk about anything? Carefully, and on more than one axis.

How do you grade something that can talk about anything? You can’t mark one right answer. So you measure a few different things — and trust none of them completely.

First axis: how surprised is it?

Give the model a chunk of text it never trained on and watch how it reacts. rolls up its surprise into one number: low means “I saw that coming,” high means “I did not.”

It’s just Book 01’s run back through an exponential. Cheap, automatic — and it only fairly compares models that chop text into the same tokens.

But being un-surprised by text isn’t the same as being right. So we also test.

Second axis: can it answer questions?

is a giant multiple-choice exam — 57 subjects, from biology to law. The model doesn’t tick a box; it spreads its belief across A–D, and its pick is whichever option it rates most likely. Score = how many picks match the key.

Try it: land on a question the model gets wrong — notice how confident the wrong pick still looks. Then tick data contamination and watch the bar leap onto the right answer for the wrong reason.

The catch: contamination

If a test question was sitting somewhere in the pretraining web-scrape, the model can recall the answer instead of working it out. That’s — and it quietly inflates every score, because popular benchmarks live on the same public web the model read.

More than one axis

Accuracy isn’t the only thing worth grading. We also weigh efficiency — how much compute and energy an answer costs — and fairness — whether the model serves everyone evenly, or quietly worse for some.

Go deeper: why one number is never enough

A benchmark is a proxy — a cheap stand-in for the thing we actually care about. Perplexity proxies “predicts text well.” MMLU proxies “knows stuff.” Neither proxies “reasons honestly,” “stays safe,” or “is cheap to run.” Optimize hard enough for any single proxy and you start gaming it instead of improving — so good evaluation means triangulating several axes and staying suspicious of any one score. (Scaling up reliably drops perplexity — Book 01’s — but a lower loss doesn’t promise a kinder or more truthful model.)

A high score means it picked the right boxes. Not that it reasoned, told the truth, or is safe to ship.

Next: when they go wrong

We’ve measured what they get right. Next we look squarely at how they go wrong.

Sources · 4

Full bibliography →