rationale:

The current trajectory of AI research makes a qualifying discovery outside the excluded fields (math, theoretical computer science, and theoretical physics) highly probable, leaving only a 3% chance of no such result by 2035. General-purpose systems are already producing top-tier results in the excluded computational fields on a monthly basis. The primary question is which empirical or applied discipline will first overcome its unique physical validation bottlenecks, attribution challenges, or classification ambiguities to produce a universally recognized "top-of-the-year" result.

Biology (24%) and ML Research (24%) These two fields are the clearest frontrunners. Biology is uniquely positioned because computational discovery via genome mining or protein design requires only short wet-lab confirmation, and frontier labs are heavily investing in proprietary biological capabilities. We have already seen credible near-misses: Claude agents recently identified the ART enzyme system with minimal human prompting anthropic.com. While priority disputes the-scientist.com and lukewarm initial reception nytimes.com kept it from the annual-best threshold, functional characterization of a similar finding would clearly qualify. ML research matches biology's odds because it requires no physical lab, operates on fast computational feedback loops, and is the highest priority for frontier labs. However, ML is constrained by structural headwinds: labs are heavily incentivized to keep their most significant automated R&D breakthroughs proprietary, and attributing a discrete, publishable breakthrough to an AI rather than its developers remains difficult importai.substack.com.

Engineering/Other (15%) and Chemistry (10%) "Engineering/other" is a robust catch-all category encompassing disciplines with fast, code-based optimization, such as applied quantum circuits and cybersecurity. Recent achievements, like a 17× reduction in quantum gates for modular inversion quera.com, highlight its potential. The main risk here is classification ambiguity: a strict resolver might categorize breakthrough cryptographic or quantum results as excluded theoretical fields arxiv.org. Chemistry also offers highly promising computational avenues. While autonomous campaigns have successfully executed thousands of reactions and improved yields molecule.one, scaling these optimization loops into a recognized theoretical or applied headline discovery has not yet materialized.

Astronomy (8%), Materials Science (7%), Medicine (5%), and Social Sciences (4%) Astronomy benefits from enormous public datasets ripe for agent analysis, but recent AI-driven anomalies still require slow spectroscopic follow-up arxiv.org, and headline discoveries currently rely on specialist pipelines rather than general models. Materials science and medicine are fundamentally limited by slow physical iteration. Despite advances in autonomous wet labs, synthesizing novel materials or proving preclinical efficacy demands rigorous, time-consuming human validation 2 sources, delaying the recognition of top-tier results. Finally, social sciences allow for inexpensive computational experimentation, but the standard for what constitutes a landmark discovery is highly subjective, and the mere volume of automated paper generation ape.socialcatalystlab.org does not satisfy the resolution criteria.