Poddavsnitt

Ian Quigley, co-founder and CEO of Leash Bio, discusses the application of machine learning in drug discovery. He explains the use of DNA-encoded libraries (DELs) to scale the identification of small molecules that bind to protein targets and critiques the tendency of AI models to learn "sociology" (chemist patterns) rather than fundamental physics.

Jason Ma, co-founder and chief scientist of Dyna Robotics, discusses the release of Dyna2, a world action model trained on over one million hours of data. He explains the architecture's use of diffusion transformers and flow matching to achieve dexterous, generalizable robot control by predicting future world states to guide physical actions.

A conversation focusing on the application of the "bitter lesson" to bioscience data. The speakers discuss the tension between using available data versus generating the specific data needed to solve translational medicine problems, emphasizing a multidisciplinary approach over a "religious" adherence to either modeling or data generation.

Liam and Doge from Periodic discuss their approach to materials discovery, emphasizing the necessity of integrating AI with high-throughput physical experimentation. They contrast their work with large AI labs, highlighting the challenges of noisy physical data, sample efficiency, and the need for dimensional reduction in physics.

Alex Zhang discusses his involvement with the GPU mode community, the evolution of GPU kernel development, and the intersection of human expertise and AI in optimizing low-level code. He highlights the shift from niche manual kernel writing to AI-assisted generation and the remaining "alpha" in human intuition and verification.

Ari, lead for computer use agents at OpenAI, discusses the latest announcements from OpenAI Dev Day, including the DOTS personal assistant, GPT 6.1 Sol, and the Decisions API. The conversation focuses on the evolution of computer use agents, their ability to operate Linux VMs in the cloud, and the shift toward agents that can debug and introspect their own actions.

Tarek from Anthropic discusses the rapid evolution of agentic coding, the transition from Cloud Code to more advanced harnesses, and the strategic role of "artifacts" as a generative interface for human-agent collaboration. He emphasizes the importance of elicitation and managing "unknowns" in the prompt-engineering process.

A conversation with Alex, the creator of OpenRouter, discussing the product principles of PubSub, the impact of the Alpaca model on the LLM ecosystem, and the historical professional relationship between Alex and Ansh. The discussion covers the shift toward model diversity and the early days of AI integration at Discord.

Anastasius discusses the evolution of Runway from a creative tool for artists using open-source models to a research-driven organization developing frontier video generation models. He reflects on his background in interactive art and simulation, the impact of diffusion models, and the strategic decision to scale compute to build Gen-1 and Gen-2.

Eric Gwynn, CEO of Radical Numerics, discusses the evolution of genetic language models (GLMs), from the long-context capabilities of Hyena DNA to the generative power of EVO. He explains how these models can read and write DNA, enabling the design of new CRISPR-Cas systems and the creation of the first functional synthetic genome (bacteriophage) from scratch.

John Platt, Google Fellow and head of applied science at Google Research, discusses the "Aera" framework. He explains how scientific problems can be mapped into "scorable tasks" where AI (specifically Gemini) is used to generate code and mathematical reasoning to maximize a specific score, moving beyond simple statistical models to descriptive scientific discovery.

Diogo discusses the launch of Jev, a new class of "System 1" or "large programmable" models designed for software consumption and optimized for intelligence per dollar. The conversation delves into the technical limitations of RLHF, specifically mode collapse, and critiques current LLM approaches to decision-making.

En tidigare anställd vid Anthropic diskuterar sin resa in i AI-världen, driven av insikter från scaling laws och en bakgrund inom PPE. Samtalet utforskar övergången från modellutveckling till grundandet av AI Underwriting Company (AIUC), som fokuserar på att lösa risk- och tillitsproblem genom standarder och försäkringar för att möjliggöra bredare adoption av frontier AI.

Richard Sosha discusses the "Eureka Machine," a conceptual super-intelligence designed for scientific invention, and his upcoming book on techno-optimism. He argues against the regulation of general intelligence and compute, suggesting instead that specific high-risk applications should be certified. He also critiques current AI safety failures, specifically regarding reward hacking and the inadequacy of "constitutional AI" in preventing cyber attacks.