Cody Coleman, co-founder and CEO of Coactive AI, explains how they leverage modern data, systems, and machine learning techniques for their multimodal asset platform and visual search tools. They discuss techniques like active learning and core set selection, and how they drive efficiency throughout the machine learning lifecycle. Cody also shares how Coactive uses multimodal embeddings for visual search and the infrastructure optimizations they've implemented to scale their systems. They conclude with advice for entrepreneurs and engineers in generative AI technologies.

Kyle Roche, founder and CEO of Griptape, discusses patterns and middleware for LLM applications. Topics include off prompt data, pipelines, Griptape's Python-based middleware stack, drivers, memory management, rule sets, DAG-based workflows, and role-based retrieval methods.

Prem Natarajan, chief scientist at Capital One, discusses AI access and inclusivity as technical challenges. They explore issues of bias, class imbalances, and the integration of research initiatives. They discuss foundation models for financial data and the importance of data quality and federated learning. They also touch on the use of LLMs and deep learning in various applications and the importance of research and impact for a bank.

Jay Emery, director of technical sales & architecture at Microsoft Azure, discusses the challenges of building LLM-based applications, including security, privacy, and performance concerns. They explore techniques like prompt tuning and fine-tuning, as well as use cases for Azure Machine Learning prompt flow and Azure ML AI Studio. Strategies for improving performance with Azure OpenAI GPT models are also discussed.

Richard Zhang, senior research scientist at Adobe Research, discusses perceptual metrics and the LPIPS paper, detection tools for fake visual content, and data attribution and concept ablation in generative AI. They explore challenges in visual generative AI, improving perceptual metrics and loss functions, controllability of generative AI systems, addressing challenges in the ecosystem, understanding the connection between synthesized images and training data, and concept ablation and opting out in the contributor domain.

Heather Gorr, MATLAB product marketing manager at MathWorks, discusses deploying AI models to hardware devices and embedded systems. Topics include factors to consider during data preparation, device constraints and latency requirements, modeling needs like explainability and robustness, verification and validation methodologies, and adapting MLOps techniques. Anecdotes about embedded AI deployments in automotive and oil & gas industries are shared.

Yoshua Bengio discusses the catastrophic risks of AI misuse. Topics include manipulation, disinformation, harm and power concentration. They explore risks associated with achieving human-level competence in AI, and challenges of defining agency and sentience. Solutions include safety guardrails, national security, bans on uncertain safety, and governance-driven AI systems.

Miriam Friedel, senior director of ML engineering at Capital One, discusses delivering ML tools in regulated environments, creating a culture of collaboration, leveraging open-source tools, building a 'unicorn' team, and the future of MLOps and enterprise AI. Examples include Rubicon for experiment tracking and Kubeflow pipeline components for scaling models.

Riley Goodside, staff prompt engineer at Scale AI, explores LLM capabilities and limitations, prompt engineering, autoregressive inference challenges, and the application of mental models in improving ChatGPT's performance.

Sara Hooker, Director at Cohere and head of Cohere For AI, joins the podcast to discuss challenges with multilingual models, the Mixture of Experts technique, common language between ML researchers and hardware architects, impact and emotional connection of language models, benefits and safety concerns of universal models, and the significance of grounded conversations in AI model development.