Structure precedes quantity.

I started out in experimental chemistry because I was in love with science. Over time, I realized that some disciplines move faster than others because the time between hypothesis and verification can be orders of magnitude shorter. That realization pulled me toward computation and theory.

Then a second realization followed: in computational science, the bottleneck is often not ideas but infrastructure. Bad software, brittle pipelines, weak benchmarks, and ugly workflows quietly determine which questions you can afford to ask and whether you can trust the answers. So these days a surprising amount of my scientific work consists of writing code.

AI has made stunning progress in language and protein structure prediction, yet we have not seen a commensurate improvement in fields like drug discovery. I do not think the answer is merely better architectures or more data. More often, we are asking models to predict properties like binding affinity or toxicity directly, when those capabilities should emerge from a model that first learns the structural problems of chemistry: reactivity, synthesis, and the logic of molecular transformation. This newsletter is where I write about that thesis, the infrastructure it demands, and systems like SynthArena, RetroCast, and the broader isChemist Protocol.