Shengran Hu, a PhD student at the University of British Columbia, dives into the world of Automated Design of Agentic Systems (ADAS). He discusses how ADAS creates agentic designs using large language models and explores emergent behaviors that arise in these systems. The conversation also highlights the balance between exploration and exploitation in design iterations and the parallels between AI systems and human society. Finally, Shengran explains the practical applications of ADAS in optimizing intelligent agents for real-world tasks.

In this episode, Peter van der Putten, director of the AI Lab at Pega and assistant professor at Leiden University, dives into the European AI Act's ethical principles and its global impact. He discusses the challenges of applying fairness metrics in AI, emphasizing the need for a risk-based regulation approach. The conversation highlights the importance of closing bias gaps in automated decision-making, particularly within finance. Listeners will gain insight into the balance between ethical data practices and real-world AI applications.

Harrison Chase, co-founder and CEO of LangChain, dives into the future of agentic systems and LLM frameworks. He elaborates on the 'spectrum of agenticness' and the vital role of cognitive architectures in real-world applications. The conversation touches on the challenges of deploying agentic systems and emphasizes the need for robust observability tools. Harrison shares insights on the evolution of Retrieval-Augmented Generation (RAG) and innovative prompting strategies, all while exploring how to transition LLM applications from prototype to production.

Siddhika Nevrekar, head of AI Hub at Qualcomm Technologies, delves into the realm of on-device AI. She highlights the shift from cloud to local device inferencing, discussing the hardware challenges developers face, including the need for powerful systems and collaboration among platforms like ONNX and TFLite. The conversation emphasizes the importance of rigorous testing across diverse devices and key metrics such as accuracy and power consumption. Nevrekar also introduces Qualcomm's AI Hub, aimed at simplifying testing and optimizing AI models for a range of applications.

Ashley Edwards, a technical staff member at Runway, dives into the cutting-edge Genie system that crafts generative interactive environments for reinforcement learning. She explains the innovative capabilities of Genie to learn from unstructured video data, facilitating interaction in 2D games. The discussion highlights core components like the latent action model and video tokenizer, as well as challenges in model integration. Edwards also touches on the implications of these technologies for education and industry, showcasing the future of video generation and interactivity.

Marius Memmel, a PhD student from the University of Washington, dives into the intricate world of sim-to-real transfer in robotics. He shares insights from his work on ASID, a framework for autonomous simulation model generation, and URDFormer, which aids in creating realistic environments. The conversation touches on challenges robots face in cluttered spaces, the significance of Fisher information for optimizing trajectories, and the innovative strategies necessary for bridging the gap between simulated and real-world robotic applications.

Hamel Husain, founder of Parlance Labs, discusses building real-world LLM products. Topics include fine-tuning LLMs, challenges in development, open-source tools like Axolotl, LoRA adapters, model optimization, systematic evaluation techniques, and the importance of data generation and curation.

Albert Gu, an assistant professor at Carnegie Mellon University, discusses his research on post-transformer architectures for multi-modal foundation models. The conversation covers the efficiency of attention mechanisms, strengths and weaknesses of transformers, tokenization in transformer pipelines, hybrid models, state update mechanisms, and the evolution of foundation models in various modalities and applications.

Amir Bar, a PhD candidate at Tel Aviv University and UC Berkeley, discusses his research on visual-based learning, focusing on self-supervised object detection and analogy reasoning for computer vision tasks. He introduces 'EgoPet,' a dataset allowing animal perspective data for robotic planning. The podcast explores the limitations of current caption-based datasets, the gap between animals and AI, and training robot policies based on animal behavior.

Sarah Bird, Chief product officer of responsible AI at Microsoft, discusses Microsoft's testing and safety measures for generative AI, focusing on red teaming, automated testing, and governance. They explore risks like adversarial inputs and fairness, using lessons from past failures like 'Tay'. The importance of aligning testing with risk assessment in generative AI systems is highlighted.