{"id":"7367f8683c17f4bf0b57b7c4e819987f723e0462b68e96d6875ed4c2309432cd","title":"Ajeya Cotra argues in Planned Obsolescence that AI capabilities outpaced her forecast","argument":"On March 5, 2026, Ajeya Cotra revisits her forecast for AI progress and says her software engineering forecast already feels too conservative. She points to Opus 4.6 solving some of METR’s longer software tasks, while warning that its estimated time horizon is noisy and the benchmark nearly saturated. Cotra argues that solo task time may become a poor guide to software capability as projects grow: large projects can be divided among workers, but shorter tasks often depend on shared context. She considers whether AI agents could use managers, records and extensive project tracking to make progress on much larger projects. Yet clean decomposition can itself take substantial work, and she says projects benefit from people understanding the bigger picture. Cotra calls full AI R&D automation a tall order, requiring progress in research judgment and creativity. Still, she says there is no solid trend she can extrapolate to rule out its happening this year. Her conclusion is that AI R&D really could be automated this year, not that it will be. Whether agent teams can sustain progress on large projects without holistic context remains unknown.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 27, 2026; originally published March 5, 2026. Cotra cautions that the time horizon estimate for the latest models is noisy, the benchmark is nearly saturated, and its confidence interval is wide. Cotra presents AI R&D automation as a possibility, not an outcome; the piece does not establish whether agent teams can sustain progress on large projects.","speaker":{"name":"Ajeya Cotra","roleAtSource":"Publisher"},"source":{"title":"I underestimated AI capabilities (again)","publisher":"Ajeya Cotra","format":"Article","url":"https://www.planned-obsolescence.org/p/i-underestimated-ai-capabilities","publishedAt":"2026-03-05T15:02:56+00:00","dateBasis":"Date stated on the publisher’s page."},"evidence":[{"text":"Cotra says her software engineering capability forecast was too conservative.","excerpt":"My forecasts for software engineering capabilities already feel much too conservative.","url":"https://www.planned-obsolescence.org/p/i-underestimated-ai-capabilities","relationship":"Source-stated"},{"text":"Cotra says Opus 4.6 solved some of METR’s longer software tasks, while warning that its estimated time horizon is noisy and the benchmark is nearly saturated.","excerpt":"Opus 4.6 was able to solve 14 of them at least some of the time","url":"https://www.planned-obsolescence.org/p/i-underestimated-ai-capabilities","relationship":"Source-stated"},{"text":"Cotra argues that large projects may be divided into smaller tasks, but that clean decomposition and shared context remain challenges.","excerpt":"Decomposing a 6 month project so cleanly and precisely that it can be executed by a team of people with no such holistic context might itself be, say, a 2 month task.","url":"https://www.planned-obsolescence.org/p/i-underestimated-ai-capabilities","relationship":"Source-stated"},{"text":"Cotra says AI R&D really could be automated this year, while describing full automation as a tall order requiring progress in research judgment and creativity.","excerpt":"AI R&D really could be automated this year.","url":"https://www.planned-obsolescence.org/p/i-underestimated-ai-capabilities","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.planned-obsolescence.org/p/i-underestimated-ai-capabilities","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"What evidence could show whether agent teams can sustain progress on large software projects without holistic context?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Ajeya Cotra argues in Planned Obsolescence that AI capabilities outpaced her forecast","summary":"On March 5, 2026, Ajeya Cotra revisits her forecast for AI progress and says her software engineering forecast already feels too conservative. She points to Opus 4.6 solving some of METR’s longer software tasks, while warning that its estimated time horizon is noisy and the benchmark nearly saturated. Cotra argues that solo task time may become a poor guide to software capability as projects grow: large projects can be divided among workers, but shorter tasks often depend on shared context. She considers whether AI agents could use managers, records and extensive project tracking to make progress on much larger projects. Yet clean decomposition can itself take substantial work, and she says projects benefit from people understanding the bigger picture. Cotra calls full AI R&D automation a tall order, requiring progress in research judgment and creativity. Still, she says there is no solid trend she can extrapolate to rule out its happening this year. Her conclusion is that AI R&D really could be automated this year, not that it will be. Whether agent teams can sustain progress on large projects without holistic context remains unknown.","question":"What evidence could show whether agent teams can sustain progress on large software projects without holistic context?","claims":[{"text":"Cotra says her software engineering capability forecast was too conservative.","excerpt":"My forecasts for software engineering capabilities already feel much too conservative.","support":"source-stated"},{"text":"Cotra says Opus 4.6 solved some of METR’s longer software tasks, while warning that its estimated time horizon is noisy and the benchmark is nearly saturated.","excerpt":"Opus 4.6 was able to solve 14 of them at least some of the time","support":"source-stated"},{"text":"Cotra argues that large projects may be divided into smaller tasks, but that clean decomposition and shared context remain challenges.","excerpt":"Decomposing a 6 month project so cleanly and precisely that it can be executed by a team of people with no such holistic context might itself be, say, a 2 month task.","support":"source-stated"},{"text":"Cotra says AI R&D really could be automated this year, while describing full automation as a tall order requiring progress in research judgment and creativity.","excerpt":"AI R&D really could be automated this year.","support":"source-stated"}],"limitations":["Text as retrieved on October 27, 2026; originally published March 5, 2026.","Cotra cautions that the time horizon estimate for the latest models is noisy, the benchmark is nearly saturated, and its confidence interval is wide.","Cotra presents AI R&D automation as a possibility, not an outcome; the piece does not establish whether agent teams can sustain progress on large projects."],"mediaId":null,"correctionReason":null,"id":"7367f8683c17f4bf0b57b7c4e819987f723e0462b68e96d6875ed4c2309432cd","publicationId":"61b695c0-3f06-7ecf-8164-4c69c75b48ba","version":1,"sourceUrl":"https://www.planned-obsolescence.org/p/i-underestimated-ai-capabilities","sourceTitle":"I underestimated AI capabilities (again)","publisher":"Ajeya Cotra","publishedAt":"2026-03-05T15:02:56+00:00","originalPublished":{"value":"2026-03-05T15:02:56+00:00","precision":"instant","basis":"backfill:article_published"},"firstObservedAt":"2026-10-27T20:13:22.992Z","createdAt":"2026-11-08T02:33:24.155Z","historical":true,"entityIds":[],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.planned-obsolescence.org/p/i-underestimated-ai-capabilities","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"638ec92f06513c737c9c23f687d6f68b387a345bfc05ef110815b6df1a8ca0bf","rubricVersion":"unrecorded","provider":"typesafe","model":"jev-1.13.0","reasonCodes":["bridge_admitted_by_model"],"confidence":0.79,"decidedAt":"2026-10-27T20:13:22.037Z"},"provenance":{"textHash":"227d298b68fd244da3a7d3a4bc68b4738a0fe37f823835bf10aa34a2ff05d795","observationId":"39a79a8213ea4de383f5fab866c18a79ab524c71a9486a1c32e1ebbbdae16aab","retrievedAt":"2026-10-27T20:15:17.297Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"805f4e22-f98f-cf28-3a05-2ba9d3e43b25"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"9f754d57-2e80-6a3e-28ea-c5267f896cdf"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-08T02:28:41.757Z","model":"anthropic/claude-opus-5.5"}},"canonicalUrl":"https://escapemeridian.com/readings/ajeya-cotra-argues-in-planned-obsolescence-that-ai-capabilities-7367f868","formats":{"html":"/readings/ajeya-cotra-argues-in-planned-obsolescence-that-ai-capabilities-7367f868","markdown":"/readings/ajeya-cotra-argues-in-planned-obsolescence-that-ai-capabilities-7367f868.md","json":"/readings/ajeya-cotra-argues-in-planned-obsolescence-that-ai-capabilities-7367f868.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":[],"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-time-horizon","name":"Time horizon","explanation":"METR's measure of AI agents' ability: the length of task, timed by how long skilled people take, that an agent completes at a set success rate.","referenceUrl":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/"},{"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"},{"id":"term-ai-agent","name":"AI agent","explanation":"An AI system that chooses and carries out actions toward a goal, often using software tools.","referenceUrl":"https://www.ibm.com/think/topics/ai-agents"},{"id":"term-confidence-interval","name":"Confidence interval","explanation":"A range of values consistent with the data at a stated level of confidence.","referenceUrl":"https://en.wikipedia.org/wiki/Confidence_interval"}],"entities":[{"id":"metr","name":"METR","kind":"organization","context":"METR studies what AI systems can do and how reliably they complete tasks, including work with human developers.","referenceUrls":["https://metr.org/"],"path":"/entities/metr","json":"/entities/metr.json","markdown":"/entities/metr.md"},{"id":"time","name":"TIME","kind":"organization","context":"TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.","referenceUrls":["https://time.com/"],"path":"/entities/time","json":"/entities/time.json","markdown":"/entities/time.md"}]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"7367f8683c17f4bf0b57b7c4e819987f723e0462b68e96d6875ed4c2309432cd","revision":"61b695c0-3f06-7ecf-8164-4c69c75b48ba","revisionKind":"publication-id","title":"Ajeya Cotra argues in Planned Obsolescence that AI capabilities outpaced her forecast","url":"https://escapemeridian.com/readings/ajeya-cotra-argues-in-planned-obsolescence-that-ai-capabilities-7367f868"},"source":{"type":"source","id":"https://www.planned-obsolescence.org/p/i-underestimated-ai-capabilities","revision":"227d298b68fd244da3a7d3a4bc68b4738a0fe37f823835bf10aa34a2ff05d795","revisionKind":"sha256-source-text","title":"I underestimated AI capabilities (again)","url":"https://www.planned-obsolescence.org/p/i-underestimated-ai-capabilities"},"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":"7367f8683c17f4bf0b57b7c4e819987f723e0462b68e96d6875ed4c2309432cd","title":"Ajeya Cotra argues in Planned Obsolescence that AI capabilities outpaced her forecast","argument":"On March 5, 2026, Ajeya Cotra revisits her forecast for AI progress and says her software engineering forecast already feels too conservative. She points to Opus 4.6 solving some of METR’s longer software tasks, while warning that its estimated time horizon is noisy and the benchmark nearly saturated. Cotra argues that solo task time may become a poor guide to software capability as projects grow: large projects can be divided among workers, but shorter tasks often depend on shared context. She considers whether AI agents could use managers, records and extensive project tracking to make progress on much larger projects. Yet clean decomposition can itself take substantial work, and she says projects benefit from people understanding the bigger picture. Cotra calls full AI R&D automation a tall order, requiring progress in research judgment and creativity. Still, she says there is no solid trend she can extrapolate to rule out its happening this year. Her conclusion is that AI R&D really could be automated this year, not that it will be. Whether agent teams can sustain progress on large projects without holistic context remains unknown.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 27, 2026; originally published March 5, 2026. Cotra cautions that the time horizon estimate for the latest models is noisy, the benchmark is nearly saturated, and its confidence interval is wide. Cotra presents AI R&D automation as a possibility, not an outcome; the piece does not establish whether agent teams can sustain progress on large projects.","speaker":{"name":"Ajeya Cotra","roleAtSource":"Publisher"},"source":{"title":"I underestimated AI capabilities (again)","publisher":"Ajeya Cotra","format":"Article","url":"https://www.planned-obsolescence.org/p/i-underestimated-ai-capabilities","publishedAt":"2026-03-05T15:02:56+00:00","dateBasis":"Date stated on the publisher’s page."},"evidence":[{"text":"Cotra says her software engineering capability forecast was too conservative.","excerpt":"My forecasts for software engineering capabilities already feel much too conservative.","url":"https://www.planned-obsolescence.org/p/i-underestimated-ai-capabilities","relationship":"Source-stated"},{"text":"Cotra says Opus 4.6 solved some of METR’s longer software tasks, while warning that its estimated time horizon is noisy and the benchmark is nearly saturated.","excerpt":"Opus 4.6 was able to solve 14 of them at least some of the time","url":"https://www.planned-obsolescence.org/p/i-underestimated-ai-capabilities","relationship":"Source-stated"},{"text":"Cotra argues that large projects may be divided into smaller tasks, but that clean decomposition and shared context remain challenges.","excerpt":"Decomposing a 6 month project so cleanly and precisely that it can be executed by a team of people with no such holistic context might itself be, say, a 2 month task.","url":"https://www.planned-obsolescence.org/p/i-underestimated-ai-capabilities","relationship":"Source-stated"},{"text":"Cotra says AI R&D really could be automated this year, while describing full automation as a tall order requiring progress in research judgment and creativity.","excerpt":"AI R&D really could be automated this year.","url":"https://www.planned-obsolescence.org/p/i-underestimated-ai-capabilities","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.planned-obsolescence.org/p/i-underestimated-ai-capabilities","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"What evidence could show whether agent teams can sustain progress on large software projects without holistic context?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Ajeya Cotra argues in Planned Obsolescence that AI capabilities outpaced her forecast","summary":"On March 5, 2026, Ajeya Cotra revisits her forecast for AI progress and says her software engineering forecast already feels too conservative. She points to Opus 4.6 solving some of METR’s longer software tasks, while warning that its estimated time horizon is noisy and the benchmark nearly saturated. Cotra argues that solo task time may become a poor guide to software capability as projects grow: large projects can be divided among workers, but shorter tasks often depend on shared context. She considers whether AI agents could use managers, records and extensive project tracking to make progress on much larger projects. Yet clean decomposition can itself take substantial work, and she says projects benefit from people understanding the bigger picture. Cotra calls full AI R&D automation a tall order, requiring progress in research judgment and creativity. Still, she says there is no solid trend she can extrapolate to rule out its happening this year. Her conclusion is that AI R&D really could be automated this year, not that it will be. Whether agent teams can sustain progress on large projects without holistic context remains unknown.","question":"What evidence could show whether agent teams can sustain progress on large software projects without holistic context?","claims":[{"text":"Cotra says her software engineering capability forecast was too conservative.","excerpt":"My forecasts for software engineering capabilities already feel much too conservative.","support":"source-stated"},{"text":"Cotra says Opus 4.6 solved some of METR’s longer software tasks, while warning that its estimated time horizon is noisy and the benchmark is nearly saturated.","excerpt":"Opus 4.6 was able to solve 14 of them at least some of the time","support":"source-stated"},{"text":"Cotra argues that large projects may be divided into smaller tasks, but that clean decomposition and shared context remain challenges.","excerpt":"Decomposing a 6 month project so cleanly and precisely that it can be executed by a team of people with no such holistic context might itself be, say, a 2 month task.","support":"source-stated"},{"text":"Cotra says AI R&D really could be automated this year, while describing full automation as a tall order requiring progress in research judgment and creativity.","excerpt":"AI R&D really could be automated this year.","support":"source-stated"}],"limitations":["Text as retrieved on October 27, 2026; originally published March 5, 2026.","Cotra cautions that the time horizon estimate for the latest models is noisy, the benchmark is nearly saturated, and its confidence interval is wide.","Cotra presents AI R&D automation as a possibility, not an outcome; the piece does not establish whether agent teams can sustain progress on large projects."],"mediaId":null,"correctionReason":null,"id":"7367f8683c17f4bf0b57b7c4e819987f723e0462b68e96d6875ed4c2309432cd","publicationId":"61b695c0-3f06-7ecf-8164-4c69c75b48ba","version":1,"sourceUrl":"https://www.planned-obsolescence.org/p/i-underestimated-ai-capabilities","sourceTitle":"I underestimated AI capabilities (again)","publisher":"Ajeya Cotra","publishedAt":"2026-03-05T15:02:56+00:00","originalPublished":{"value":"2026-03-05T15:02:56+00:00","precision":"instant","basis":"backfill:article_published"},"firstObservedAt":"2026-10-27T20:13:22.992Z","createdAt":"2026-11-08T02:33:24.155Z","historical":true,"entityIds":[],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.planned-obsolescence.org/p/i-underestimated-ai-capabilities","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"638ec92f06513c737c9c23f687d6f68b387a345bfc05ef110815b6df1a8ca0bf","rubricVersion":"unrecorded","provider":"typesafe","model":"jev-1.13.0","reasonCodes":["bridge_admitted_by_model"],"confidence":0.79,"decidedAt":"2026-10-27T20:13:22.037Z"},"provenance":{"textHash":"227d298b68fd244da3a7d3a4bc68b4738a0fe37f823835bf10aa34a2ff05d795","observationId":"39a79a8213ea4de383f5fab866c18a79ab524c71a9486a1c32e1ebbbdae16aab","retrievedAt":"2026-10-27T20:15:17.297Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"805f4e22-f98f-cf28-3a05-2ba9d3e43b25"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"9f754d57-2e80-6a3e-28ea-c5267f896cdf"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-08T02:28:41.757Z","model":"anthropic/claude-opus-5.5"}}}],"audits":[{"publicationId":"61b695c0-3f06-7ecf-8164-4c69c75b48ba","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The piece is a single first-person essay by Ajeya Cotra on her Planned Obsolescence site. The reading attributes the forecasts and arguments to her throughout and does not present them as established fact.","excerpt":"On Jan 14th, I made predictions about AI progress in 2026."},{"dimension":"summary","status":"uncertain","explanation":"The summary is broadly faithful, but key context is missing. METR is never explained: it is the AI evaluation group where Cotra now works. Opus 4.6 is not identified as a Claude model. The terms AI R&D and saturated are not defined in the reading or in definitions. The reading drops her concrete numbers: a 24-hour median forecast and a guess of over 100 hours by year end. It also omits her 10% probability. Finally, it leaves out her counterweight that decomposition might work surprisingly well because agents are cheap and patient. Smallest repair: briefly explain METR, add the 100-hour guess and the 10% figure, and add her 'surprisingly well' caveat.","excerpt":"But it might work surprisingly well for a surprisingly large class of software projects."},{"dimension":"date","status":"pass","explanation":"The stated date, March 5, 2026, matches originalPublished and the dateline. modifiedAt equals the publication time, so there is no sign of a later edit. The March 3 bug correction mentioned in the text predates publication. The year-end automation claim is reported as Cotra's possibility, not as an outcome. The kind is publisher-page, so no archive disclosure is needed.","excerpt":"Mar 05, 2026"},{"dimension":"support","status":"pass","explanation":"Each claim is directly supported by an excerpt from the source. The reading keeps her caveats about noise, saturation and the wide confidence interval. It frames automation as 'could', not 'will'. The question about agent teams working without holistic context follows from her own discussion and is not a leading claim.","excerpt":"Fully automating AI R&D still seems like a tall order."},{"dimension":"media","status":"pass","explanation":"No media was selected and no verified candidates were supplied, so there is nothing to misidentify.","excerpt":null}],"createdAt":"2026-11-08T02:28:41.757Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}