Will artificial superintelligence (ASI) be created before January 1, 2050?

Absent a strict global ban, the baseline trajectory points strongly toward the creation of ASI before 2050, yielding a 79% probability. Frontier AI capabilities are advancing rapidly, driven by substantial capital and massive compute scaling. Epoch reports frontier language-model training compute has grown 5x annually since 2020, alongside a 3x annual improvement in pre-training compute efficiency epoch.ai. This has led to sharp performance leaps, such as SWE-bench Verified scores jumping from 60% to near 100% in a single year hai.stanford.edu. Timelines for AGI and the subsequent leap to ASI have compressed significantly. Major lab executives project extremely powerful AI or AGI within the next 3 to 10 years (e.g., Amodei pointing to 2026-2027, Hassabis estimating a 50% chance within the decade as of early 2026 [fortune]). Aggregated expert expectations roughly align, with the 2023 AI Impacts survey placing a 50% probability on high-level machine intelligence by roughly 2047 arxiv.org, and quantitative platforms estimating AGI by the early 2030s metaculus.com with the AGI-to-ASI transition taking roughly 26 to 44 months metaculus.com[057e1]. DeepMind's recent 'From AGI to ASI' report also outlines multiple plausible transition pathways, including recursive self-improvement and multi-agent collectives deepmind.google. The probability is kept from being higher by physical bottlenecks (data, energy, and semiconductor constraints), the possibility of a paradigm plateau, and the strict resolution requirement for a widely-accepted determination that the system substantially surpasses human capabilities across virtually all domains. If a US-China treaty banning the creation of ASI is enacted and adaptively enforced globally, the probability falls sharply to 21%. The treaty, modeled on the MIRI Technical Governance Team proposal arxiv.orgarxiv.org[techgov 7ecba9], leverages the US and China's dominance over advanced semiconductor supply chains cset.georgetown.educongress.gov to impose strict hardware and compute limits arxiv.org. Crucially, the condition specifies this enforcement is adaptive, capable of closing loopholes like sub-threshold decentralized training through VRAM, network traffic, and procurement monitoring arxiv.org. This robust enforcement infrastructure effectively neutralizes the primary pathways to ASI: massive compute scaling by frontier labs and unmonitored sovereign projects. However, the probability remains non-trivial at 21% due to three primary leakage paths over the 24-year horizon. First, the MIRI-style proposal is explicitly a moratorium of unlimited duration rather than a permanent ban. It is designed to be lifted once alignment is solved, leaving a distinct possibility that a globally monitored, legally authorized safe ASI is developed before 2050. Second, maintaining flawless, active interdiction and cooperation between the US and China over 24 years is historically improbable; competitive pressures or regime defection could erode enforcement. Third, despite adaptive monitoring, radical algorithmic efficiency breakthroughs could sufficiently lower compute barriers, allowing rogue actors to develop ASI using stockpiled or indigenous hardware before enforcement mechanisms can adapt. In sum, an adaptive global ban suppresses the massive industrial compute pathway that dominates the baseline, but the likelihood of geopolitical erosion, low-compute breakthroughs, and eventual legal authorization of a safe ASI drives the remaining 21% probability.