Will there be an AI system by the end of 2030 that is smarter than the top human experts in 90%+ of fields?

probability: 36%

rationale:

This probability rests on balancing the extraordinary acceleration of frontier AI against an exceptionally strict resolution threshold. The question requires an artificial superintelligence—comparable to a "country of geniuses"—that not only matches but demonstrably outperforms the best human experts in at least 90% of all intellectual and professional fields. Crucially, this must be "widely judged" to be true by the end of 2030. Beating average professionals or achieving superhuman performance solely in STEM is not enough; the system must conquer tacit, high-context domains, and the broader community must formally recognize it.

As of late 2026, the technical trajectory is advancing at a blistering pace. Hard academic benchmarks that once seemed formidable are rapidly saturating: FrontierMath Tier 4 (v2) has reached 100% epoch.ai, GPQA Diamond sits at roughly 96.5% epoch.ai, and Humanity's Last Exam (HLE) has hit 74.3% 2 sources. Systems are also producing research-grade results, such as formalizing Fermat's Last Theorem and solving open physics problems 2 sources. Critically, AI is increasingly assisting in its own development. Evidence suggests that AI-driven R&D progress has accelerated to roughly 1.9x its previous rate metr.org, creating a compounding feedback loop that makes reaching extreme capability levels by 2030 highly plausible.

However, academic excellence does not seamlessly translate into broad professional dominance. Real-world, open-ended work remains a significant bottleneck. While the Remote Labor Index (RLI) recently saw a rapid jump to a 41.3% automation rate labs.scale.com, this is still well below human baselines, let alone top-expert standards across diverse professions 2 sources. Evaluators explicitly warn that success on constrained software or math tasks should not be extrapolated to holistic, high-context jobs metr.org. Furthermore, the introduction of even harder evaluations, like the FrontierMath Erdős open problems where top systems score just 4.4% epoch.ai, demonstrates that elite autonomous research still has substantial headroom.

Beyond technical hurdles, the "widely judged" requirement introduces severe friction. Even if the raw capability to dominate 90% of fields exists by 2029 or 2030, building credible consensus will be slow. In qualitative fields like law, management, humanities, and the arts, there are no clean benchmarks for top-expert performance. Disputes over AI credit and evaluation methodologies will inevitably drag out recognition quantamagazine.org. Additionally, progress faces potential headwinds from deliberate lab pacing commitments darioamodei.com and government interventions, which have already disrupted frontier releases thenextweb.com. The distinction between internal achievement and widespread recognition is essential: widely recognized ASI requires the capability to have been achieved, plus sufficient evidence and consensus before the deadline. Its probability must therefore sit below the likelihood of achieving the capability internally, consistent with the related forecasts .

Ultimately, there are four more years for compounding, accelerated AI progress to unfold. Expert and community timelines for general AI and superintelligence frequently cluster around the 2028–2030 window 2 sources, suggesting the underlying capability is entirely within reach. Yet, the strict demands of near-universal top-expert coverage combined with the lag of forming broad consensus in hard-to-measure fields make this an exceptionally high bar. Weighing the explosive technical momentum against these structural drags, and distinguishing widely recognized ASI from capability achieved internally, results in an estimate of 36%.