Advertise on podcast: Generally Intelligent

Deep learning works extraordinarily well. And we still largely don't know why. A new paper from Jamie Simon, Daniel Kunin, and 12 co-authors argues that a scientific theory of deep learning is emerging, and coins a name for the emerging field: learning mechanics. We sat down with Jamie and Dan on Generally Intelligent to talk about what a physics of deep learning would actually look like, why now, and what's left to figure out. 00:03:05 Learning mechanics as the physics to mechanistic interpretability's biology 00:04:13 Why deep learning needs a theory 00:07:07 Why deep learning is uniquely hard to engineer 00:12:11 How a week in the woods became a paper 00:25:59 The barrier to theory isn't opacity, but complexity 00:36:26 Deep learning's first gas law 00:47:22 Why more particles makes the problem easier 00:56:22 The discretization hypothesis 01:01:50 The strongest signal that a compact theory exists 01:05:07 The Platonic Representation Hypothesis 01:15:41 Why learning mechanics and mech interp need each other 01:25:29 Theory as safety infrastructure Read the paper Transcript and links Learning Mechanics website Full transcript: https://imbueai.substack.com/p/geoffrey-litt Generally Intelligent is a podcast by ⁠⁠Imbue⁠⁠, a research company building toward a future where AI agents are open and accountable to their users, so people have more power in the digital world. WebsiteSubstack⁠X⁠LinkedInYouTube

Geoffrey Litt is a design engineer at Notion working on malleable software: computing environments where anyone can adapt their software to meet their needs and their lives. Before joining Notion, he was a researcher at the independent lab, Ink & Switch, where he explored the future of computing. He did his PhD at MIT on programming interfaces. Most of his work circles around a very simple but powerful question: how can everyday people shape the software they use like clay so that humans can have more power and agency in the world? In this conversation, Geoffrey and Kanjun discuss: Technical, economic, and infrastructural barriers to malleable softwareInventing new UI components for the AI agePrinciples for agent-human collaborationHow AI affects the creative process...and more! Full transcript: https://imbueai.substack.com/p/geoffrey-litt Generally Intelligent is a podcast by ⁠Imbue⁠, a research company building toward a future where AI agents are open and accountable to their users, so people have more power in the digital world. Website: ⁠https://imbue.com/⁠ Substack: ⁠https://ideas.imbue.com/ LinkedIn: ⁠https://www.linkedin.com/company/imbue_ai/⁠ X: ⁠@imbue_ai⁠ YouTube: ⁠https://www.youtube.com/@imbue_ai/

Welcome back to Generally Intelligent! We’re excited to relaunch our podcast—still featuring thoughtful conversations on building AI, but now with an expanded lens on its economic, societal, political, and human impacts. Matt Boulos leads policy and safety at Imbue, where he shapes the responsible development of AI coding tools that make software creation broadly accessible. His work centers on understanding what technological power means for individual liberty and advocates for the legal and institutional frameworks we need to protect our freedom. Matt is a lawyer, computer scientist, and founder. A full transcript is available on our Substack: https://imbueai.substack.com/matt-boulos/ Highlights: AI’s four core challenges Governing lawless digital spaces Why abundance is not enough without libertyFreedom as deep enablement and deep protectionThe role of technologists in shaping society Generally Intelligent is a podcast by Imbue, an independent research company developing a better way to build personal software. Our mission is to empower humans in the age of AI by creating powerful computing tools controlled by individuals. Website: https://imbue.com/ Substack: https://imbueai.substack.com/ LinkedIn: https://www.linkedin.com/company/imbue_ai/ X: @imbue_ai Bluesky: https://bsky.app/profile/imbue-ai.bsky.social YouTube: https://www.youtube.com/@imbue_ai/

Rylan Schaeffer is a PhD student at Stanford studying the engineering, science, and mathematics of intelligence. He authored the paper “Are Emergent Abilities of Large Language Models a Mirage?”, as well as other interesting refutations in the field that we’ll talk about today. He previously interned at Meta on the Llama team, and at Google DeepMind. Generally Intelligent is a podcast by Imbue where we interview researchers about their behind-the-scenes ideas, opinions, and intuitions that are hard to share in papers and talks. About Imbue Imbue is an independent research company developing AI agents that mirror the fundamentals of human-like intelligence and that can learn to safely solve problems in the real world. We started Imbue because we believe that software with human-level intelligence will have a transformative impact on the world. We’re dedicated to ensuring that that impact is a positive one. We have enough funding to freely pursue our research goals over the next decade, and our backers include Y Combinator, researchers from OpenAI, Astera Institute, and a number of private individuals who care about effective altruism and scientific research. Our research is focused on agents for digital environments (ex: browser, desktop, documents), using RL, large language models, and self supervised learning. We’re excited about opportunities to use simulated data, network architecture search, and good theoretical understanding of deep learning to make progress on these problems. We take a focused, engineering-driven approach to research. Website: https://imbue.com LinkedIn: https://www.linkedin.com/company/imbue_ai/ Twitter/X: @imbue_ai

Ari Morcos is the CEO of DatologyAI, which makes training deep learning models more performant and efficient by intervening on training data. He was at FAIR and DeepMind before that, where he worked on a variety of topics, including how training data leads to useful representations, lottery ticket hypothesis, and self-supervised learning. His work has been honored with Outstanding Paper awards at both NeurIPS and ICLR. Generally Intelligent is a podcast by Imbue where we interview researchers about their behind-the-scenes ideas, opinions, and intuitions that are hard to share in papers and talks. About Imbue Imbue is an independent research company developing AI agents that mirror the fundamentals of human-like intelligence and that can learn to safely solve problems in the real world. We started Imbue because we believe that software with human-level intelligence will have a transformative impact on the world. We’re dedicated to ensuring that that impact is a positive one. We have enough funding to freely pursue our research goals over the next decade, and our backers include Y Combinator, researchers from OpenAI, Astera Institute, and a number of private individuals who care about effective altruism and scientific research. Our research is focused on agents for digital environments (ex: browser, desktop, documents), using RL, large language models, and self supervised learning. We’re excited about opportunities to use simulated data, network architecture search, and good theoretical understanding of deep learning to make progress on these problems. We take a focused, engineering-driven approach to research. Website: ⁠https://imbue.com/⁠ LinkedIn: ⁠https://www.linkedin.com/company/imbue-ai/⁠ Twitter: @imbue_ai

Percy Liang is an associate professor of computer science and statistics at Stanford. These days, he’s interested in understanding how foundation models work, how to make them more efficient, modular, and robust, and how they shift the way people interact with AI—although he’s been working on language models for long before foundation models appeared. Percy is also a big proponent of reproducible research, and toward that end he’s shipped most of his recent papers as executable papers using the CodaLab Worksheets platform his lab developed, and published a wide variety of benchmarks. Generally Intelligent is a podcast by Imbue where we interview researchers about their behind-the-scenes ideas, opinions, and intuitions that are hard to share in papers and talks. About Imbue Imbue is an independent research company developing AI agents that mirror the fundamentals of human-like intelligence and that can learn to safely solve problems in the real world. We started Imbue because we believe that software with human-level intelligence will have a transformative impact on the world. We’re dedicated to ensuring that that impact is a positive one.We have enough funding to freely pursue our research goals over the next decade, and our backers include Y Combinator, researchers from OpenAI, Astera Institute, and a number of private individuals who care about effective altruism and scientific research.Our research is focused on agents for digital environments (ex: browser, desktop, documents), using RL, large language models, and self supervised learning. We’re excited about opportunities to use simulated data, network architecture search, and good theoretical understanding of deep learning to make progress on these problems. We take a focused, engineering-driven approach to research. Website: ⁠https://imbue.com/⁠ LinkedIn: ⁠https://www.linkedin.com/company/imbue-ai/⁠ Twitter: ⁠@imbue_ai

Seth Lazar is a professor of philosophy at the Australian National University, where he leads the Machine Intelligence and Normative Theory (MINT) Lab. His unique perspective bridges moral and political philosophy with AI, introducing much-needed rigor to the question of what will make for a good and just AI future. Generally Intelligent is a podcast by Imbue where we interview researchers about their behind-the-scenes ideas, opinions, and intuitions that are hard to share in papers and talks. About Imbue Imbue is an independent research company developing AI agents that mirror the fundamentals of human-like intelligence and that can learn to safely solve problems in the real world. We started Imbue because we believe that software with human-level intelligence will have a transformative impact on the world. We’re dedicated to ensuring that that impact is a positive one.We have enough funding to freely pursue our research goals over the next decade, and our backers include Y Combinator, researchers from OpenAI, Astera Institute, and a number of private individuals who care about effective altruism and scientific research.Our research is focused on agents for digital environments (ex: browser, desktop, documents), using RL, large language models, and self supervised learning. We’re excited about opportunities to use simulated data, network architecture search, and good theoretical understanding of deep learning to make progress on these problems. We take a focused, engineering-driven approach to research. Website: https://imbue.com/ LinkedIn: https://www.linkedin.com/company/imbue-ai/ Twitter: @imbue_ai

Tri Dao is a PhD student at Stanford, co-advised by Stefano Ermon and Chris Re. He’ll be joining Princeton as an assistant professor next year. He works at the intersection of machine learning and systems, currently focused on efficient training and long-range context. About Generally Intelligent  We started Generally Intelligent because we believe that software with human-level intelligence will have a transformative impact on the world. We’re dedicated to ensuring that that impact is a positive one.   We have enough funding to freely pursue our research goals over the next decade, and our backers include Y Combinator, researchers from OpenAI, Astera Institute, and a number of private individuals who care about effective altruism and scientific research.   Our research is focused on agents for digital environments (ex: browser, desktop, documents), using RL, large language models, and self supervised learning. We’re excited about opportunities to use simulated data, network architecture search, and good theoretical understanding of deep learning to make progress on these problems. We take a focused, engineering-driven approach to research.   Learn more about us Website: https://generallyintelligent.com/ LinkedIn: linkedin.com/company/generallyintelligent/  Twitter: @genintelligent

Jamie Simon is a 4th year Ph.D. student at UC Berkeley advised by Mike DeWeese, and also a Research Fellow with us at Generally Intelligent. He uses tools from theoretical physics to build fundamental understanding of deep neural networks so they can be designed from first-principles. In this episode, we discuss reverse engineering kernels, the conservation of learnability during training, infinite-width neural networks, and much more. About Generally Intelligent  We started Generally Intelligent because we believe that software with human-level intelligence will have a transformative impact on the world. We’re dedicated to ensuring that that impact is a positive one.   We have enough funding to freely pursue our research goals over the next decade, and our backers include Y Combinator, researchers from OpenAI, Astera Institute, and a number of private individuals who care about effective altruism and scientific research.   Our research is focused on agents for digital environments (ex: browser, desktop, documents), using RL, large language models, and self supervised learning. We’re excited about opportunities to use simulated data, network architecture search, and good theoretical understanding of deep learning to make progress on these problems. We take a focused, engineering-driven approach to research.   Learn more about us Website: https://generallyintelligent.com/ LinkedIn: linkedin.com/company/generallyintelligent/  Twitter: @genintelligent

Bill Thompson is a cognitive scientist and an assistant professor at UC Berkeley. He runs an experimental cognition laboratory where he and his students conduct research on human language and cognition using large-scale behavioral experiments, computational modeling, and machine learning. In this episode, we explore the impact of cultural evolution on human knowledge acquisition, how pure biological evolution can lead to slow adaptation and overfitting, and much more. About Generally Intelligent  We started Generally Intelligent because we believe that software with human-level intelligence will have a transformative impact on the world. We’re dedicated to ensuring that that impact is a positive one.   We have enough funding to freely pursue our research goals over the next decade, and our backers include Y Combinator, researchers from OpenAI, Astera Institute, and a number of private individuals who care about effective altruism and scientific research.   Our research is focused on agents for digital environments (ex: browser, desktop, documents), using RL, large language models, and self supervised learning. We’re excited about opportunities to use simulated data, network architecture search, and good theoretical understanding of deep learning to make progress on these problems. We take a focused, engineering-driven approach to research.   Learn more about us Website: https://generallyintelligent.com/ LinkedIn: linkedin.com/company/generallyintelligent/  Twitter: @genintelligent

Ben Eysenbach is a PhD student from CMU and a student researcher at Google Brain. He is co-advised by Sergey Levine and Ruslan Salakhutdinov and his research focuses on developing RL algorithms that get state-of-the-art performance while being more simple, scalable, and robust. Recent problems he’s tackled include long horizon reasoning, exploration, and representation learning. In this episode, we discuss designing simpler and more principled RL algorithms, and much more. About Generally Intelligent  We started Generally Intelligent because we believe that software with human-level intelligence will have a transformative impact on the world. We’re dedicated to ensuring that that impact is a positive one.   We have enough funding to freely pursue our research goals over the next decade, and our backers include Y Combinator, researchers from OpenAI, Astera Institute, and a number of private individuals who care about effective altruism and scientific research.   Our research is focused on agents for digital environments (ex: browser, desktop, documents), using RL, large language models, and self supervised learning. We’re excited about opportunities to use simulated data, network architecture search, and good theoretical understanding of deep learning to make progress on these problems. We take a focused, engineering-driven approach to research.   Learn more about us Website: https://generallyintelligent.com/ LinkedIn: linkedin.com/company/generallyintelligent/  Twitter: @genintelligent

Jim Fan is a research scientist at NVIDIA and got his PhD at Stanford under Fei-Fei Li. Jim is interested in building generally capable autonomous agents, and he recently published MineDojo, a massively multiscale benchmarking suite built on Minecraft, which was an Outstanding Paper at NeurIPS. In this episode, we discuss the foundation models for embodied agents, scaling data, and why prompt engineering will become irrelevant.   About Generally Intelligent  We started Generally Intelligent because we believe that software with human-level intelligence will have a transformative impact on the world. We’re dedicated to ensuring that that impact is a positive one.   We have enough funding to freely pursue our research goals over the next decade, and our backers include Y Combinator, researchers from OpenAI, Astera Institute, and a number of private individuals who care about effective altruism and scientific research.   Our research is focused on agents for digital environments (ex: browser, desktop, documents), using RL, large language models, and self supervised learning. We’re excited about opportunities to use simulated data, network architecture search, and good theoretical understanding of deep learning to make progress on these problems. We take a focused, engineering-driven approach to research.   Learn more about us Website: https://generallyintelligent.com/ LinkedIn: linkedin.com/company/generallyintelligent/  Twitter: @genintelligent

Sergey Levine, an assistant professor of EECS at UC Berkeley, is one of the pioneers of modern deep reinforcement learning. His research focuses on developing general-purpose algorithms for autonomous agents to learn how to solve any task. In this episode, we talk about the bottlenecks to generalization in reinforcement learning, why simulation is doomed to succeed, and how to pick good research problems.

Noam Brown is a research scientist at FAIR. During his Ph.D. at CMU, he made the first AI to defeat top humans in No Limit Texas Hold 'Em poker. More recently, he was part of the team that built CICERO which achieved human-level performance in Diplomacy. In this episode, we extensively discuss ideas underlying both projects, the power of spending compute at inference time, and much more.

Sugandha Sharma is a Ph.D. candidate at MIT advised by Prof. Ila Fiete and Prof. Josh Tenenbaum. She explores the computational and theoretical principles underlying higher cognition in the brain by constructing neuro-inspired models and mathematical tools to discover how the brain navigates the world, or how to construct memory mechanisms that don’t exhibit catastrophic forgetting. In this episode, we chat about biologically inspired neural architectures, how memory could be implemented, why control theory is underrated and much more.