andrew blevins

This account is verified. Learn more

@Andrewdblevins

CTO Leash Biosciences

Salt Lake City, UT Joined April 2009

1K Following 545 Followers 354 Posts

Latest posts 5

Pinned post

andrew blevins @Andrewdblevins Oct 14

Teaching AI to cure cancer will require lab-in-the-loop data. But models learn poorly from actions they wouldn’t have chosen. This is why we premeasure a very large space of possible experiments, so models can learn to run them on-policy. New paper: github.com/Leash-Labs/min…

andrew blevins @Andrewdblevins Oct 19

Some people are asking why we go through all this when we could just train the model on all the data, or jump straight to our best hit in the dataset.
We don't want to give the model a fish, we are trying to teach it to fish.

Ian Quigley @allmeasures Oct 19

Last week we released Minos, our framework of premeasured experimental worlds for agentic science.

Minos is not for obtaining the best scientific result, it's for understanding how AI agents make scientific decisions and maybe help them make better ones.

Let's dig in!

1/

View quoted post on X

andrew blevins @Andrewdblevins Dec 19, 2025

I scored 4/5 on the Chemist Style Quiz. Can you spot the outlier? leash-labs.github.io/chemist-style-…

leash-labs.github.io Chemist Style Quiz Spot the outlier molecule and test your ability to recognize chemist style.

andrew blevins @Andrewdblevins Dec 17, 2025

We created a leaderboard ranking chemists on ChEMBL by how distinct and recognizable their molecules are.

andrew blevins @Andrewdblevins Dec 17, 2025

A very simple ML model can look at a molecule and identify the chemist that made it in ChEMBL. 60% top-5 accuracy on a 1815 class classification problem. ML practitioners should be wary of data leaks from models memorizing the intentions of the chemists making the molecules.