Delta Podcast

Delta Institute

Supporting exceptional researchers and engineers, from academia to industry and beyond.

  1. 28 May

    Ep. 57: Kevin Wang, NeurIPS Best Paper Author and OpenAI Researcher

    Kevin Wang is the first author of the NeurIPS 2025 Best Paper, titled "1000 Layer Networks for Self-Supervised RL: Scaling Depth Can Enable New Goal-Reaching Capabilities". He's currently a researcher at OpenAI, where he works on RL/reasoning. Before coming to OpenAI, Kevin studied CS at Princeton.Delta Institute (deltainstitutes.org) supports exceptional researchers and engineers, from academia to industry and beyond. They host technical events to bring great people together, a podcast that gives industry/academic leaders a platform to share their experiences, a small fellows program that builds a tight-knit community of exceptional people, and a grant program that provides compute/mentorship for research projects.Timestamps:00:00 Introduction00:26 Overview of the 1000 Layer Networks Paper00:42 Motivation and Background of the Research01:37 Self-Supervised Reinforcement Learning Paradigm04:16 Challenges and Innovations in Data Scaling06:23 Hindsight Experience Replay and Its Impact08:56 Classification vs Regression in Reinforcement Learning12:25 Training Stability and Architectural Components14:23 Key Results and Performance Gains17:23 Qualitative Behaviors and Representation Learning19:44 Scaling Depth and Batch Size23:06 Limits of Scaling in Reinforcement Learning23:55 Exploring Actor Loss and Layer Depth in Training24:51 Scaling Layers for Complex Tasks28:28 Challenges and Innovations in Deep Network Training30:36 Future Directions in Reinforcement Learning37:32 Personal Journey and Career Path

    46 min
  2. 28 May

    Ep. 56: Grace Li, Design Arena Co-Creator and Arcada Labs Co-Founder

    Grace Li is the co-founder of Arcada Labs (arcada.dev), creators of Design Arena, Prediction Arena, and Social Arena. Arcada's vision is to build portals that bridge AI to the real world by building real-world evaluations for things that are often hard to benchmark, like design, ability to make money in prediction markets, and ability to write posts that do well on social media. Before starting Arcada, Grace studied CS at Harvard and spent some time at Apple.Delta Institute (deltainstitutes.org) supports exceptional researchers and engineers, from academia to industry and beyond. They host technical events to bring great people together, a podcast that gives industry/academic leaders a platform to share their experiences, a small fellows program that builds a tight-knit community of exceptional people, and a grant program that provides compute/mentorship for research projects.Timestamps:00:00 Introduction00:53 Founder's Background and Journey01:45 Startup Path and Early Experiences03:53 Pivoting to Game Engine Development06:03 Scaling and Expanding Arenas09:53 Building and Measuring Model Capabilities13:14 Agent Runner and Open Source Contributions14:33 Future Vision and Customer Focus18:56 Evaluating Next JS Models: Static vs. Live Benchmarks20:20 Challenges of Crowdsourced Benchmarks21:28 Understanding User and Researcher Needs23:23 Introduction to Prediction Arena24:30 Technical Implementation and Investor Skepticism26:45 Model Performance and Future Improvements31:28 Exploring Social Arena36:08 Company Culture and Future Goals37:54 Conclusion and Final Thoughts

    38 min

About

Supporting exceptional researchers and engineers, from academia to industry and beyond.

You Might Also Like