About this time last summer, Leash Bio started an internal project to teach language models a chemistry notation called SMILES. Ian Quigley of Leash Bio said the company used supervised fine-tuning, which means training the models further on worked examples with known correct answers. Leash did this "because none of the existing frontier models were any good at it." Frontier models are the most capable systems from the leading AI labs. Quigley pointed to one test in particular. At the time, he said, none of the Anthropic or OpenAI models could convert a SMILES string into a non-canonical SMILES string very effectively.
Quigley described the project in a follow-up conversation on the Latent Space podcast, published October 26, 2026. He placed events relative to that conversation ("last summer," "around February") and gave no exact dates.
A molecule written as a walk
The conversation paused to define the notation for non-chemists: SMILES is a string language that describes a chemical precisely. Quigley explained why it is harder than it looks. Molecules can branch, form rings, or do both. A SMILES string is "basically a graph traversal across the molecule written down into a one-dimensional string." It records a walk from atom to atom and writes each step down along the way.
The walk can start at a different atom, so one molecule can have many valid SMILES strings. The Daylight SMILES documentation gives ethanol as an example: CCO, OCC and C(O)C all describe the same molecule, with parentheses marking a branch. Paired digits mark where a ring closes. Canonicalization picks one standard string out of all the valid ones, so a given molecule always gets the same string. A non-canonical string is any of the other valid versions.
Quigley said "it's nice to be fluent in transferring from one to the other." Ordinary software handles this easily, and he named the open-source package RDKit. RDKit's official guide shows a SMILES string being read into a molecular graph and written back out in canonical form by default. Combining a rootedAtAtom setting with canonical=False makes the output start from a chosen atom. That is the kind of rewrite Quigley said frontier models could not do well.
Leash still wanted the language models to do the conversion themselves. Quigley gave two reasons: it would be faster, and the models "weren't as good at using tools on their own back then." Calling an outside program like RDKit was not yet something they did well.
The February turn
"And nobody could do it," Quigley said. "And then around February, like a lot of other things, it became clear that they could start to actually work with chemistry in a meaningful way." He gave that change as one reason Leash moved in its current direction. The company now has AI agents carry out iterative tasks. Each agent looks at molecules, learns something about them and uses that to choose better molecules in the next round.
Learning, or using?
The discussion then turned to what this shows. The suggestion was that a model able to do this must be internalizing some understanding of the physical nature of molecules. Quigley agreed only "a little bit" and pointed to an experiment. In an ablation, researchers remove one piece of a system's input and watch what changes. When Leash removes the information about the molecules, "these decisions really suffer," he said. "So they're learning something."
Then he corrected himself: "Or at least they're using it. I guess that's a better way." A model's weights are the numbers set during training, and they hold what it has learned over the long term. Leash's agents work with those weights unchanged, so the agents use whatever information they are given while they work. "We're not changing weights here," Quigley said, "but they're definitely using the information as to make future choices."