Genesis Molecular AI's PEARL model is doing the most interesting diffusion research in the i...
Genesis Molecular AI has a model called PEARL that is quietly doing some of the most interesting diffusion research in the industry, and none of it involves language models. On a recent Latent Space podcast, co-founder Evan Feinberg and CTO Sergey Edunov, who previously led Llama 2 training and Llama 3 pretraining at Meta, explained why small molecule drug discovery is finally producing usable results after roughly a decade of underwhelming machine learning attempts.
The unlock is not another transformer variant. It is diffusion applied to 3D molecular structure prediction. PEARL routinely clears the accuracy threshold needed to make real drug discovery decisions, which is a much harder bar than winning on public benchmarks. Edunov was blunt about this distinction on the podcast, saying the broader ML community has often been willing to call AI slop good enough when the numbers looked decent on paper. In drug discovery that gap matters, because a model that is 80 percent right about where atoms sit in space is not useful to a medicinal chemist deciding which compound to synthesize next. PEARL is being built against the standard of actual lab decisions, not leaderboard scores.
The architectural point is worth sitting with. Transformers have not produced a genuinely new primitive since the original 2017 paper. Mixture of experts, longer context windows, and various attention tweaks all scale the same core idea. Diffusion is the first new generative primitive in years that opens up a domain language models fundamentally cannot touch, which is geometry in continuous 3D space where atoms have to land in physically valid positions relative to each other. Text tokens do not have bond angles. Chat models do not know what a torsional strain feels like. Diffusion, which learned to generate images by denoising, turns out to be a natural fit for generating and refining 3D structure because the underlying math handles continuous spatial data gracefully.
For anyone building in a technical vertical, the practical lesson is that the next capability leap in your domain may not come from a bigger or better tuned chat model. It comes from matching the generative primitive to the shape of your data. Protein folding, materials, robotics motion planning, and molecular design all live in geometric spaces where diffusion or something diffusion adjacent is likely to keep pulling ahead.
What is worth watching from here is whether Genesis and a few other biotech first labs can turn PEARL's accuracy into an actual approved drug on a timeline that validates the approach commercially. That is the benchmark that will decide whether diffusion in science becomes a durable pattern or a promising detour.
Originally posted on LinkedIn.