Why the Best AI Agents Abandon Dead Ends Fast
Ian Quigley of Leash Bio finds that top AI agents succeed by abandoning bad molecular neighborhoods instead of rationalizing weak results.
An Eval for AI Scientists That Knows the Answer Key - Ian Quigley, Leash Bio
With Brandon, RJ, Ian Quigley · Sunday, November 1, 2026
Ian Quigley of Leash Bio discusses Minos, a framework of pre-measured biological environments that lets AI agents explore chemical spaces with zero physical latency. By benchmarking language models against completely mapped ground-truth datasets, Leash evaluates scientific reasoning, explore-versus-exploit policies, and structure-activity relationship navigation. The discussion covers model performance tiers, the evolution of LLM chemical comprehension, and the future of amortized lab-in-the-loop systems for reinforcement learning.
Key takeaways
- Leash Bio built Minos, an evaluation framework using pre-measured chemical spaces to benchmark scientific AI agents without physical wet-lab delays. Read more →
- Conventional lab-in-the-loop drug discovery is bottlenecked by physical synthesis and assay cycles, which creates tiny datasets that lack counterfactual density. Read more →
- Leash Bio built Minos to test automated chemistry agents inside environments mapped in advance with DNA-encoded libraries. Read more →
- In late 2023, Leash Bio attempted supervised fine-tuning on OpenAI and Anthropic models because frontier LLMs failed basic SMILES string manipulations. Read more →
4 articles from this episode
Ian Quigley Scales Leash Bio to 20 Biological Environments a Day
Ian Quigley of Leash Bio pre-measures chemical spaces so AI agents can train on wet lab data with zero physical latency.
Why Leash Bio Pre-Measures Lab Data for AI Agents
Ian Quigley explains how Leash Bio uses pre-measured chemical spaces to benchmark AI agents across 90,000 rollouts without lab delays.
Frontier Models Gained Native Chemistry Fluency in 2024
Ian Quigley explains how frontier models learned SMILES strings in early 2024, enabling AI agents to design molecules without fine-tuning.
More Latent Space episodes
- The Recommendation Engine for Commodity Traders — Anekha Sokhal, MoshiOct 25 · 4 articles
- The Programmable Data Center — Evan Conrad, SF ComputeOct 25 · 6 articles
- AI in NYC: How Tenex Builds an AI-Native Engineering Team — Alex Lieberman & CJ HessOct 25 · 4 articles
- Why AI Compute Comes in Rectangles — Erik Bernhardsson, ModalOct 25 · 5 articles