{"id":"3d8a8e11bcb6233b0cb12713c25795ee0f0c524b4389d44c3c53fd6d4901dda8","title":"MIT Technology Review reports on OpenAI’s supervision research and Google DeepMind’s mathematical discovery","argument":"MIT Technology Review’s December 15, 2023, newsletter reports on two developments involving large language models. Will Douglas Heaven says OpenAI’s superalignment team described a technique in which a less powerful model supervises a more powerful one, and suggests it might be a small step toward humans supervising superhuman machines. The newsletter notes that this work assumes machines will eventually surpass human intelligence, a position many researchers question.\nHeaven also reports that Google DeepMind used FunSearch to find a solution to a famous unsolved problem in pure mathematics. The researchers say it produced verifiable, valuable information that did not previously exist. The newsletter’s point is that large language models might contribute to discovery, despite their reputation for making things up—but only if they are coaxed in the right way and most of their output is discarded. The account does not specify the problem, the verification process, or how broadly the approach might work.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 30, 2026; originally published December 15, 2023. The newsletter does not name the mathematical problem or describe how Google DeepMind verified the reported solution. The newsletter does not provide details of the supervision technique or evidence about how well it generalizes beyond the described work.","speaker":{"name":"MIT Technology Review","roleAtSource":"Publisher"},"source":{"title":"The Download: beyond CRISPR, and OpenAI’s superalignment findings","publisher":"MIT Technology Review","format":"Article","url":"https://www.technologyreview.com/2023/12/15/1085460/the-download-beyond-crispr-and-openais-superalignment-findings/","publishedAt":"2023-12-15T13:10:00Z","dateBasis":"Date from the publisher’s own records, through its API."},"evidence":[{"text":"OpenAI’s superalignment team described a technique for having a less powerful large language model supervise a more powerful one, and suggested it might help explore how humans could supervise superhuman machines.","excerpt":"the team describes a technique that lets a less powerful large language model supervise a more powerful one—and suggests that this might be a small step toward figuring out how humans might supervise superhuman machines.","url":"https://www.technologyreview.com/2023/12/15/1085460/the-download-beyond-crispr-and-openais-superalignment-findings/","relationship":"Source-stated"},{"text":"The newsletter says OpenAI’s team takes machines’ eventual superiority over humans as given, although many researchers question whether machines will match or surpass human intelligence.","excerpt":"While many researchers still question whether machines will ever match human intelligence, let alone outmatch it, OpenAI’s team takes machines’ eventual superiority as given.","url":"https://www.technologyreview.com/2023/12/15/1085460/the-download-beyond-crispr-and-openais-superalignment-findings/","relationship":"Source-stated"},{"text":"Google DeepMind says FunSearch produced a solution to a famous unsolved problem and yielded verifiable, valuable information that did not previously exist.","excerpt":"The researchers say it is the first time a large language model has been used to discover a solution to a long-standing scientific puzzle—producing verifiable and valuable new information that did not previously exist.","url":"https://www.technologyreview.com/2023/12/15/1085460/the-download-beyond-crispr-and-openais-superalignment-findings/","relationship":"Source-stated"},{"text":"The newsletter argues that FunSearch suggests large language models can make discoveries when guided in a particular way and most of their outputs are discarded.","excerpt":"It shows that they can indeed make discoveries—if they are coaxed just so, and if you throw out the majority of what they come up with.","url":"https://www.technologyreview.com/2023/12/15/1085460/the-download-beyond-crispr-and-openais-superalignment-findings/","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.technologyreview.com/2023/12/15/1085460/the-download-beyond-crispr-and-openais-superalignment-findings/","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"What problem did FunSearch solve, and how did Google DeepMind verify its solution?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"MIT Technology Review reports on OpenAI’s supervision research and Google DeepMind’s mathematical discovery","summary":"MIT Technology Review’s December 15, 2023, newsletter reports on two developments involving large language models. Will Douglas Heaven says OpenAI’s superalignment team described a technique in which a less powerful model supervises a more powerful one, and suggests it might be a small step toward humans supervising superhuman machines. The newsletter notes that this work assumes machines will eventually surpass human intelligence, a position many researchers question.\nHeaven also reports that Google DeepMind used FunSearch to find a solution to a famous unsolved problem in pure mathematics. The researchers say it produced verifiable, valuable information that did not previously exist. The newsletter’s point is that large language models might contribute to discovery, despite their reputation for making things up—but only if they are coaxed in the right way and most of their output is discarded. The account does not specify the problem, the verification process, or how broadly the approach might work.","question":"What problem did FunSearch solve, and how did Google DeepMind verify its solution?","claims":[{"text":"OpenAI’s superalignment team described a technique for having a less powerful large language model supervise a more powerful one, and suggested it might help explore how humans could supervise superhuman machines.","excerpt":"the team describes a technique that lets a less powerful large language model supervise a more powerful one—and suggests that this might be a small step toward figuring out how humans might supervise superhuman machines.","support":"source-stated"},{"text":"The newsletter says OpenAI’s team takes machines’ eventual superiority over humans as given, although many researchers question whether machines will match or surpass human intelligence.","excerpt":"While many researchers still question whether machines will ever match human intelligence, let alone outmatch it, OpenAI’s team takes machines’ eventual superiority as given.","support":"source-stated"},{"text":"Google DeepMind says FunSearch produced a solution to a famous unsolved problem and yielded verifiable, valuable information that did not previously exist.","excerpt":"The researchers say it is the first time a large language model has been used to discover a solution to a long-standing scientific puzzle—producing verifiable and valuable new information that did not previously exist.","support":"source-stated"},{"text":"The newsletter argues that FunSearch suggests large language models can make discoveries when guided in a particular way and most of their outputs are discarded.","excerpt":"It shows that they can indeed make discoveries—if they are coaxed just so, and if you throw out the majority of what they come up with.","support":"source-stated"}],"limitations":["Text as retrieved on October 30, 2026; originally published December 15, 2023.","The newsletter does not name the mathematical problem or describe how Google DeepMind verified the reported solution.","The newsletter does not provide details of the supervision technique or evidence about how well it generalizes beyond the described work."],"mediaId":null,"correctionReason":null,"id":"3d8a8e11bcb6233b0cb12713c25795ee0f0c524b4389d44c3c53fd6d4901dda8","publicationId":"78d0fcc5-1f93-6900-e28d-30c729d3da02","version":1,"sourceUrl":"https://www.technologyreview.com/2023/12/15/1085460/the-download-beyond-crispr-and-openais-superalignment-findings/","sourceTitle":"The Download: beyond CRISPR, and OpenAI’s superalignment findings","publisher":"MIT Technology Review","publishedAt":"2023-12-15T13:10:00Z","originalPublished":{"value":"2023-12-15T13:10:00Z","precision":"instant","basis":"backfill:api_published"},"firstObservedAt":"2026-10-30T21:18:35.843Z","createdAt":"2026-11-22T04:18:12.513Z","historical":true,"entityIds":[],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.technologyreview.com/2023/12/15/1085460/the-download-beyond-crispr-and-openais-superalignment-findings/","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"91e7eb330d955622f67783246dd35f12da3e7d240dd4a586ba78f4cdc2184d28","rubricVersion":"unrecorded","provider":"anthropic","model":"claude-opus-5-5","reasonCodes":["bridge_admitted_by_model","second_opinion_categorical"],"confidence":1,"decidedAt":"2026-10-30T21:18:32.893Z"},"provenance":{"textHash":"ee8841c1cb90f4b8ccf3b2858699b12c1e29121178ffa0c694b9a19c4bfa6f94","observationId":"78df2c855f7f9003fdc7622442d1f9f635ab58c5d81b293bde6caab3945c865c","retrievedAt":"2026-10-30T21:57:48.300Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"f0797528-0e3e-6a56-8ab2-2ef39c144c6f"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"a97b249d-aa96-1c63-f345-0643d5f44a9a"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-22T04:14:36.939Z","model":"anthropic/claude-opus-5.5"}},"canonicalUrl":"https://escapemeridian.com/readings/mit-technology-review-reports-on-openai-s-supervision-research-3d8a8e11","formats":{"html":"/readings/mit-technology-review-reports-on-openai-s-supervision-research-3d8a8e11","markdown":"/readings/mit-technology-review-reports-on-openai-s-supervision-research-3d8a8e11.md","json":"/readings/mit-technology-review-reports-on-openai-s-supervision-research-3d8a8e11.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.","maintainedGuideLinks":[],"conditions":[],"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-llm","name":"Large language model","explanation":"A language model with many learned numerical settings, used to process and generate language.","referenceUrl":"https://developers.google.com/machine-learning/glossary#large-language-model"}],"entities":[{"id":"openai","name":"OpenAI","kind":"organization","context":"OpenAI develops AI models and products and publishes research, release announcements and evaluations.","referenceUrls":["https://openai.com/news/"],"path":"/entities/openai","json":"/entities/openai.json","markdown":"/entities/openai.md"},{"id":"google-deepmind","name":"Google DeepMind","kind":"organization","context":"Google DeepMind develops AI systems for general tasks and scientific research, including AlphaFold.","referenceUrls":["https://deepmind.google/blog/","https://deepmind.google/science/alphafold/"],"path":"/entities/google-deepmind","json":"/entities/google-deepmind.json","markdown":"/entities/google-deepmind.md"}]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"3d8a8e11bcb6233b0cb12713c25795ee0f0c524b4389d44c3c53fd6d4901dda8","revision":"78d0fcc5-1f93-6900-e28d-30c729d3da02","revisionKind":"publication-id","title":"MIT Technology Review reports on OpenAI’s supervision research and Google DeepMind’s mathematical discovery","url":"https://escapemeridian.com/readings/mit-technology-review-reports-on-openai-s-supervision-research-3d8a8e11"},"source":{"type":"source","id":"https://www.technologyreview.com/2023/12/15/1085460/the-download-beyond-crispr-and-openais-superalignment-findings/","revision":"ee8841c1cb90f4b8ccf3b2858699b12c1e29121178ffa0c694b9a19c4bfa6f94","revisionKind":"sha256-source-text","title":"The Download: beyond CRISPR, and OpenAI’s superalignment findings","url":"https://www.technologyreview.com/2023/12/15/1085460/the-download-beyond-crispr-and-openais-superalignment-findings/"},"status":"available","items":[],"next":null,"scope":"Explicit citations in public readings to this reading. Shared topics are not citations. Other record types need an explicit reference adapter."},"revisions":[{"id":"3d8a8e11bcb6233b0cb12713c25795ee0f0c524b4389d44c3c53fd6d4901dda8","title":"MIT Technology Review reports on OpenAI’s supervision research and Google DeepMind’s mathematical discovery","argument":"MIT Technology Review’s December 15, 2023, newsletter reports on two developments involving large language models. Will Douglas Heaven says OpenAI’s superalignment team described a technique in which a less powerful model supervises a more powerful one, and suggests it might be a small step toward humans supervising superhuman machines. The newsletter notes that this work assumes machines will eventually surpass human intelligence, a position many researchers question.\nHeaven also reports that Google DeepMind used FunSearch to find a solution to a famous unsolved problem in pure mathematics. The researchers say it produced verifiable, valuable information that did not previously exist. The newsletter’s point is that large language models might contribute to discovery, despite their reputation for making things up—but only if they are coaxed in the right way and most of their output is discarded. The account does not specify the problem, the verification process, or how broadly the approach might work.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 30, 2026; originally published December 15, 2023. The newsletter does not name the mathematical problem or describe how Google DeepMind verified the reported solution. The newsletter does not provide details of the supervision technique or evidence about how well it generalizes beyond the described work.","speaker":{"name":"MIT Technology Review","roleAtSource":"Publisher"},"source":{"title":"The Download: beyond CRISPR, and OpenAI’s superalignment findings","publisher":"MIT Technology Review","format":"Article","url":"https://www.technologyreview.com/2023/12/15/1085460/the-download-beyond-crispr-and-openais-superalignment-findings/","publishedAt":"2023-12-15T13:10:00Z","dateBasis":"Date from the publisher’s own records, through its API."},"evidence":[{"text":"OpenAI’s superalignment team described a technique for having a less powerful large language model supervise a more powerful one, and suggested it might help explore how humans could supervise superhuman machines.","excerpt":"the team describes a technique that lets a less powerful large language model supervise a more powerful one—and suggests that this might be a small step toward figuring out how humans might supervise superhuman machines.","url":"https://www.technologyreview.com/2023/12/15/1085460/the-download-beyond-crispr-and-openais-superalignment-findings/","relationship":"Source-stated"},{"text":"The newsletter says OpenAI’s team takes machines’ eventual superiority over humans as given, although many researchers question whether machines will match or surpass human intelligence.","excerpt":"While many researchers still question whether machines will ever match human intelligence, let alone outmatch it, OpenAI’s team takes machines’ eventual superiority as given.","url":"https://www.technologyreview.com/2023/12/15/1085460/the-download-beyond-crispr-and-openais-superalignment-findings/","relationship":"Source-stated"},{"text":"Google DeepMind says FunSearch produced a solution to a famous unsolved problem and yielded verifiable, valuable information that did not previously exist.","excerpt":"The researchers say it is the first time a large language model has been used to discover a solution to a long-standing scientific puzzle—producing verifiable and valuable new information that did not previously exist.","url":"https://www.technologyreview.com/2023/12/15/1085460/the-download-beyond-crispr-and-openais-superalignment-findings/","relationship":"Source-stated"},{"text":"The newsletter argues that FunSearch suggests large language models can make discoveries when guided in a particular way and most of their outputs are discarded.","excerpt":"It shows that they can indeed make discoveries—if they are coaxed just so, and if you throw out the majority of what they come up with.","url":"https://www.technologyreview.com/2023/12/15/1085460/the-download-beyond-crispr-and-openais-superalignment-findings/","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.technologyreview.com/2023/12/15/1085460/the-download-beyond-crispr-and-openais-superalignment-findings/","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"What problem did FunSearch solve, and how did Google DeepMind verify its solution?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"MIT Technology Review reports on OpenAI’s supervision research and Google DeepMind’s mathematical discovery","summary":"MIT Technology Review’s December 15, 2023, newsletter reports on two developments involving large language models. Will Douglas Heaven says OpenAI’s superalignment team described a technique in which a less powerful model supervises a more powerful one, and suggests it might be a small step toward humans supervising superhuman machines. The newsletter notes that this work assumes machines will eventually surpass human intelligence, a position many researchers question.\nHeaven also reports that Google DeepMind used FunSearch to find a solution to a famous unsolved problem in pure mathematics. The researchers say it produced verifiable, valuable information that did not previously exist. The newsletter’s point is that large language models might contribute to discovery, despite their reputation for making things up—but only if they are coaxed in the right way and most of their output is discarded. The account does not specify the problem, the verification process, or how broadly the approach might work.","question":"What problem did FunSearch solve, and how did Google DeepMind verify its solution?","claims":[{"text":"OpenAI’s superalignment team described a technique for having a less powerful large language model supervise a more powerful one, and suggested it might help explore how humans could supervise superhuman machines.","excerpt":"the team describes a technique that lets a less powerful large language model supervise a more powerful one—and suggests that this might be a small step toward figuring out how humans might supervise superhuman machines.","support":"source-stated"},{"text":"The newsletter says OpenAI’s team takes machines’ eventual superiority over humans as given, although many researchers question whether machines will match or surpass human intelligence.","excerpt":"While many researchers still question whether machines will ever match human intelligence, let alone outmatch it, OpenAI’s team takes machines’ eventual superiority as given.","support":"source-stated"},{"text":"Google DeepMind says FunSearch produced a solution to a famous unsolved problem and yielded verifiable, valuable information that did not previously exist.","excerpt":"The researchers say it is the first time a large language model has been used to discover a solution to a long-standing scientific puzzle—producing verifiable and valuable new information that did not previously exist.","support":"source-stated"},{"text":"The newsletter argues that FunSearch suggests large language models can make discoveries when guided in a particular way and most of their outputs are discarded.","excerpt":"It shows that they can indeed make discoveries—if they are coaxed just so, and if you throw out the majority of what they come up with.","support":"source-stated"}],"limitations":["Text as retrieved on October 30, 2026; originally published December 15, 2023.","The newsletter does not name the mathematical problem or describe how Google DeepMind verified the reported solution.","The newsletter does not provide details of the supervision technique or evidence about how well it generalizes beyond the described work."],"mediaId":null,"correctionReason":null,"id":"3d8a8e11bcb6233b0cb12713c25795ee0f0c524b4389d44c3c53fd6d4901dda8","publicationId":"78d0fcc5-1f93-6900-e28d-30c729d3da02","version":1,"sourceUrl":"https://www.technologyreview.com/2023/12/15/1085460/the-download-beyond-crispr-and-openais-superalignment-findings/","sourceTitle":"The Download: beyond CRISPR, and OpenAI’s superalignment findings","publisher":"MIT Technology Review","publishedAt":"2023-12-15T13:10:00Z","originalPublished":{"value":"2023-12-15T13:10:00Z","precision":"instant","basis":"backfill:api_published"},"firstObservedAt":"2026-10-30T21:18:35.843Z","createdAt":"2026-11-22T04:18:12.513Z","historical":true,"entityIds":[],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.technologyreview.com/2023/12/15/1085460/the-download-beyond-crispr-and-openais-superalignment-findings/","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"91e7eb330d955622f67783246dd35f12da3e7d240dd4a586ba78f4cdc2184d28","rubricVersion":"unrecorded","provider":"anthropic","model":"claude-opus-5-5","reasonCodes":["bridge_admitted_by_model","second_opinion_categorical"],"confidence":1,"decidedAt":"2026-10-30T21:18:32.893Z"},"provenance":{"textHash":"ee8841c1cb90f4b8ccf3b2858699b12c1e29121178ffa0c694b9a19c4bfa6f94","observationId":"78df2c855f7f9003fdc7622442d1f9f635ab58c5d81b293bde6caab3945c865c","retrievedAt":"2026-10-30T21:57:48.300Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"f0797528-0e3e-6a56-8ab2-2ef39c144c6f"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"a97b249d-aa96-1c63-f345-0643d5f44a9a"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-22T04:14:36.939Z","model":"anthropic/claude-opus-5.5"}}}],"audits":[{"publicationId":"78d0fcc5-1f93-6900-e28d-30c729d3da02","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The issue is a single MIT Technology Review newsletter with blurbs written in its own words, so it counts as one piece. The OpenAI and FunSearch items are both signed Will Douglas Heaven. The discovery claims are correctly credited to the researchers.","excerpt":"The researchers say it is the first time a large language model has been used to discover a solution to a long-standing scientific puzzle"},{"dimension":"summary","status":"uncertain","explanation":"The reading names OpenAI's superalignment team as a central actor but never explains the term. The source defines it as an in-house initiative to prevent a superintelligence, a hypothetical computer that outsmarts humans, from going rogue. Without that, a general reader cannot see why weak-to-strong supervision matters. FunSearch is also never identified as Google DeepMind's new tool. Smallest repair: add one clause giving the source's definition of superalignment, and one calling FunSearch DeepMind's new tool. Minor point: the reading says LLMs 'might contribute to discovery', which softens the source's 'shows that they can indeed make discoveries'.","excerpt":"the firm’s in-house initiative dedicated to preventing a superintelligence—a hypothetical future computer that can outsmart humans—from going rogue"},{"dimension":"date","status":"pass","explanation":"December 15, 2023 matches the UTC day of originalPublished. The source is a publisher page, so no archive disclosure is needed. Later Deep Dive teasers on the page, such as OpenAI's 377 math findings, are neither used nor relied on. No hindsight framing appears.","excerpt":null},{"dimension":"support","status":"pass","explanation":"Each claim is directly supported by the quoted source text. The listed limitations about the unnamed problem, the verification and generalization are accurate gaps. The question follows from those gaps without assuming an answer. The reading adds no unsupported benefits or costs.","excerpt":"It shows that they can indeed make discoveries—if they are coaxed just so, and if you throw out the majority of what they come up with."},{"dimension":"media","status":"pass","explanation":"No media is selected and no verified candidates were supplied.","excerpt":null}],"createdAt":"2026-11-22T04:14:36.939Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}