Mark Humphries, Professor of History at Wilfred Laurier University, discusses his use of AI technology in the field of history, including his experience with GPT-3.5. He explores the potential of AI in transcribing and digitizing historical documents, uncovering hidden patterns, and improving productivity in historical research. They also discuss the limitations and challenges of using AI models for historical document retrieval and the impact of AI on writing skills. The conversation concludes with a discussion on the long-term consequences of AI and the immediate threat it poses.

Yohei Nakajima, creator of BabyAGI and GP at Untapped Capital, discusses the opportunity for AI to strengthen human understanding, insights on investing in AI, automating tasks, building internal tools, exploring different projects and technologies, using AI for workflow optimization and idea generation, prototyping an autonomous startup founder, advancements in AI agents, AI video platforms, regulating AI progress, and automating human work.

Anton Troynikov, cofounder of Chroma, discusses the importance of keeping the retrieval-augmented generation (RAG) loop in house, Chroma being in “wartime” mode, and the challenge of incorporating data never before in a database. They delve into the shift from agents to RAG in the AI space, scaling challenges of a retrieval system, best practices for improving search results, increasing data and creating business value, improving the stack and solving RAG problems, the importance of evaluations and trade-offs, searching and retrieving data, and model reactions and generalizability.

AI researcher Trey Kollmer returns to discuss reducing global compute needs by 10%, the benefits of analogical prompting, and how compressive historical records can increase LLM memory and retention abilities.

AI researcher Trey Kollmer joins Nathan Labenz to discuss Microsoft's Self-Taught Optimizer research, H100 chips supercharging GPT-4 level compute, LLM representation of space and time, OpenAI's revenue growth and potential price drop, GPT-4 performance and knowledge cutoff, LOMA2's impact on studying complex concepts, and the importance of loss functions in AI system design.

The podcast discusses AI jailbreaks, including the Calvin and Hobbes case. It covers monitoring and controlling model behavior, detecting and controlling middle layer representations, language model performance in different languages, polysementicity in neural networks, detecting deception in AI networks, and new training techniques.

James Yu, Founder of Sudowrite, an AI writing tool, discusses how he started Sudowrite after using GPT-3 for his own fiction writing. They explore the evolution of Sudowrite, advancements in AI models for deeper narrative understanding, pricing and economic challenges of AI models for writing, and the future of AI in creativity and entertainment.

Swyx of Latent Space and Nathan, former college dorm mate of Mark Zuckerberg, discuss AI engineers, tools for developers, and tensions between OpenAI and developers. They also touch on the emergence of AI engineers, finding and joining AI communities, progress in AI, and differences between open source and close source companies. They explore the potential dominance of OpenAI and upcoming events in the AI field.

Aravind Srinivas, CEO of Perplexity AI, discusses Perplexity's growth to millions of queries per day and their competition against Google. They also talk about the unique positioning of their product and the limitations of AI models.

Zvi Mowshowitz, author of Don't Worry About the Vase, talks about OpenAI's recent developments, Amazon's investment, and the Roman Empire. They discuss topics like perplexity, deepfakes, software bundling, and the challenges of training AI models on ongoing data feeds. The episode also explores language model performance and pricing, the code interpreter's revolutionary capabilities, and defending against misinformation. They touch on the reversal curse in training language models and discuss Amazon's investment and non-standard investment terms.