将 AI 阐述为正常技术，与乌托邦/反乌托邦的超级智能观点形成对比。 Articulates AI as normal technology, contrasting with utopian/dystopian views of superintelligence.
区分 AI 方法、应用和采用，认为变革性影响将是缓慢的（数十年尺度）。 Distinguishes AI methods, applications, and adoption, arguing transformative impacts will be slow (decades).
预测人类劳动将转向 AI 控制任务，即使有先进 AI 也不会变得多余。 Predicts human labor will shift to AI control tasks, not become superfluous, even with advanced AI.
拒绝超级智能框架，聚焦于能力和权力而非智能作为风险分析的关键。 Rejects superintelligence framing, focusing on power and capability instead of intelligence as key for risk analysis.
认为 AI 控制可通过现有工程学科（网络安全、系统安全）和技术 AI 安全来解决。 Argues AI control is tractable via existing engineering disciplines (cybersecurity, system safety) and technical AI safety.
倡导将韧性和减少不确定性作为政策目标，警告剧烈干预可能使结果恶化。 Advocates for resilience and reducing uncertainty as policy goals, warning drastic interventions may worsen outcomes.
局限 · Limitations
排除军事 AI 分析，限制了在国防领域的适用性。 Excludes military AI from analysis, limiting applicability to defense contexts.
依赖历史类比（电力、互联网），可能无法捕捉 AI 独特的速度或规模。 Relies on historical analogies (electricity, internet) that may not capture AI's unique speed or scale.
关于缓慢扩散的预测可能低估未来 AI 突破的颠覆性潜力。 Predictions about slow diffusion may underestimate disruptive potential of future AI breakthroughs.
假设市场力量和监管将充分激励安全，这在所有场景中可能不成立。 Assumes market forces and regulation will adequately incentivize safety, which may not hold in all scenarios.
未量化灾难性风险的概率，对最坏结果留下不确定性。 Does not quantify probabilities of catastrophic risks, leaving uncertainty about worst-case outcomes.
论文章节 · Sections（共 27）
概述Overview
作为潜在超级智能的 AI 愿景的替代方案An alternative to the vision of AI as a potential superintelligence
人工智能与民主自由Artificial Intelligence and Democratic Freedoms(https://knightcolumbia.org/research/artificial-intelligence-and-democratic-freedoms)
第一部分：进步的速度Part I: The Speed of Progress
AI 在安全关键领域的扩散缓慢AI diffusion in safety-critical areas is slow
扩散受限于人类、组织和制度变革的速度Diffusion is limited by the speed of human, organizational, and institutional change
外部世界给 AI 创新设定了速度限制The External world puts a speed limit on AI innovation
基准测试无法衡量实际效用Benchmarks do not measure real-world utility
经济影响可能是渐进的Economic impacts are likely to be gradual
AI 方法进展的速度限制Speed limits to progress in AI methods
第二部分：拥有先进 AI 的世界可能是什么样子Part II: What a World With Advanced AI Might Look Like
人类能力不受生物学限制Human ability is not constrained by biology
游戏对超级智能可能性的误导性直觉Games provide misleading intuitions about the possibility of superintelligence
控制有多种形式Control comes in many flavors
第三部分：风险Part III: Risks
军备竞赛是一个老问题Arms races are an old problem
针对滥用的主要防御必须位于模型下游The primary defenses against misuse must be located downstream of models
AI 对防御有用AI is useful for defense
灾难性失调是一种推测性风险Catastrophic misalignment is a speculative risk
历史表明，普通 AI 可能引入多种系统性风险History suggests normal AI may introduce many kinds of systemic risks
第四部分：政策Part IV: Policy
不确定性下的政策制定挑战The challenge of policy making under uncertainty
将减少不确定性作为政策目标Reducing uncertainty as a policy goal
韧性论证The case for resilience
防扩散难以执行且导致单点故障Nonproliferation is infeasible to enforce and leads to single points of failure