rationale:

The most likely first resolution is better-than-expert prediction of reaction yields (32%), followed by a 10× better catalyst (16%) and chemical accuracy for strongly correlated systems (14%). The ranking is driven by two primary constraints: the speed at which a solution can be empirically verified, and the strict requirement that a general-purpose AI system produces the central idea. Evaluated alongside related timelines, these constraints continue to favor reaction yield prediction as the most probable initial breakthrough, without materially changing the broader outlook. While AI has made breathtaking, rapid strides in theoretical mathematics—such as solving local smoothing and refuting the Exponential Time Hypothesis Working with academic communities to build scientific ...—chemistry requires slow experimental loops, community consensus, and rigorous prospective validation.

Reaction yield prediction (32%) leads because it is the most benchmark-like problem and thus highly accessible to data-driven AI methodologies. Early general-AI precursors are already emerging: a recent comparison reported GPT-6 Astra outperforming PhD chemists on yield-threshold classification across six reaction classes Why Molecule.one's Maria predicts reaction outcomes up .... However, this was a provider-run test rather than a broadly accepted prospective blind challenge. Significant hurdles remain, particularly regarding noisy yield data and generalization, as random splits often overestimate performance and incomplete conditions limit transfer Yield Smarter, Not Harder: Good Practices for Machine .... Nonetheless, a general research agent orchestrating an autonomous lab could plausibly close this gap relatively soon given current high-throughput capacities of around 110 reactions a day OpenAI and Molecule.one report a near-autonomous AI ....

Catalyst improvement (16%) and chemical accuracy (14%) represent the middle tier, benefiting directly from the scaling of autonomous labs and algorithmic improvements. AI-run platforms have already synthesized durable iridium-free catalysts How an AI-run lab cracked open green hydrogen's catalyst ... [2609.30133] AI-guided high-throughput discovery of iridium, though the lead candidates from platforms like Lila Sciences do not yet claim a 10× improvement [2609.30133] AI-guided high-throughput discovery of iridium. Similarly, while machine-learned density functionals like Skala show strong performance on main-group chemistry, they do not yet claim chemical accuracy for strongly correlated systems Broadening access to Skala creates a faster path ... Broadening access to Skala creates a faster path ... https://huggingface.co/microsoft/skala-1.1. Both problems face strict attribution hurdles: breakthroughs must be definitively credited to the central idea of a general AI, rather than specialized Bayesian optimizers or purpose-built ML functionals acting alone.

None by 2035 (16%) remains a substantial possibility. Although general AI capability is compounding rapidly—with costs falling roughly 47% per quarter—chemistry breakthroughs are not bottlenecked by pure intelligence. The primary risks are attribution disputes between general and specialist systems, and the sociological friction of the requirement that chemists "widely accept" a result. It is highly plausible that AI achieves super-expert capabilities in chemistry by 2030, but that strict attribution rules, verification delays, or academic conservatism prevent any specific outcome from fulfilling the exact resolution criteria before January 1, 2035.

The remaining options are constrained by structural or procedural bottlenecks. A new, widely adopted reaction class (9%) could be discovered early, but wide adoption naturally takes years. Blind crystal-structure prediction (6%) is gated by the CCDC's unannounced schedule for an 8th blind test The 7th CSP Blind Test https://www.ccdc.cam.ac.uk/community/ccdc-for-the-community/partnerships-and-initiatives/csp-blind-test/, and winning requires overcoming unprecedented challenges with disorder and close polymorph energies The seventh blind test of crystal structure prediction - PMC - NIH https://doi.org/10.1107/S2052520624008679. Finally, both the nitrogenase or photosystem II mechanism (4%) and prebiotic self-replication (3%) face monumental experimental validation requirements. Current progress on FeMoco, for instance, only addresses the resting-state energy rather than the full catalytic cycle Key Chemistry Question Answered, No Quantum Computer ... Classical computational simulation of the FeMo-cofactor ... Classical computational simulation of the FeMo-cofactor ....