Andriy Burkov talks down dishonest hype and sets realistic expectations for when LLMs, if properly and critically applied, are useful. Although maybe not as AI agents.

Andriy and Kimberly discuss how he uses LLMs as an author; LLMs as unapologetic liars; how opaque training data impacts usability; not knowing if LLMs will save time or waste it; error-prone domains; when language fluency is useless; how expertise maximizes benefit; when some idea is better than no idea; limits of RAG; how LLMs go off the rails; why prompt engineering is not enough; using LLMs for rapid prototyping; and whether language models make good AI agents (in the strictest sense of the word).

Andriy Burkov holds a PhD in Artificial Intelligence and is the author of The Hundred Page Machine Learning and Language Models books. His Artificial Intelligence Newsletter reaches 870,000+ subscribers. Andriy was previously the Machine Learning Lead at Talent Neuron and the Director of Data Science (ML) at Gartner. He has never been a Ukrainian footballer.

Related Resources

• The Hundred Page Language Models Book: https://thelmbook.com/
• The Hundred Page Machine Learning Book: https://themlbook.com/
• True Positive Weekly (newsletter): https://aiweekly.substack.com/

A transcript of this episode is here.