How we work
Built for each science.
Held to one standard.
A membrane free-energy calculation, a sweep of a million genomes and a crystal stability screen do not yield to the same machinery, and we do not pretend they do. Every domain gets its own harness, built around its own physics, data and controls. What they share is the part that decides whether a result can be trusted: the same rules of evidence, and the compute to apply them at scale.
The idea
Why each science gets its own harness
A single pipeline stretched over every field looks efficient on a slide. In practice it hides the decisions that settle whether a result is real, because those decisions are different in every field.
The physics is different
Molecular dynamics, genome-scale sequence search, density-functional theory and cohort statistics have almost nothing in common below the surface. Each harness is written around the method its field actually trusts.
The controls are different
What counts as a positive and a negative control is set by each field's own literature. We build them from primary sources for every domain and pin them, instead of borrowing a generic set.
The ways to be fooled are different
Thin statistics, database fields that disagree with their own papers, a candidate that turns out to be a published lookalike. Each harness is audited against the specific traps of its field.
The standard of proof is not
Negative controls, three-way verdicts, prior-art checks and provenance on every number apply everywhere. The code is separate; the rules a result must survive are the same.
When one harness teaches us a new way to be wrong, the lesson is written down as a rule and checked into the others by hand. We do not assume two pipelines agree until we have compared them.
For example
What goes in, what comes out
Drugs
Molecules in, binding and permeation energies out, ranked into the handful worth synthesising.
Proteins
A target surface in, sequences and backbones out, filtered before a single plasmid is ordered.
Biomarkers
Multi-omic cohorts in, signatures out, held to a site the model has never seen.
Materials
A property specification in, compositions and structures out, screened against physics before a furnace is switched on.
Layers
What we build with
Foundation Models
We use and develop foundation models capable of understanding molecular, biological, and scientific data.
Generative Discovery
Instead of only predicting the properties of existing candidates, our models propose entirely new ones, optimized toward several objectives at once.
Simulation and Prediction
Physics and learned models estimate properties before expensive experiments begin.
Autonomous Research Agents
Agents run parts of the scientific workflow end to end, and keep a research memory across cycles.
Large-Scale Compute
Discovery needs scale. AI3 Discovery is built on the AI infrastructure and large-scale model serving expertise of AI3.
Discovery loop
AI × Simulation × Experiment
Each experiment creates new data. Each new data point improves the next discovery cycle. The result is a research program that compounds.
Understand
Collect scientific knowledge and biological data.
Generate
AI proposes new hypotheses and candidates.
Predict
Models estimate likely properties and outcomes.
Select
The most promising candidates are prioritized.
Validate
Candidates are tested in simulation or in the laboratory.
Learn
Results return to the model, and the loop repeats.
Engineering
Built to be checked
A discovery system is only as good as the results it refuses to accept. Our pipelines are designed so that a wrong answer is caught by the system, not by a reader.
Negative controls by default
Every gate runs against a deliberately broken system. A gate that cannot fail is not a gate.
Three-way verdicts
Pass, fail, and indeterminate. A test that passes because the data is thin is reported as indeterminate, together with the smallest effect it could have resolved.
Cross-vendor adjudication
Independent model families score the same candidate. Disagreement raises priority instead of being averaged away.
Provenance on every number
Each reported value carries the system, force field, and run that produced it. Numbers without a track are not quotable.
Bring us a problem worth the computation.
Joint programs, platform access, and data collaborations.