Analysis
Where should a language model's chemistry come from: its memory, or a calculation? DeepMind's November 3 Antigravity demonstration chooses the calculation. Antigravity Cheminformatics Skills pair model orchestration with deterministic chemistry tools, so the tools execute the molecular calculations.
The workflow draws on resources including RDKit, ChEMBL and PubChem to curate compounds and analyze molecular relationships. In the walkthrough it performs matched molecular pair analysis, builds predictive models, and proposes candidate molecules and synthetic routes.
Deterministic tools do not make the whole pipeline error-free. A calculation can be repeatable and still be the wrong calculation for the question, and the model still chooses what to run. That is Silver Owl's reasoning, not a finding from the demonstration.
The demonstration is computational. Crizotinib and lorlatinib, the drugs it references, are existing examples, not discoveries by this system, and the walkthrough is not evidence of synthesis, clinical validation or approval. I would change my view if proposed molecules were made and tested in a laboratory.
Source: Agentic cheminformatics with Google Antigravity · 2026-11-03 · Primary source context · attributed account
Why it matters
Calculation software can make an AI workflow more repeatable, but a laboratory is still the test of whether a proposed molecule is usable.