{"id":"millidge-learning-beyond-demonstrations","title":"A clear target lets a model try what people have not demonstrated","topicIds":["intelligence-and-discovery"],"domain":"science","speaker":{"name":"Beren Millidge","roleAtSource":"Zyphra CTO, introduced in the panel","expertise":"Researcher developing open models at Zyphra, as introduced in the episode.","claimBoundary":"A technical argument about learning; it does not demonstrate a result in a specific business setting.","declaredInterests":[{"text":"Works at Zyphra, which develops open models.","url":"https://www.dwarkesh.com/p/john-beren-charlie"}],"bioSources":[{"title":"Panel introduction","url":"https://www.dwarkesh.com/p/john-beren-charlie"}]},"source":{"title":"AI researchers debate how close we are to recursive self-improvement","publisher":"Dwarkesh Podcast","format":"Panel transcript","url":"https://www.dwarkesh.com/p/john-beren-charlie","publishedAt":"2026-09-11","dateBasis":"Publication date displayed on the official episode page.","anchor":{"label":"28:06 — Learning through testable attempts","url":"https://www.youtube.com/watch?v=PrSf7IOYu-I&t=1686","timestampSeconds":1686},"accessNote":"Publisher transcript read on 23 September 2026, New York time. Timestamp marks a chapter, not an exact quote. Audio/video not independently listened to."},"argument":"Millidge argues that models can attempt harder goals than people can demonstrate when success is verifiable. Whether that learning transfers far enough to remove human guidance remains uncertain.","claimType":"interpretation","evidence":[{"text":"Millidge contrasts copying demonstrated solutions with learning from attempts in environments that can check success. He remains uncertain whether that learning transfers far enough to remove human guidance.","url":"https://www.dwarkesh.com/p/john-beren-charlie","kind":"interpretive","relationship":"supports"}],"assumptions":["Editorial reconstruction: the outcome can be checked and captures the purpose of the work."],"whatWouldChangeView":{"attribution":"Editorial test, not a stated concession","text":"Test whether improvements against the chosen target also improve outcomes on new cases, including failures the target does not directly measure."},"limits":"An easy-to-score task can omit the reason the work mattered. This argument is a reason to test transfer, not assume it.","discussionQuestion":"Where could your team state the desired result clearly enough to test approaches nobody has tried?","relatedSignalIds":["research-panel-choosing-next-question","developer-productivity-measurement"],"editorialComparisons":[],"sourceMedia":{"status":"found","sourceUrl":"https://www.dwarkesh.com/p/john-beren-charlie","checkedAt":"2026-10-15T20:07:33.560800Z","items":[{"type":"video","url":"https://www.youtube.com/watch?v=PrSf7IOYu-I&t=1686","sourceUrl":"https://www.dwarkesh.com/p/john-beren-charlie","attribution":"Dwarkesh Podcast","basis":"publisher-embed","availability":"verified","verifiedAt":"2026-10-15T20:07:33.560800Z","label":"Watch from 28:06","preview":{"url":"https://i.ytimg.com/vi/PrSf7IOYu-I/hqdefault.jpg","alt":"Episode thumbnail: AI researchers debate how close we are to recursive self-improvement","credit":"Dwarkesh Patel / YouTube"}}],"note":"Opens the recorded source chapter; the linked chapter may begin before the passage discussed."},"canonicalUrl":"https://escapemeridian.com/readings/millidge-learning-beyond-demonstrations","formats":{"html":"/readings/millidge-learning-beyond-demonstrations","markdown":"/readings/millidge-learning-beyond-demonstrations.md","json":"/readings/millidge-learning-beyond-demonstrations.json"},"attribution":"Headline and sourced paraphrase by Escape Meridian. Inclusion is not endorsement.","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":[],"entities":[{"id":"beren-millidge","name":"Beren Millidge","kind":"person","context":"Beren Millidge is introduced as Zyphra's chief technology officer developing open models in the September 2026 Dwarkesh panel on AI research.","referenceUrls":["https://www.dwarkesh.com/p/john-beren-charlie"],"path":"/entities/beren-millidge","json":"/entities/beren-millidge.json","markdown":"/entities/beren-millidge.md"},{"id":"dwarkesh-podcast","name":"Dwarkesh Podcast","kind":"show","context":"Dwarkesh Podcast features long conversations with researchers, builders and thinkers about technology, history and the future.","referenceUrls":["https://www.dwarkesh.com/"],"path":"/entities/dwarkesh-podcast","json":"/entities/dwarkesh-podcast.json","markdown":"/entities/dwarkesh-podcast.md"}]},"withdrawn":false,"revisions":[],"audits":[],"correctionUrl":"https://escapemeridian.com/corrections/"}