作者以自身 2010 年参与美国情报体系预测竞赛为例，强调提升准确度的要素：开放吸收多方观点、数据与数理素养、谦逊并从错误学习；且多位预测者的聚合通常优于单一高手。但 AI 因历史数据稀少与非线性风险更难：例如「千禧年奖」7 题的解出速度约为每 25 年 1 题，难以直接外推到 AI 解题能力。
为可操作性，作者提出以短中期指标替代「何时达到 AGI」：LEAP 在数月内收集逾 300 名专家资料，平均预期 2030 年 AI 将用掉美国用电 7%（2024 年估计 1%，约增至 7 倍），同年 18% 工时受 AI 协助（2024 年 9 月为 2%，约增至 9 倍）；并给出 2040 年 AI 至少达到「千年级技术」的重要性之机率为 32%。计划未来 3 年每月更新预测并以奖金鼓励准确与可辩护的解释，供电网、就业与政策提早因应极端情境。
The essay argues AI debate rewards confident slogans over testable detail: some claim AI will exceed human intelligence by 2030, others say its impact will be 10× bigger than the Industrial Revolution and 10× faster, while skeptics warn of a multitrillion-dollar investment bubble. The author urges replacing proclamations with measurable, policy-relevant probabilistic forecasts that can be checked against reality and used to judge track records.
Drawing on a 2010 US intelligence forecasting tournament, the author highlights traits linked to better prediction: open-minded synthesis, numeracy and data hunger, humility, and learning from error; aggregating many forecasters typically beats even top analysts. AI is harder because history is thin and effects may be nonlinear: even benchmarking against the Millennium Prize Problems (7 total) is shaky when solutions have averaged about 1 every 25 years.
Instead of “when AGI,” the proposal is to forecast short- and medium-term indicators. Early LEAP results from 300+ experts estimate AI uses 7% of US electricity in 2030 versus ~1% in 2024 (about 7×), and AI assists 18% of work hours in 2030 versus 2% in Sep 2024 (about 9×). Experts assign a 32% chance that by 2040 AI ranks at least a “technology of the millennium.” Forecasts update monthly for 3 years, with prizes to reward accuracy and well-grounded explanations so policymakers can prepare for downside tails.
