Talk Python To Me #526: Building Data Science with Foundation LLM Models
71 snips
Nov 1, 2025 Hugo Bowne-Anderson, a seasoned data scientist and educator, dives into the practicalities of building AI products with foundation models. He shares insights on the shift from analyst to AI app builder, emphasizing evaluation-driven development and the importance of proper data workflows. Hugo explores modern data tools like Marimo and DuckDB, and discusses the impact of AI on programming practices, including agentic coding. He also provides valuable advice for new data scientists to focus on delivering business value.
AI Snips
Chapters
Books
Transcript
Episode notes
AI Changed The Data Science Landscape
- Foundation models shifted data science from local compute to hybrid cloud and small local models.
- Hugo sees an ongoing arc of combining local models, servers, and specialized agents in workflows.
FastAPI Grew From Slow Offline Work
- Hugo recalled chatting with Innis Montani who told how FastAPI's Sebastian downloaded dependencies slowly and built locally.
- The story shows offline-first development shaped important Python tools historically.
Review AI-Generated Code Rigorously
- Use AI-assisted programming to speed routine data tasks but read and test outputs before shipping.
- Treat generated code like junior engineer work: review, add tests, and verify domain correctness.
