The Derby Mill Series

• Welcome to the Era of Experience (The Derby Mill Series ep 10)

  Derby Mill co-host Richard Sutton and his former student, David Silver, recently published a paper about the future of artificial intelligence, called Welcome to the Era of Experience. So in this episode the show’s other hosts—Ajay Agrawal, Sendhil Mullainathan and Niamh Gavin—take their chance to interview Rich about the essay, and provide their take on its implications.Today’s large language models (LLMs) are trained on human-generated data. So far, this has led to the development of incredible capabilities, such as mastering complex games like backgammon or chess, or absorbing content created by humans and creating fascinating new iterations of art.While the evolution of LLMs—from AlphaZero (2017) to ChatGPT (2022) to DeepSeek (2025) and beyond—can make it seem as though their possibilities are endless, the agents remain constrained by the scope of the data they are given. In the paper, Silver and Sutton write that “in key domains such as mathematics, coding, and science, the knowledge extracted from human data is rapidly approaching a limit.” Consequently, AI agents will have to be trained on other data, such as their own experiences, which could lead to rapid innovation and superhuman capabilities—a time period which Silver and Sutton refer to as the “age of experience.”This episode, a roundtable discussion, focuses on the following quotes pulled from the paper:* Why now? "This will become possible, as outlined above, when agents are able to autonomously act and observe in streams of real-world experience, and where the rewards may be flexibly connected to any of an abundance of grounded, real-world signals."* Why science? "Perhaps most transformative will be the acceleration of scientific discovery."* Human-like vs superhuman AIs. "This era of experience will likely be characterised by agents and environments that, in addition to learning from vast quantities of experiential data, will break through the limitations of human-centric AI systems... Furthermore, the pursuit of this agenda by the AI community will spur new innovations in these directions that rapidly progress AI towards truly superhuman agents.”GUESTS AND HOSTSAjay Agrawal, co-founder and partner, Intrepid Growth PartnersRichard Sutton, senior advisor, Intrepid Growth Partners, 2024 Turing Award recipient, pioneer of reinforcement learning and professor, University of AlbertaSendhil Mullainathan, senior advisor, Intrepid Growth Partners, MacArthur Genius grant recipient and professor, MITNiamh Gavin, senior advisor, Intrepid Growth Partners, Applied AI scientist, CEO, Emergent PlatformsLINKSDerby Mill show websiteRead Rich Sutton’s latest paper Welcome to the Era of ExperienceRich Sutton’s 2019 paper The Bitter LessonCo-founder of OpenAI, Ilya Sutskever, says AI reasoning power will become less predictableListen to our previous episode about DeepSeekCheck out co-author David Silver’s websiteRich Sutton’s home page. Follow Rich on XSendhil Mullainathan’s website. Follow Sendhil on XBe sure to catch every episode of The Derby Mill Series by subscribing on the following platforms: YouTube // Spotify // Apple Podcasts DISCUSSION POINTS00:00 Introduction01:34 Context about the paper02:56 Chronology of AI paradigms03:10 Why now?06:09 Niamh’s chronology of AI development14:10 Why science?20:36 Sendhil on scientific research and AI27:07 Grounded vs. ungrounded rewards29:21 Rich on RL temporal difference errors31:10 Human-like vs. superhuman AIs36:40 Final commentsNugget 01 - AI for Scientific DiscoveryNugget 02 - Is Science like RL?Nugget 03 - The Value of ExperienceDISCLAIMERThe content of this podcast is for informational and educational purposes only and should not be construed as marketing, solicitation, or an offer to buy or sell any securities or investments. The opinions expressed in this video are those of the participants and do not necessarily reflect the views of Intrepid Growth Partners or its affiliates. Any discussion of specific companies, technologies, or industries is for illustrative purposes and does not constitute investment advice. Viewers are encouraged to consult with their own financial, legal, and tax advisors before making any investment decisions. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit insights.intrepidgp.com

  --------  

  39:27
• Cancer Detection (The Derby Mill Series ep 09)

  Skin Analytics is a UK company using AI to automate the diagnosis of serious skin conditions, starting with skin cancer. Its core product, DERM, is the only Class III CE mark AI medical device for autonomous dermatology in the UK’s health system. Used on more than 150,000 real-world patients, DERM achieves 99.8% negative predictive value, outperforming dermatologists. The company is expanding into general dermatology and launching in the EU and US.In the future, Skin Analytics intends to create a dermatology AI platform that is able to diagnose and treat a broader range of conditions. Based on a diverse sampling of low-cost data, the company intends its platform to transition from self-supervised to unsupervised learning, enabling ubiquitous, low-friction health monitoring.This episode features the Intrepid team exploring such questions as:* What would it take to build healthcare around AI abundance, not human bottlenecks?* How might one frame an approach to reach 99% automation in dermatological triage?* What are the tradeoffs between sensitivity, specificity, and health system efficiency?* How could reward systems (RL or pathway-based optimization) be introduced?* What’s the potential of self-supervised learning across multiple medical modalities?GUESTS AND HOSTSNeil Daly, founder and director, Skin AnalyticsJack Greenhalgh, AI director, Skin AnalyticsAjay Agrawal, co-founder and partner, Intrepid Growth PartnersRichard Sutton, senior advisor, Intrepid Growth Partners, 2024 Turing Award recipient, pioneer of reinforcement learning and professor, University of AlbertaSendhil Mullainathan, senior advisor, Intrepid Growth Partners, MacArthur Genius grant recipient and professor, MITNiamh Gavin, senior advisor, Intrepid Growth Partners, Applied AI scientist, CEO, Emergent PlatformsLINKSDerby Mill show website: insights.intrepidgp.com/podcastSkin Analytics website and explainer videoRich Sutton’s home page. Follow Rich on XSendhil Mullainathan’s website. Follow Sendhil on XBe sure to catch every episode of The Derby Mill Series by subscribing on the following platforms: YouTube // Spotify // Apple Podcasts DISCUSSION POINTS00:00 Introduction01:24 Meet the team: Skin Analytics06:12 The lead-up to image recognition10:29 Patient drop-off post-referral14:03 Getting classification right18:47 Integrating into the healthcare system22:36 Cancer detection in the limit27:55 At-home cancer detection34:10 Making dermatology RL-able45:00 Using data as proxies for other diagnoses50:21 Early detection vs. overdiagnosis55:07 Higher rates of cancer detection advantages57:00 What took so long?59:07 Final remarksNugget 01 - Sensors Reveal Hidden Data in the SkinTraditionally, dermatology has been rate-limited by the human eye and optical sensors. So incorporating a variety of additional sensors to collect more diverse and comprehensive data can open the door to a new kind of pre-primary care, potentially revealing more information about internal conditions like hypertension or liver disease.Nugget 02 - The Economic Model Behind At-Home DiagnosesThere's a massive direct-to-consumer interest in skin health, which opens the door to a potential expansion of at-home skin-monitoring apps that could be used beyond only in primary care settings. But overdiagnoses risk overwhelming the healthcare system. In order to avoid case buildup, these apps require an economic model that leverages medical systems and consumer trust.Nugget 03 - Redesigning the Treatment DelayWhat prevents people from accessing treatment is not the diagnostic delay (which often involves a lengthy wait for results), but rather the delay in seeking help: People tend to wait for a reason to address an issue, which increases the risk of lowering the survival rate as a disease spreads.DISCLAIMERIntrepid GP is an investor in Skin Analytics. The content of this podcast is for informational and educational purposes only and should not be construed as marketing, solicitation, or an offer to buy or sell any securities or investments. The opinions expressed in this video are those of the participants and do not necessarily reflect the views of Intrepid Growth Partners or its affiliates. Any discussion of specific companies, technologies, or industries is for illustrative purposes and does not constitute investment advice. Viewers are encouraged to consult with their own financial, legal, and tax advisors before making any investment decisions. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit insights.intrepidgp.com

  --------  

  1:02:00
• Customer Support Unpacked (The Derby Mill Series ep 08)

  In this unpacked episode, the team further expands its discussion of themes that came up in episode seven, which explored the automation of customer support with artificial intelligence. Our guests in that episode were a duo that is leading efforts in that space: CEO Mike Murchison and chief product and technology officer Mike Gozzo from Ada.In this episode, Intrepid Growth Partners cofounder and partner Ajay Agrawal leads the discussion with Intrepid Senior Advisors Rich Sutton (Turing Award winner), Sendhil Mullainathan (MacArthur genius grant recipient) and Niamh Gavin (CEO, Emergent Platforms).In the previous episode, we learned that Ada’s north star is “percent automated resolutions”, or the percentage of customer inquiries that are fully resolved by AI without human intervention. One challenge is that Ada relies on large language models (LLMs) rather than action-based goals, often requiring human agents to step in when confidence is low.“It’s a mistake to think that [Ada’s AI agents] have goals,” says Sutton. ”What we have instead … is we have [AI agents] mimicking people.”All of which raises the question of how customer support will evolve as this technology advances towards the limit.Our team also debates the need for clear, objective measures of AI performance and the challenges of achieving true goal-oriented AI systems.Our panel of experts:Ajay Agrawal, co-founder and partner, Intrepid Growth PartnersRichard Sutton, 2024 Turing Award recipient, pioneer of reinforcement learning and professor, University of AlbertaSendhil Mullainathan, MacArthur Genius grant recipient and professor, MITNiamh Gavin, Applied AI scientist, CEO, Emergent PlatformsSutton, Mullainathan and Gavin are all Intrepid Growth Partners’ senior advisors.LINKSAda websiteThis episode extends the discussion from Derby Mill episode 07: Customer Support Rich Sutton’s home page. Follow Rich on XSendhil Mullainathan’s website. Follow Sendhil on XBe sure to catch every episode by subscribing on the following platforms:YouTube // Spotify // Apple PodcastsDISCUSSION POINTS00:00 Introduction and opening credit02:00 Ada refresher03:46 Clip: Testing harness06:50 Clip discussion begins10:15 What are goal-based objectives?13:30 Is this the year of the agent?17:40 What makes agents goal-oriented19:20 Decision-making fundamentals in AI21:27 Clip: Automating system improvement over time23:32 Clip discussion begins30:13 Automating the evaluation process34:08 What could Ada look like in the limit?41:03 Closing remarksNUGGET 01: Vertical vs. Horizontal CompetitionFine-tuning used to be costly and impractical, pushing companies to open-source solutions—only to revert to OpenAI due to complexity. Now, companies like Ada build on top of model providers, offering flexibility while managing AI’s complexity. Niamh discusses the competitiveness between verticalized AI (industry-specific applications) and horizontal AI (broad sector models).NUGGET 02: The Challenge of InterpretabilityAda's evaluations rely on human judgment. The challenge here is interpretability—determining whether an outcome is truly good without direct human input. Rich Sutton offers potential solutions, including using reinforcement learning with human feedback (RLHF) as a proxy measure trained on high-quality data.NUGGET 03: Benchmarking vs. Deployment in the FieldNiamh and Sendhil discuss how, despite concerns about hallucinations in AI-powered customer service, CEOs adopt GenAI more for signaling competence than for real effectiveness.DISCLAIMERThe content of this podcast is for informational and educational purposes only and should not be construed as marketing, solicitation, or an offer to buy or sell any securities or investments. The opinions expressed in this video are those of the participants and do not necessarily reflect the views of Intrepid Growth Partners or its affiliates. Any discussion of specific companies, technologies, or industries is for illustrative purposes and does not constitute investment advice. Viewers are encouraged to consult with their own financial, legal, and tax advisors before making any investment decisions. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit insights.intrepidgp.com

  --------  

  43:04
• Customer Support (The Derby Mill Series ep 07)

  Meet Ada, a Canadian AI agent platform automating the resolution of customer service interactions. When customers have complex requests—such as resetting passwords, checking order status, or requesting a refund—Ada uses large language models to radically reduce the amount of human effort required to fulfill the customer’s inquiry.Here, Ada CEO Mike Murchison and Chief Product & Technology Officer Mike Gozzo join the Derby Mill podcast to discuss the intersection of AI and customer support—and where the technology may go, at the limit.Our panel of experts:Ajay Agrawal, co-founder and partner, Intrepid Growth PartnersRichard Sutton, 2024 Turing Award recipient, pioneer of reinforcement learning and professor, University of AlbertaSendhil Mullainathan, MacArthur Genius grant recipient and professor, MITNiamh Gavin, Applied AI scientist, CEO, Emergent PlatformsLINKSAda websiteAda CEO Mike Murchison LinkedInAda Chief Product & Technology Officer Mike Gozzo LinkedInRich Sutton’s home page. Follow Rich on X.Sendhil Mullainathan’s website. Follow Sendhil on X.Sendhil’s article on Algorithms Need Managers, Too published in the Harvard Business Review Be sure to catch every episode by subscribing on the following platforms:YouTube // Spotify // Apple PodcastsDISCUSSION POINTS00:00 Introduction01:49 Meet Ada, the company automating customer support05:05 Customer service & books: an analogy05:41 Murchison describes automated resolution08:50 Human feedback for automated improvement 23:17 LLMs in customer service26:10 The difference between language and action26:34 Ada’s use of LLMs30:01 Murchison on how “deterministic” Ada’s actions are30:59 Improving decision quality37:06 Protecting against LLM’s unreliability 44:40 Closing remarksNUGGET 01: Human Feedback for Automated ImprovementAda describes the role of humans "coaching" their AIs. Why this is one of the first areas for "automated improvement," and how can the preference data they are collecting through the coaching process be used to "drive automated improvements throughout the entire system."NUGGET 02: Decision-Making QualityRich Sutton asks how Ada improves the quality of the system's decisions, and questions the role of humans vs. AI in terms of evaluating versus improving the quality of decisions.NUGGET 03: DistillationGiven the cost and latency virtues of smaller models, when do we anticipate applications to use large foundation models at the limit? Is the Ada case a good example of using large models to bootstrap a commercial solution en route to smaller, more specialized models?DISCLAIMER The content of this podcast is for informational and educational purposes only and should not be construed as marketing, solicitation, or an offer to buy or sell any securities or investments. The opinions expressed in this video are those of the participants and do not necessarily reflect the views of Intrepid Growth Partners or its affiliates. Any discussion of specific companies, technologies, or industries is for illustrative purposes and does not constitute investment advice. Viewers are encouraged to consult with their own financial, legal, and tax advisors before making any investment decisions. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit insights.intrepidgp.com

  --------  

  51:06
• Mining Exploration Unpacked (The Derby Mill Series ep 06)

  In the first episode of The Derby Mill Series to post since our co-host Rich Sutton won the Turing Award, the highest distinction in the field of computer science, Sutton joins our panel of experts to further discuss the application of artificial intelligence to the task of mining exploration. What ensues is a remarkable conversation about the increasing relevance of cheap and easily discoverable data sources in an age dominated by artificial intelligence and reinforcement learning.The analogy begins with mining, where core samples may provide an element of ground truth but are, at the same time, tough and very expensive to get. Other data sources, such as aerial imagery, chips and dust, are cheap and more easily available. So can the pattern-recognition abilities of artificial intelligence elevate the relevance of that lower-fidelity, more easily available information?Extending the analysis, Intrepid Senior Advisors Niamh Gavin and Sendhil Mullainathan draw parallels with health care. Apple Watch’s skin sensors are certainly less accurate than an annually drawn blood test. But as the Watch conducts its tests numerous times a day, and as AI better recognizes troubling sensor patterns, the cheaper Watch data could become just as important as more expensive medical diagnostics.Generalizing to other areas, Intrepid partner Ajay Agrawal notes that lower-fidelity data that is more easily available could become as informative as high-fidelity data that is tougher to extract. Our experts’ ultimate prediction, then, observes that cheap data plus artificial intelligence could transform the fundamental economics of many different industries.Our panel of expertsRichard Sutton, 2024 winner of the Turing Award, pioneer of reinforcement learning and professor, University of AlbertaSendhil Mullainathan, MacArthur Genius grant recipient and professor, MITNiamh Gavin, Applied AI scientist, CEO, Emergent PlatformsAjay Agrawal, co-founder and partner, Intrepid Growth PartnersLINKSIntrepid Growth Partners’ Senior Advisor Rich Sutton wins the Turing Award. NY Times. Financial Times. Betakit.This episode extends a discussion in Derby Mill episode 05: Mining Exploration.Referenced in this episode: Sendhil Mullainathan's heart attack studyyRich Sutton’s home page. Follow Rich on XSendhil Mullainathan’s website. Follow Sendhil on XBe sure to catch every episode by subscribing on the following platforms:YouTube // Spotify // Apple PodcastsDISCUSSION POINTS00:00 Introductions and opening credit01:43 Recap of Mining Exploration06:01 Clips: Mining industry transformation09:52 Niamh on narrowing the search zone12:18 Rich on sequential decisions and pattern recognition15:08 Sendhil on using supervised learning to train predictors19:34 Niamh on non-invasive markers21:52 Signals in healthcare vs. mining25:02 The combination of human + AI27:47 A new age of data analysis29:50 Data sources and reinforcement learning35:53 A cognitive barrier for data42:24 The indicator analogy45:58 Closing remarksNUGGETS (short excerpts from the full episode)NUGGET 01: Sum > Whole of Its PartsNiamh Gavin argues that human + AI intelligence is better than either in isolation.NUGGET 02: A More Interactive Feedback ProcessRich Sutton advocates for an awareness that important mining exploration problems require a wide diversity of data inputs.NUGGET 03: The Less Invasive, Far Cheaper Data Axis“What if I had a sweat test that was 10% as good as a blood test?” asks Sendhil Mullainathan, noting the way AI can make use of data from cheap and more easily available diagnostics to improve numerous different industries.NUGGET 04: AI and the Increasing Relevance of Little TestsExpert intelligence is expensive today, notes Rich Sutton, but as computational power decreases, AI will help to bolster the importance of all sorts of cheap and easy tests.DISCLAIMERThe content of this podcast is for informational and educational purposes only and should not be construed as marketing, solicitation, or an offer to buy or sell any securities or investments. The opinions expressed in this video are those of the participants and do not necessarily reflect the views of Intrepid Growth Partners or its affiliates. Any discussion of specific companies, technologies, or industries is for illustrative purposes and does not constitute investment advice. Viewers are encouraged to consult with their own financial, legal, and tax advisors before making any investment decisions. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit insights.intrepidgp.com

  --------  

  46:46

Plus de podcasts Technologies

Podcasts tendance de Technologies

À propos de The Derby Mill Series