The reading room · Digna Legi
He Wants AI Benchmarking To Be a Science, Not a Dark Art
/100
80–100: high value. 70–79: worth the time. Below 70: below the usual publication threshold.
Evidence-reviewed score based on available publisher text. The interview substantiates the argument presented, but the underlying empirical studies and proposed preparation protocol are not supplied for verification.
Scores reflect one reader’s profile, not an objective quality rating. Best is a separate personal selection.
How scoring works →This brief · about 3 min with detail
Why read this
Hardt argues that unreliable benchmark scores once produced reliable model rankings, but unequal preparation now threatens even that narrower achievement.
AI brief · Checked against source text
The main idea
Moritz Hardt argues that AI benchmarks historically identified better models even when their numerical scores failed to measure dependable real-world ability: rankings often survived changes in test data that substantially changed accuracy. He says this accidental reliability depended on models sharing training data, a condition absent from today’s large language models, whose undisclosed training mixtures can reward preparation for a particular test over broader capability.
Go a little deeper
A public test becomes part of development
The textbook safeguard is to reserve test data and use it only once to compare models. Public leaderboards undermine that separation: builders can repeatedly adjust their systems in response to test performance. Hardt’s distinction between a measuring instrument and a competition matters because competitors actively respond to the rules. The benchmark therefore helps shape the behavior it claims merely to observe.
Standardize preparation, not just the exam
Hardt’s proposed response is to train each model briefly on benchmark-relevant material before testing it. Since evaluators cannot reconstruct undisclosed training histories, they would instead apply a common preparation protocol to reduce unequal familiarity. This accepts optimization as part of evaluation. Its intended achievement is narrower than making scores meaningful: restoring a comparison in which the stronger model is more likely to win.
Selection and explanation are different achievements
A benchmark can help select a better-performing system without revealing how that system works. Hardt acknowledges the scientific cost of this arrangement: repeated victories provide indicators of progress while leaving mechanisms obscure. Yet he resists treating explanation as a prerequisite for useful competition. His position leaves room for benchmarks to contribute to scientific knowledge over time without mistaking the leaderboard itself for that knowledge.
A case from the article
The altitude-training analogy
Hardt recounts a favored American runner losing the 1968 Olympic 1,500-meter race to a Kenyan runner better prepared for Mexico City’s elevation. The response, in his telling, was to permit altitude training for everyone. The analogy locates unfairness in unequal preparation rather than preparation itself: restoring a useful contest may require broadening access to an advantage instead of prohibiting it.
How the case is made
An edited interview develops the argument through historical observations about model rankings, an Olympic analogy, and a proposed evaluation protocol.
Where the idea has limits
Hardt describes ranking stability as an empirical pattern without a comprehensive theory; his proposed preparation protocol remains a possible remedy, not an established solution.
A question to take away · from Digna Legi
When comparing competing systems, how much of the apparent advantage comes from greater capability, and how much from unequal preparation for the comparison?
What the original adds
The interview also traces benchmarking to a 1959 character-recognition competition and uses the CASP protein-folding benchmark to illustrate how competition can advance science before mechanisms are understood.
About this brief
AI-written, then separately checked for source support, useful detail and clarity. The author’s claims and our editorial question are kept separate. The original remains the author’s work. How we select and summarise →
How was this brief?
Rate this summary, separately from the author’s article.
Optional. Saved in this browser; shared only if you allow analytics.
How was the original article?
Rate the author’s original after reading it.
Optional. Saved in this browser; shared only if you allow analytics.
Digna legi. Worth reading.