scikit-learn creator on agentic data science: what to delegate to agents and how to verify
hugobowne · x · 2026-10-05
Vanishing Gradients host Hugo Bowne-Anderson talks with Gaël Varoquaux (scikit-learn, Inria, Probabl), Shipra Arora (Bain) and Luca Fiaschi (PyMC Labs) about the state of agentic data science: what to delegate and how to verify.
- Agents can write code, run analysis and produce convincing explanations — but how do you check they answered the right question and used data appropriately?
- Verification splits roles: programmatic checks, agent critique of agent work, and human judgment; decision risk shapes which checks can be delegated.
- Varoquaux: "Agents are like coffee. They can get you to do stupid things faster with more energy."
- Also covers common agent mistakes, which checks can themselves be delegated, and how workflows must evolve as models improve. 59 min episode.
Related event: scikit-learn Founder Discusses Agentic Data Science on Podcast(2 posts)→
More from coding & agent
- Qwen Flash Next Q4 hits 17.5 tks on Mac Mini M5 with carousel expert streaming and dual-SSD tricks — turtleninja99 · 2026-10-05
- Agents can test your code in ways that took weeks to build before — and catch human-introduced bugs — remilouf · 2026-10-05
- Viral meme: AI agents with full computer access will indeed modify your files — smtabatabaie · 2026-10-05
- Ethan Mollick: We need team email systems built for humans and AI agents to share — emollick · 2026-10-05
- 7 trade-offs every AI engineer makes, from accuracy-vs-speed to privacy-vs-personalization — goyalshaliniuk · 2026-10-05
- GPT-6.1 Sol + Blender MCP turns Gundam photos into a textured 3D model in under 10 minutes — sidahuj · 2026-10-05