New structures, candidate materials and mathematical conjectures. Computation expands the search; experiments and proofs test the results.
AlphaProof solved three of five non-geometry problems from the 2024 International Mathematical Olympiad, while AlphaGeometry 2 handled one geometry problem. The combined score matched a silver medal, but required formalized statements and days of computation.

Models trained on published synthesis records classified proposed inorganic crystals and suggested preparation methods and starting materials. Strong held-out results do not replace a laboratory synthesis, especially for less common methods.

Chemma combined reaction prediction with active learning to help chemists find a ligand and solvent in 15 laboratory runs, reaching 67% isolated yield. It is one reaction study with human oversight, not a general autonomous chemist.

AlphaFold 3 extends structure prediction beyond proteins to interactions involving DNA, RNA, small molecules and antibodies. Its benchmark gains can guide experiments, but a predicted structure is not proof of binding or biological effect.

Researchers implemented an anomaly detector as decision trees on an FPGA and measured 30-nanosecond inference with low hardware use. It is a fast prototype tested on simulated collider scenarios, not evidence that new physics has been discovered.
