™Why a 4-Billion-Point Kaggle Dataset Still Couldn't Beat Chance — Ian Quigley, Leash Bio (Latent Space)
Latent Space podcast episode of Oct 14, 2026, 1:49:56. The host interviews Ian Quigley of Leash Bio on how DNA-encoded libraries produce high-throughput protein–small-molecule interaction data; the BELKA Kaggle competition (about 4B measurements, ~2,000 teams) where no team beat chance on an unseen
Key points Tap a time to jump
Introduction: Ian Quigley and Leash Bio
How DNA-encoded libraries work
Barcodes and controls to catch DNA-binding false positives
Designing the BELKA Kaggle competition: ~4B measurements, hidden evaluations
What the Kaggle results taught: no team beat chance among ~2,000
Clever Hans: predicting a molecule's chemist from its structure
Language models and the scientific-reasoning sandbox Minos
From DEL hits to mouse PK/PD in six months on an inflammation program
Comments
More to watch
All →- 1:21:05Robots That Predict the Future: Dyna-2 & Physical AI — Jason Ma, Dyna Robotics
- 12:09Gemini 4 Argon: 10 incredible AI demos (1M token output in action)
- 1:33:00All-In Podcast E292: Is Claude Conscious? OpenAI's Math Backlash, France Riots
- 32:41The Bitter Lesson of Biology: Why Scaling AI Isn't Enough — Pushmeet Kohli (DeepMind) & Sal Candido (Biohub)
Latest news
All →- Oct 15xAI Releases Grok Voice 3.0, Topping Speech-to-Speech Benchmarks at Unchanged Price
- Oct 15Google Opens Frontier Models to 200 European Scientists with $3M in Compute Credits
- Oct 15Australia's Apate Sends Thousands of AI Agents Undercover to Fight Scammers
- Oct 15Figma Launches Specs, a Design-Rule System for AI Agents
- Oct 15Cursor vs Windsurf: AI Code Editors Face Off
- Oct 15Anthropic Reveals $16.8B Q3 Revenue Ahead of a Potential $2T IPO
Today
Oct 15 Thursday- Brief
- Quiz
- Call