Beyond the Hype: What AI Actually Can (and Can't) Do Jodie Burchell & Michelle Frost

The podcast features an interview with Jodie Bertrell, a JetBrains advocate who transitioned from clinical psychology and biostatistics to data science and NLP, emphasizing ethical AI and scientific rigor since 2016. The discussion explores AIs evolution, focusing on generative models based on transformer architectures and their dominance in current conversations, while highlighting gaps in terminology and definitions of AI over time. It addresses ethical challenges, such as debunking misconceptions about large language models (LLMs), the intersection of psychology and data science in shaping responsible AI, and the importance of addressing bias as a multifaceted issue affecting generalizability, performance, and liability. The conversation also critiques the overemphasis on generative AI at the expense of foundational machine learning knowledge, stressing that even "paint-by-numbers" tasks require robust data preprocessing, domain expertise, and awareness of biases in data collection.

The episode delves into the accessibility paradox, noting that while pre-trained models and tooling make AI seem approachable, their complexity creates barriers for practitioners. It underscores the necessity of classical machine learning fundamentals for effective AI use, particularly in tasks like vector search and embeddings, where domain expertise and data quality are critical. Ethical considerations remain central, with discussions on the complexity of defining fairness in AI and the challenges of aligning governance frameworks with practical product development. The conversation also examines historical patterns in AI development, drawing parallels to past "AI summers" and "winters," while cautioning against conflating narrow AI advancements (like GPT models) with the elusive goal of artificial general intelligence (AGI). The dialogue concludes by emphasizing the current utility of LLMs in specific software development tasks, such as improving code understanding and productivity, while advocating for realistic expectations and a focus on practical applications over overambitious goals.