Quanta Magazine · Oct 16, 2026
He Wants AI Benchmarking To Be a Science, Not a Dark Art
Markdown · JSON · Original source
In Quanta Magazine, John Pavlus interviews computer scientist Moritz Hardt about why AI benchmarks—standardized tests used to compare models—can mislead. Hardt argues that a score on one test set is not a dependable measure of a model’s general ability, while rankings between models have often been more stable. He says that signal is now less trustworthy for large language models because their training data are unknown and developers may tune them for familiar benchmarks. His proposed remedy is to prepare models on benchmark-relevant material under a shared protocol before testing them, an idea he compares to athletes training at altitude. He presents this as a possible way to make comparisons fairer, not as a proven solution. Benchmarks can indicate progress, Hardt says, but they do not explain how models work.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
Quanta Magazine · Article
Original source date: Oct 16, 2026 (2026-10-16T13:25:08+00:00).
Date basis: Date stated by the publisher.
Added to Escape Meridian: 2026-10-16T17:01:17.298Z. First observed: 2026-10-16T13:30:52.465Z. Retrieved: 2026-10-16T17:00:33.583Z.
Limits
This is an interview presenting Hardt’s perspective; the source does not establish that his proposed preparation protocol has been tested or validated. Hardt says the observed stability of model rankings is empirical and that there is no comprehensive theory explaining it.
Open question
How could benchmark designers standardize the proposed preparation across models without making the evaluation another test that models can be specifically optimized for?
Source-stated
Hardt argues that a benchmark score can change substantially when the test set changes slightly, so an absolute score should not be read as a general measure of ability.
If you look at a model’s accuracy numbers on one benchmark, and then you change the test set only slightly, you get completely different accuracy numbers.
Source-stated
Hardt says model rankings have often remained more stable across different datasets than individual benchmark scores, but that this pattern is empirical and lacks a comprehensive theory.
And we don’t yet have a good comprehensive theory for why this happens. It’s mostly an empirical phenomenon.
Source-stated
Hardt argues that unknown training data for large language models make benchmark comparisons less reliable, because developers can prepare models for particular tests.
In fact, model builders can strategically assemble data mixes that best promote benchmark performance.
Source-stated
Hardt proposes giving each model some benchmark-relevant preparation before testing, so models face a more equal evaluation; the interview does not establish that this approach works.
we can train them for a little bit on benchmark-relevant data, just to make sure that we’re leveling the playing field somewhat.
Source-stated
Hardt says benchmarking may contribute to scientific knowledge, but it does not explain how AI models work.
We don’t know how these models do it, and benchmarking won’t tell us.
Source and contributor context
Quanta Magazine · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Source-associated media was found; no additional asset was selected for this reading. The original source remains available.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Quanta Magazine
Quanta Magazine explains developments in mathematics, physics, biology and computer science through original reporting.
Editorial provenance
Version 1. Publication ID: 419c38df-aad4-4323-822e-227fa662b2b8.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-10-16T17:01:17.298Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: useful_perspective.
Observation: 0198027d7991cbf42eb2a788580840a984a409016645240db397419d39420ccf. Source text hash: 9ba46a7e591ab32e5616b210bf8b2334047d3701c41a8bc2683274ef679cef6f.
Revision and source-check history
Version 1 · 2026-10-16T17:01:17.298Z
He Wants AI Benchmarking To Be a Science, Not a Dark Art
In Quanta Magazine, John Pavlus interviews computer scientist Moritz Hardt about why AI benchmarks—standardized tests used to compare models—can mislead. Hardt argues that a score on one test set is not a dependable measure of a model’s general ability, while rankings between models have often been more stable. He says that signal is now less trustworthy for large language models because their training data are unknown and developers may tune them for familiar benchmarks. His proposed remedy is to prepare models on benchmark-relevant material under a shared protocol before testing them, an idea he compares to athletes training at altitude. He presents this as a possible way to make comparisons fairer, not as a proven solution. Benchmarks can indicate progress, Hardt says, but they do not explain how models work.
Original source
Quanta Magazine · Article
Original source date: Oct 16, 2026 (2026-10-16T13:25:08+00:00).
Date basis: Date stated by the publisher.
Added to Escape Meridian: 2026-10-16T17:01:17.298Z. First observed: 2026-10-16T13:30:52.465Z. Retrieved: 2026-10-16T17:00:33.583Z.
Limits
This is an interview presenting Hardt’s perspective; the source does not establish that his proposed preparation protocol has been tested or validated. Hardt says the observed stability of model rankings is empirical and that there is no comprehensive theory explaining it.
Open question
How could benchmark designers standardize the proposed preparation across models without making the evaluation another test that models can be specifically optimized for?
Source-stated
Hardt argues that a benchmark score can change substantially when the test set changes slightly, so an absolute score should not be read as a general measure of ability.
If you look at a model’s accuracy numbers on one benchmark, and then you change the test set only slightly, you get completely different accuracy numbers.
Source-stated
Hardt says model rankings have often remained more stable across different datasets than individual benchmark scores, but that this pattern is empirical and lacks a comprehensive theory.
And we don’t yet have a good comprehensive theory for why this happens. It’s mostly an empirical phenomenon.
Source-stated
Hardt argues that unknown training data for large language models make benchmark comparisons less reliable, because developers can prepare models for particular tests.
In fact, model builders can strategically assemble data mixes that best promote benchmark performance.
Source-stated
Hardt proposes giving each model some benchmark-relevant preparation before testing, so models face a more equal evaluation; the interview does not establish that this approach works.
we can train them for a little bit on benchmark-relevant data, just to make sure that we’re leveling the playing field somewhat.
Source-stated
Hardt says benchmarking may contribute to scientific knowledge, but it does not explain how AI models work.
We don’t know how these models do it, and benchmarking won’t tell us.
Source and contributor context
Quanta Magazine · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Source-associated media was found; no additional asset was selected for this reading. The original source remains available.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Quanta Magazine
Quanta Magazine explains developments in mathematics, physics, biology and computer science through original reporting.
Editorial provenance
Version 1. Publication ID: 419c38df-aad4-4323-822e-227fa662b2b8.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-10-16T17:01:17.298Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: useful_perspective.
Observation: 0198027d7991cbf42eb2a788580840a984a409016645240db397419d39420ccf. Source text hash: 9ba46a7e591ab32e5616b210bf8b2334047d3701c41a8bc2683274ef679cef6f.
passed · 2026-10-16T17:01:17.298Z · anthropic/claude-opus-5.5
Publication: 419c38df-aad4-4323-822e-227fa662b2b8
attribution — pass: The draft correctly presents this as a Quanta Magazine interview by John Pavlus with Moritz Hardt. It attributes the arguments about benchmarks, rankings, LLM training data and altitude-style preparation to Hardt.
Quanta spoke to Hardt about what went wrong, and right, with benchmarks, and where to take them next.
summary — pass: The title matches the source. The summary defines benchmarks in plain language and identifies the actors. It accurately conveys Hardt's three main points: absolute scores are unreliable, rankings are historically stable, and unknown LLM training data is a problem. It also covers his proposed shared preparation protocol and the limits of what benchmarks explain. The question follows from the proposal without leading the reader.
You would have a test set, and you would also have some kind of protocol for preparing each of the models before you test them — that’s the altitude training.
date — pass: The originalPublished value of 2026-10-16 matches the article byline date. Nothing in the source indicates an earlier first publication.
By John Pavlus October 16, 2026
support — pass: Each claim is grounded in Hardt's stated views, and the draft frames them as his arguments. The draft is appropriately careful: it notes the remedy is unvalidated and that the stability of rankings is empirical. No unsupported benefits or costs are implied.
If, instead of absolute numbers, you look at rankings — relative comparisons between models — those rankings tend to be very stable across different datasets.
media — pass: The selected media is a verified direct link to the original Quanta article, which gives readers the easiest access to the source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.