On January 7th, we launched our very own podcast. Making a Mind is hosted by Dr. Danielle Perszyk, cognitive scientist and head of the human-computer interaction (HCI) team at Amazon’s AGI Lab. The podcast explores the science of intelligence through conversations with leading AI researchers, examining the scientific breakthroughs required to realize useful general intelligence. It’s designed for both technical audiences and general listeners curious about AI development—anyone interested in the minds behind the machines and how AI agents are being built to work alongside humans to augment our capacity. We’ve got a bunch of guests from our in-house AGI Lab as well as upcoming episodes featuring peers from across the AI industry.
Episode 1: A History of Modern Agents with Kelsey Szot
"Something that is 30% reliable or 60% reliable or even 80% reliable is 0% useful," Kelsey Szot, former co-founder of Adept and current product lead at the AGI Lab, explains. In this episode of Making a Mind, Kelsey traces the evolution of language models to modern agents capable of taking real-world action, reviewing a range of topics from how to solve the agent reliability problem, to generalizability, to future job transformation. You can watch the full episode with Kelsey below.
Episode 2: Building RL Gyms to Shape Agent Learning with Jason Laster
"Environments aren't a side project—they're as fundamental as the model itself," emphasizes Jason Laster, the AGI Lab's engineering lead for building RL gyms. In this episode of Making a Mind, Jason argues that training environments deserve investment equal to models and compute. In his conversation with Danielle, Jason shares the essential properties of high-fidelity simulation environments and the complexity of the world of browser training. You can watch to the full episode with Jason below. If you’re interested in learning more about RL gyms, you can also read this story from Jason about the unseen work of building AI agents.
Episode 3: Giving Agents the Ability to See with Matthew Elkherj
"Perception is the foundation everything else is built on," Matthew Elkherj explains. In this conversation, Matthew and Danielle dive into why UI understanding is fundamentally different from traditional computer vision, the distinct reliability requirements agents face when interpreting digital interfaces, and why perceptual hallucinations can sometimes be a feature rather than a bug. They also discuss the critical role synthetic gym environments play in training agents to handle the complexity and variability of real-world interfaces—and why solving perception is inseparable from solving the broader challenge of building reliable AI teammates that augment human capacity. You can watch the full episode with Matthew below.
Note: This episode was recorded in August 2025.
Episode 4: Improving Agent Reliability with Reinforcement Learning with Deniz Birlikci
"Reliability, not accuracy, is the true bottleneck for web agents," Deniz explains. In this conversation, Deniz and Danielle dive into the critical role of robust verification systems in validating agent behavior, the specific failure models that RL attempts to fix, and the extraordinary complexity of orchestrating live browsers with perception and actuation stacks. They also discuss why so much has to go right to make an agent work—and how RL is establishing the foundation for agents that can execute complex workflows reliably alongside humans in production environments.
Episode 5: Agent Learning Curriculums with Anirudh Chakravarthy
In this conversation, Ani introduces a training paradigm where two AI agents work together: one explores the web to discover tasks at the frontier of its current capabilities, while the other learns from these challenges. Ani and Danielle discuss the parallels between human development and agent learning, how agents represent a new form of embodied intelligence distinct from language models, and what this means for the future of human-AI collaboration. You can watch the full episode with Ani below.
Stay tuned for more as we publish new episodes every other week across YouTube, Amazon Music, Apple Podcasts, and Spotify.
