{"id":"c48b97604606a345359a32dfe7cdae73fa43ee886e6e00f594af75440c5c76ed","title":"Quanta Magazine: Moritz Hardt wants AI benchmarking to be a science, not a dark art","argument":"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 fixed collection of test cases 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 less trustworthy for the models discussed in the interview 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. How could benchmark designers standardize the proposed preparation without making the evaluation another test that models can be specifically optimized for?","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","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.","speaker":{"name":"Quanta Magazine","roleAtSource":"Publisher"},"source":{"title":"He Wants AI Benchmarking To Be a Science, Not a Dark Art | Quanta Magazine","publisher":"Quanta Magazine","format":"Article","url":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","publishedAt":"2026-10-16T13:25:08+00:00","dateBasis":"Date stated by the publisher."},"evidence":[{"text":"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.","excerpt":"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.","url":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","relationship":"Source-stated"},{"text":"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.","excerpt":"And we don’t yet have a good comprehensive theory for why this happens. It’s mostly an empirical phenomenon.","url":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","relationship":"Source-stated"},{"text":"Hardt argues that unknown training data for large language models make benchmark comparisons less reliable, because developers can prepare models for particular tests.","excerpt":"In fact, model builders can strategically assemble data mixes that best promote benchmark performance.","url":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","relationship":"Source-stated"},{"text":"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.","excerpt":"we can train them for a little bit on benchmark-relevant data, just to make sure that we’re leveling the playing field somewhat.","url":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","relationship":"Source-stated"},{"text":"Hardt says benchmarking may contribute to scientific knowledge, but it does not explain how AI models work.","excerpt":"We don’t know how these models do it, and benchmarking won’t tell us.","url":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","relationship":"Source-stated"}],"sourceMedia":{"status":"found","sourceUrl":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","checkedAt":"2026-10-21T09:15:59.533Z","items":[{"type":"image","url":"https://www.quantamagazine.org/wp-content/uploads/2026/10/Moritz-Hardt-QA-cr.Jerrit-Peinelt-Lede-1536x864.webp","sourceUrl":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","attribution":"Quanta Magazine","basis":"publisher-embed","availability":"verified","label":"The computer scientist Moritz Hardt calls benchmarks “the mistake that made machine learning.” He’s studied where they went wrong, and why they’ve managed to d…","verifiedAt":"2026-10-21T09:15:59.806Z"}],"note":"Media declared by the original publisher; availability checked separately from the editorial reading."},"assumptions":[],"whatWouldChangeView":{"attribution":"Open question","text":"How could benchmark designers standardize the proposed preparation across models without making the evaluation another test that models can be specifically optimized for?"},"editorialComparisons":[],"live":{"title":"Quanta Magazine: Moritz Hardt wants AI benchmarking to be a science, not a dark art","summary":"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 fixed collection of test cases 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 less trustworthy for the models discussed in the interview 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. How could benchmark designers standardize the proposed preparation without making the evaluation another test that models can be specifically optimized for?","question":"How could benchmark designers standardize the proposed preparation across models without making the evaluation another test that models can be specifically optimized for?","claims":[{"text":"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.","excerpt":"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.","support":"source-stated"},{"text":"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.","excerpt":"And we don’t yet have a good comprehensive theory for why this happens. It’s mostly an empirical phenomenon.","support":"source-stated"},{"text":"Hardt argues that unknown training data for large language models make benchmark comparisons less reliable, because developers can prepare models for particular tests.","excerpt":"In fact, model builders can strategically assemble data mixes that best promote benchmark performance.","support":"source-stated"},{"text":"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.","excerpt":"we can train them for a little bit on benchmark-relevant data, just to make sure that we’re leveling the playing field somewhat.","support":"source-stated"},{"text":"Hardt says benchmarking may contribute to scientific knowledge, but it does not explain how AI models work.","excerpt":"We don’t know how these models do it, and benchmarking won’t tell us.","support":"source-stated"}],"limitations":["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."],"mediaId":"912c74a65d82a0c87012a9f5a11897558c74872f38784fd5990326adf3ed195e","correctionReason":"The title and summary wording were revised to attribute the proposal to Hardt and clarify the interview’s scope, replacing “test set” with “fixed collection of test cases” and narrowing the statement about trustworthiness.","id":"c48b97604606a345359a32dfe7cdae73fa43ee886e6e00f594af75440c5c76ed","publicationId":"5a968895-02fd-4393-80bc-c3e10412782e","version":2,"sourceUrl":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","sourceTitle":"He Wants AI Benchmarking To Be a Science, Not a Dark Art | Quanta Magazine","publisher":"Quanta Magazine","publishedAt":"2026-10-16T13:25:08+00:00","originalPublished":{"value":"2026-10-16T13:25:08+00:00","precision":"instant","basis":"publisher_page"},"firstObservedAt":"2026-10-16T13:30:52.465Z","createdAt":"2026-10-21T09:16:33.216Z","historical":true,"entityIds":["quanta-magazine"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"912c74a65d82a0c87012a9f5a11897558c74872f38784fd5990326adf3ed195e","kind":"article","url":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","sourceUrl":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-10-21T09:15:59.806Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."},{"id":"f44386c972671b66381704eeab8c3a05f6f3499ccf7bdea0135d93ff77c13e0e","kind":"image","url":"https://www.quantamagazine.org/wp-content/uploads/2026/10/Moritz-Hardt-QA-cr.Jerrit-Peinelt-Lede-1536x864.webp","sourceUrl":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","label":"The computer scientist Moritz Hardt calls benchmarks “the mistake that made machine learning.” He’s ","provenance":"publisher_page","verification":{"status":"verified","checkedAt":"2026-10-21T09:15:59.806Z","method":"publisher_association_and_head","contentType":"image/webp","cacheUntil":1792574159806},"recommendationReason":"Observed on the acquired publisher source; availability checked."}],"sourceMedia":{"status":"found","sourceUrl":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","checkedAt":"2026-10-21T09:15:59.533Z","items":[{"type":"image","url":"https://www.quantamagazine.org/wp-content/uploads/2026/10/Moritz-Hardt-QA-cr.Jerrit-Peinelt-Lede-1536x864.webp","sourceUrl":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","attribution":"Quanta Magazine","basis":"publisher-embed","availability":"verified","label":"The computer scientist Moritz Hardt calls benchmarks “the mistake that made machine learning.” He’s studied where they went wrong, and why they’ve managed to d…","verifiedAt":"2026-10-21T09:15:59.806Z"}],"note":"Media declared by the original publisher; availability checked separately from the editorial reading."},"editorialDecision":{"outcome":"accept","reasonCodes":["useful_perspective"],"rubricVersion":"stream-admission-v5","mediaId":"912c74a65d82a0c87012a9f5a11897558c74872f38784fd5990326adf3ed195e","confidence":0.96,"provider":"vercel","model":"typesafe-ai/jev","inputHash":"6a37240c860fcd0f33a7bf338feac3ed4f683bdcb03f6b552e42524e151496a2","decidedAt":"2026-10-21T09:16:00.607Z"},"provenance":{"textHash":"9ba46a7e591ab32e5616b210bf8b2334047d3701c41a8bc2683274ef679cef6f","observationId":"0198027d7991cbf42eb2a788580840a984a409016645240db397419d39420ccf","retrievedAt":"2026-10-21T09:15:59.533Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"2f5439a9-e872-4c45-8c30-c06815b30431"},"reviewer":{"model":"gpt-6-luna","runId":"733e166f-046a-41aa-8be7-f44a3ca8ee4c"}},"audit":{"status":"passed","reviewedAt":"2026-10-21T09:16:33.216Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[],"authorRole":"article-author","hosts":[],"status":"missing","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[],"participationCoverage":{"total":0,"returned":0,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","snapshotId":"fcdc033cd465885c6bb01e46991772d9de91eca02fee1302445596aa0b9ced64","contentSha256":"86dffe298cbbbb44dca0a03a69ea495931b97c350fca384c681b18f41cac2670","textHash":"9ba46a7e591ab32e5616b210bf8b2334047d3701c41a8bc2683274ef679cef6f","observedAt":"2026-10-21T09:15:59.533Z","basis":"publisher-page-metadata","sourceLinks":31,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/c48b97604606a345359a32dfe7cdae73fa43ee886e6e00f594af75440c5c76ed/source"}},"canonicalUrl":"https://escapemeridian.com/readings/quanta-magazine-moritz-hardt-wants-ai-benchmarking-to-be-a-c48b9760","formats":{"html":"/readings/quanta-magazine-moritz-hardt-wants-ai-benchmarking-to-be-a-c48b9760","markdown":"/readings/quanta-magazine-moritz-hardt-wants-ai-benchmarking-to-be-a-c48b9760.md","json":"/readings/quanta-magazine-moritz-hardt-wants-ai-benchmarking-to-be-a-c48b9760.json"},"attribution":"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.","readerContext":{"version":"reader-context-v1","attribution":"Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.","terms":[{"id":"term-benchmark","name":"Benchmark","explanation":"A defined set of tests used to compare systems. Results depend on the tasks and measures selected.","referenceUrl":"https://arxiv.org/abs/2211.09110"}],"entities":[{"id":"quanta-magazine","name":"Quanta Magazine","kind":"organization","context":"Quanta Magazine explains developments in mathematics, physics, biology and computer science through original reporting.","referenceUrls":["https://www.quantamagazine.org/"],"path":"/entities/quanta-magazine","json":"/entities/quanta-magazine.json","markdown":"/entities/quanta-magazine.md"}]},"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"c48b97604606a345359a32dfe7cdae73fa43ee886e6e00f594af75440c5c76ed","revision":"5a968895-02fd-4393-80bc-c3e10412782e","revisionKind":"publication-id","title":"Quanta Magazine: Moritz Hardt wants AI benchmarking to be a science, not a dark art","url":"https://escapemeridian.com/readings/quanta-magazine-moritz-hardt-wants-ai-benchmarking-to-be-a-c48b9760"},"source":{"type":"source","id":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","revision":"9ba46a7e591ab32e5616b210bf8b2334047d3701c41a8bc2683274ef679cef6f","revisionKind":"sha256-source-text","title":"He Wants AI Benchmarking To Be a Science, Not a Dark Art | Quanta Magazine","url":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/"},"status":"available","items":[],"next":null,"scope":"Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter."},"revisions":[{"id":"c48b97604606a345359a32dfe7cdae73fa43ee886e6e00f594af75440c5c76ed","title":"Quanta Magazine: Moritz Hardt wants AI benchmarking to be a science, not a dark art","argument":"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 fixed collection of test cases 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 less trustworthy for the models discussed in the interview 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. How could benchmark designers standardize the proposed preparation without making the evaluation another test that models can be specifically optimized for?","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","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.","speaker":{"name":"Quanta Magazine","roleAtSource":"Publisher"},"source":{"title":"He Wants AI Benchmarking To Be a Science, Not a Dark Art | Quanta Magazine","publisher":"Quanta Magazine","format":"Article","url":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","publishedAt":"2026-10-16T13:25:08+00:00","dateBasis":"Date stated by the publisher."},"evidence":[{"text":"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.","excerpt":"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.","url":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","relationship":"Source-stated"},{"text":"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.","excerpt":"And we don’t yet have a good comprehensive theory for why this happens. It’s mostly an empirical phenomenon.","url":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","relationship":"Source-stated"},{"text":"Hardt argues that unknown training data for large language models make benchmark comparisons less reliable, because developers can prepare models for particular tests.","excerpt":"In fact, model builders can strategically assemble data mixes that best promote benchmark performance.","url":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","relationship":"Source-stated"},{"text":"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.","excerpt":"we can train them for a little bit on benchmark-relevant data, just to make sure that we’re leveling the playing field somewhat.","url":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","relationship":"Source-stated"},{"text":"Hardt says benchmarking may contribute to scientific knowledge, but it does not explain how AI models work.","excerpt":"We don’t know how these models do it, and benchmarking won’t tell us.","url":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","relationship":"Source-stated"}],"sourceMedia":{"status":"found","sourceUrl":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","checkedAt":"2026-10-21T09:15:59.533Z","items":[{"type":"image","url":"https://www.quantamagazine.org/wp-content/uploads/2026/10/Moritz-Hardt-QA-cr.Jerrit-Peinelt-Lede-1536x864.webp","sourceUrl":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","attribution":"Quanta Magazine","basis":"publisher-embed","availability":"verified","label":"The computer scientist Moritz Hardt calls benchmarks “the mistake that made machine learning.” He’s studied where they went wrong, and why they’ve managed to d…","verifiedAt":"2026-10-21T09:15:59.806Z"}],"note":"Media declared by the original publisher; availability checked separately from the editorial reading."},"assumptions":[],"whatWouldChangeView":{"attribution":"Open question","text":"How could benchmark designers standardize the proposed preparation across models without making the evaluation another test that models can be specifically optimized for?"},"editorialComparisons":[],"live":{"title":"Quanta Magazine: Moritz Hardt wants AI benchmarking to be a science, not a dark art","summary":"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 fixed collection of test cases 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 less trustworthy for the models discussed in the interview 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. How could benchmark designers standardize the proposed preparation without making the evaluation another test that models can be specifically optimized for?","question":"How could benchmark designers standardize the proposed preparation across models without making the evaluation another test that models can be specifically optimized for?","claims":[{"text":"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.","excerpt":"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.","support":"source-stated"},{"text":"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.","excerpt":"And we don’t yet have a good comprehensive theory for why this happens. It’s mostly an empirical phenomenon.","support":"source-stated"},{"text":"Hardt argues that unknown training data for large language models make benchmark comparisons less reliable, because developers can prepare models for particular tests.","excerpt":"In fact, model builders can strategically assemble data mixes that best promote benchmark performance.","support":"source-stated"},{"text":"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.","excerpt":"we can train them for a little bit on benchmark-relevant data, just to make sure that we’re leveling the playing field somewhat.","support":"source-stated"},{"text":"Hardt says benchmarking may contribute to scientific knowledge, but it does not explain how AI models work.","excerpt":"We don’t know how these models do it, and benchmarking won’t tell us.","support":"source-stated"}],"limitations":["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."],"mediaId":"912c74a65d82a0c87012a9f5a11897558c74872f38784fd5990326adf3ed195e","correctionReason":"The title and summary wording were revised to attribute the proposal to Hardt and clarify the interview’s scope, replacing “test set” with “fixed collection of test cases” and narrowing the statement about trustworthiness.","id":"c48b97604606a345359a32dfe7cdae73fa43ee886e6e00f594af75440c5c76ed","publicationId":"5a968895-02fd-4393-80bc-c3e10412782e","version":2,"sourceUrl":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","sourceTitle":"He Wants AI Benchmarking To Be a Science, Not a Dark Art | Quanta Magazine","publisher":"Quanta Magazine","publishedAt":"2026-10-16T13:25:08+00:00","originalPublished":{"value":"2026-10-16T13:25:08+00:00","precision":"instant","basis":"publisher_page"},"firstObservedAt":"2026-10-16T13:30:52.465Z","createdAt":"2026-10-21T09:16:33.216Z","historical":true,"entityIds":["quanta-magazine"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"912c74a65d82a0c87012a9f5a11897558c74872f38784fd5990326adf3ed195e","kind":"article","url":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","sourceUrl":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-10-21T09:15:59.806Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."},{"id":"f44386c972671b66381704eeab8c3a05f6f3499ccf7bdea0135d93ff77c13e0e","kind":"image","url":"https://www.quantamagazine.org/wp-content/uploads/2026/10/Moritz-Hardt-QA-cr.Jerrit-Peinelt-Lede-1536x864.webp","sourceUrl":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","label":"The computer scientist Moritz Hardt calls benchmarks “the mistake that made machine learning.” He’s ","provenance":"publisher_page","verification":{"status":"verified","checkedAt":"2026-10-21T09:15:59.806Z","method":"publisher_association_and_head","contentType":"image/webp","cacheUntil":1792574159806},"recommendationReason":"Observed on the acquired publisher source; availability checked."}],"sourceMedia":{"status":"found","sourceUrl":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","checkedAt":"2026-10-21T09:15:59.533Z","items":[{"type":"image","url":"https://www.quantamagazine.org/wp-content/uploads/2026/10/Moritz-Hardt-QA-cr.Jerrit-Peinelt-Lede-1536x864.webp","sourceUrl":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","attribution":"Quanta Magazine","basis":"publisher-embed","availability":"verified","label":"The computer scientist Moritz Hardt calls benchmarks “the mistake that made machine learning.” He’s studied where they went wrong, and why they’ve managed to d…","verifiedAt":"2026-10-21T09:15:59.806Z"}],"note":"Media declared by the original publisher; availability checked separately from the editorial reading."},"editorialDecision":{"outcome":"accept","reasonCodes":["useful_perspective"],"rubricVersion":"stream-admission-v5","mediaId":"912c74a65d82a0c87012a9f5a11897558c74872f38784fd5990326adf3ed195e","confidence":0.96,"provider":"vercel","model":"typesafe-ai/jev","inputHash":"6a37240c860fcd0f33a7bf338feac3ed4f683bdcb03f6b552e42524e151496a2","decidedAt":"2026-10-21T09:16:00.607Z"},"provenance":{"textHash":"9ba46a7e591ab32e5616b210bf8b2334047d3701c41a8bc2683274ef679cef6f","observationId":"0198027d7991cbf42eb2a788580840a984a409016645240db397419d39420ccf","retrievedAt":"2026-10-21T09:15:59.533Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"2f5439a9-e872-4c45-8c30-c06815b30431"},"reviewer":{"model":"gpt-6-luna","runId":"733e166f-046a-41aa-8be7-f44a3ca8ee4c"}},"audit":{"status":"passed","reviewedAt":"2026-10-21T09:16:33.216Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[],"authorRole":"article-author","hosts":[],"status":"missing","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[],"participationCoverage":{"total":0,"returned":0,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","snapshotId":"fcdc033cd465885c6bb01e46991772d9de91eca02fee1302445596aa0b9ced64","contentSha256":"86dffe298cbbbb44dca0a03a69ea495931b97c350fca384c681b18f41cac2670","textHash":"9ba46a7e591ab32e5616b210bf8b2334047d3701c41a8bc2683274ef679cef6f","observedAt":"2026-10-21T09:15:59.533Z","basis":"publisher-page-metadata","sourceLinks":31,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/c48b97604606a345359a32dfe7cdae73fa43ee886e6e00f594af75440c5c76ed/source"}}},{"id":"c48b97604606a345359a32dfe7cdae73fa43ee886e6e00f594af75440c5c76ed","title":"He Wants AI Benchmarking To Be a Science, Not a Dark Art","argument":"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.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","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.","speaker":{"name":"Quanta Magazine","roleAtSource":"Publisher"},"source":{"title":"He Wants AI Benchmarking To Be a Science, Not a Dark Art | Quanta Magazine","publisher":"Quanta Magazine","format":"Article","url":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","publishedAt":"2026-10-16T13:25:08+00:00","dateBasis":"Date stated by the publisher."},"evidence":[{"text":"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.","excerpt":"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.","url":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","relationship":"Source-stated"},{"text":"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.","excerpt":"And we don’t yet have a good comprehensive theory for why this happens. It’s mostly an empirical phenomenon.","url":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","relationship":"Source-stated"},{"text":"Hardt argues that unknown training data for large language models make benchmark comparisons less reliable, because developers can prepare models for particular tests.","excerpt":"In fact, model builders can strategically assemble data mixes that best promote benchmark performance.","url":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","relationship":"Source-stated"},{"text":"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.","excerpt":"we can train them for a little bit on benchmark-relevant data, just to make sure that we’re leveling the playing field somewhat.","url":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","relationship":"Source-stated"},{"text":"Hardt says benchmarking may contribute to scientific knowledge, but it does not explain how AI models work.","excerpt":"We don’t know how these models do it, and benchmarking won’t tell us.","url":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","relationship":"Source-stated"}],"sourceMedia":{"status":"found","sourceUrl":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","checkedAt":"2026-10-16T17:00:33.583Z","items":[{"type":"image","url":"https://www.quantamagazine.org/wp-content/uploads/2026/10/Moritz-Hardt-QA-cr.Jerrit-Peinelt-Social.jpg","sourceUrl":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","attribution":"Quanta Magazine","basis":"publisher-metadata","availability":"verified","label":"Man sitting on a green couch.","verifiedAt":"2026-10-16T17:00:33.987Z"}],"note":"Media declared by the original publisher; availability checked separately from the editorial reading."},"assumptions":[],"whatWouldChangeView":{"attribution":"Open question","text":"How could benchmark designers standardize the proposed preparation across models without making the evaluation another test that models can be specifically optimized for?"},"editorialComparisons":[],"live":{"title":"He Wants AI Benchmarking To Be a Science, Not a Dark Art","summary":"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.","question":"How could benchmark designers standardize the proposed preparation across models without making the evaluation another test that models can be specifically optimized for?","claims":[{"text":"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.","excerpt":"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.","support":"source-stated"},{"text":"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.","excerpt":"And we don’t yet have a good comprehensive theory for why this happens. It’s mostly an empirical phenomenon.","support":"source-stated"},{"text":"Hardt argues that unknown training data for large language models make benchmark comparisons less reliable, because developers can prepare models for particular tests.","excerpt":"In fact, model builders can strategically assemble data mixes that best promote benchmark performance.","support":"source-stated"},{"text":"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.","excerpt":"we can train them for a little bit on benchmark-relevant data, just to make sure that we’re leveling the playing field somewhat.","support":"source-stated"},{"text":"Hardt says benchmarking may contribute to scientific knowledge, but it does not explain how AI models work.","excerpt":"We don’t know how these models do it, and benchmarking won’t tell us.","support":"source-stated"}],"limitations":["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."],"mediaId":"912c74a65d82a0c87012a9f5a11897558c74872f38784fd5990326adf3ed195e","correctionReason":null,"id":"c48b97604606a345359a32dfe7cdae73fa43ee886e6e00f594af75440c5c76ed","publicationId":"419c38df-aad4-4323-822e-227fa662b2b8","version":1,"sourceUrl":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","sourceTitle":"He Wants AI Benchmarking To Be a Science, Not a Dark Art | Quanta Magazine","publisher":"Quanta Magazine","publishedAt":"2026-10-16T13:25:08+00:00","originalPublished":{"value":"2026-10-16T13:25:08+00:00","precision":"instant","basis":"publisher_page"},"firstObservedAt":"2026-10-16T13:30:52.465Z","createdAt":"2026-10-16T17:01:17.298Z","historical":false,"entityIds":["quanta-magazine"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"912c74a65d82a0c87012a9f5a11897558c74872f38784fd5990326adf3ed195e","kind":"article","url":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","sourceUrl":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-10-16T17:00:33.987Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."},{"id":"6163120bad61fbb9b74d40217762c69ee0cc3c4cd0b3276bfcbd9d123cc4c498","kind":"image","url":"https://www.quantamagazine.org/wp-content/uploads/2026/10/Moritz-Hardt-QA-cr.Jerrit-Peinelt-Social.jpg","sourceUrl":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","label":"Man sitting on a green couch.","provenance":"publisher_page","verification":{"status":"verified","checkedAt":"2026-10-16T17:00:33.987Z","method":"publisher_association_and_head","contentType":"image/jpeg","cacheUntil":1792170033987},"recommendationReason":"Observed on the acquired publisher source; availability checked."}],"sourceMedia":{"status":"found","sourceUrl":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","checkedAt":"2026-10-16T17:00:33.583Z","items":[{"type":"image","url":"https://www.quantamagazine.org/wp-content/uploads/2026/10/Moritz-Hardt-QA-cr.Jerrit-Peinelt-Social.jpg","sourceUrl":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","attribution":"Quanta Magazine","basis":"publisher-metadata","availability":"verified","label":"Man sitting on a green couch.","verifiedAt":"2026-10-16T17:00:33.987Z"}],"note":"Media declared by the original publisher; availability checked separately from the editorial reading."},"editorialDecision":{"outcome":"accept","reasonCodes":["useful_perspective"],"rubricVersion":"stream-admission-v5","mediaId":"912c74a65d82a0c87012a9f5a11897558c74872f38784fd5990326adf3ed195e","confidence":0.97,"provider":"vercel","model":"typesafe-ai/jev","inputHash":"b891ad076fb5fd5380c2a6f14d229a99c988e830b019767a5026aafde2415acc","decidedAt":"2026-10-16T17:00:35.898Z"},"provenance":{"textHash":"9ba46a7e591ab32e5616b210bf8b2334047d3701c41a8bc2683274ef679cef6f","observationId":"0198027d7991cbf42eb2a788580840a984a409016645240db397419d39420ccf","retrievedAt":"2026-10-16T17:00:33.583Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"afbc53a0-05f3-4724-8ed8-9db543139e9e"},"reviewer":{"model":"gpt-6-luna","runId":"7b45dfbf-f4d5-4633-8f3e-ed9152bd4a83"}},"audit":{"status":"passed","reviewedAt":"2026-10-16T17:01:17.298Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[],"authorRole":"article-author","hosts":[],"status":"missing","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[],"participationCoverage":{"total":0,"returned":0,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/","snapshotId":"fcdc033cd465885c6bb01e46991772d9de91eca02fee1302445596aa0b9ced64","contentSha256":"86dffe298cbbbb44dca0a03a69ea495931b97c350fca384c681b18f41cac2670","textHash":"9ba46a7e591ab32e5616b210bf8b2334047d3701c41a8bc2683274ef679cef6f","observedAt":"2026-10-21T09:15:59.533Z","basis":"publisher-page-metadata","sourceLinks":31,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/c48b97604606a345359a32dfe7cdae73fa43ee886e6e00f594af75440c5c76ed/source"}}}],"audits":[{"publicationId":"419c38df-aad4-4323-822e-227fa662b2b8","verdict":"passed","findings":[{"dimension":"attribution","status":"pass","explanation":"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.","excerpt":"Quanta spoke to Hardt about what went wrong, and right, with benchmarks, and where to take them next."},{"dimension":"summary","status":"pass","explanation":"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.","excerpt":"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."},{"dimension":"date","status":"pass","explanation":"The originalPublished value of 2026-10-16 matches the article byline date. Nothing in the source indicates an earlier first publication.","excerpt":"By\nJohn Pavlus\nOctober 16, 2026"},{"dimension":"support","status":"pass","explanation":"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.","excerpt":"If, instead of absolute numbers, you look at rankings — relative comparisons between models — those rankings tend to be very stable across different datasets."},{"dimension":"media","status":"pass","explanation":"The selected media is a verified direct link to the original Quanta article, which gives readers the easiest access to the source.","excerpt":null}],"createdAt":"2026-10-16T17:01:17.298Z","model":"anthropic/claude-opus-5.5"},{"publicationId":"5a968895-02fd-4393-80bc-c3e10412782e","verdict":"passed","findings":[{"dimension":"attribution","status":"pass","explanation":"The draft correctly identifies this as a Quanta Magazine interview by John Pavlus with computer scientist Moritz Hardt. It attributes the arguments and proposal to Hardt rather than presenting them as established fact.","excerpt":"Quanta spoke to Hardt about what went wrong, and right, with benchmarks, and where to take them next."},{"dimension":"summary","status":"pass","explanation":"The summary is understandable to a general reader. It defines benchmarks, explains the score-versus-ranking distinction, the concern about unknown training data, and the altitude-training analogy. It notes the remedy is unproven. The question follows directly from the proposed protocol and is not leading.","excerpt":"If everyone optimizes, or “altitude trains,” the absolute numbers will be inflated, but you can restore the validity of the rankings."},{"dimension":"date","status":"pass","explanation":"The originalPublished value of 2026-10-16 matches the article byline date and the feed entry. No earlier publication date is indicated in the source.","excerpt":"By\nJohn Pavlus\nOctober 16, 2026"},{"dimension":"support","status":"pass","explanation":"Each claim is source-stated and paired with a matching excerpt. The limitations accurately reflect Hardt's own caveats about the lack of theory and the untested remedy. No unsupported benefits or impacts are asserted.","excerpt":"But it’s lucky that it worked as well as it did — we had no reason to expect this to be the case."},{"dimension":"media","status":"pass","explanation":"The selected media is the verified direct link to the original Quanta article. This is the most direct way for readers to access the source.","excerpt":null}],"createdAt":"2026-10-21T09:16:33.216Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}