大规模预训练近期已成为创建通用生成模型（如 GPT-3、Megatron-Turing NLG、Gopher 等）的技术。本文强调了这类模型的一个反直觉特性及其政策影响：它们在广泛的训练分布上具有可预测的损失（体现在“缩放定律”中），但具体能力、输入和输出却不可预测。我们认为，高层次的可预测性和有用能力的表象推动了模型的快速发展，而不可预测性则使模型部署的后果难以预料。我们通过文献和实际观察中的例子展示了这种组合如何导致社会危害，并进行了两项新实验以说明不可预测性带来的危害。此外，我们分析了这些矛盾特性如何给模型开发者带来部署动机和挑战。最后，我们提出了一系列可能的干预措施，以增加这些模型产生有益影响的机会。本文旨在帮助理解和管理 AI 系统的政策制定者、关注工作潜在政策影响的技术人员，以及希望分析、批评和开发大型生成模型的学者。
Large-scale pre-training has recently emerged as a technique for creating capable, general purpose, generative models such as GPT-3, Megatron-Turing NLG, Gopher, and many others. In this paper, we highlight a counterintuitive property of such models and discuss the policy implications of this property. Namely, these generative models have an unusual combination of predictable loss on a broad training distribution (as embodied in their "scaling laws"), and unpredictable specific capabilities, inputs, and outputs. We believe that the high-level predictability and appearance of useful capabilities drives rapid development of such models, while the unpredictable qualities make it difficult to anticipate the consequences of model deployment. We go through examples of how this combination can lead to socially harmful behavior with examples from the literature and real world observations, and we also perform two novel experiments to illustrate our point about harms from unpredictability. Furthermore, we analyze how these conflicting properties combine to give model developers various motivations for deploying these models, and challenges that can hinder deployment. We conclude with a list of possible interventions the AI community may take to increase the chance of these models having a beneficial impact. We intend this paper to be useful to policymakers who want to understand and regulate AI systems, technologists who care about the potential policy impact of their work, and academics who want to analyze, critique, and potentially develop large generative models.
核心贡献 · Key contributions
识别出大型生成模型中可预测的缩放定律与不可预测的具体能力之间的矛盾组合。 Identifies paradoxical combination of predictable scaling laws and unpredictable specific capabilities in large generative models.
通过新颖实验证明有害能力（如带偏见的再犯预测）在规模扩大时突然涌现。 Demonstrates abrupt emergence of harmful capabilities like recidivism prediction with bias at scale via novel experiments.
分析可预测性如何推动快速发展，而不可预测性如何阻碍后果预判。 Analyzes how predictability drives rapid development while unpredictability hinders consequence anticipation.
记录开放输入和输出导致未知能力及有害输出（如有毒性随规模增长）。 Documents open-ended inputs and outputs leading to unknown capabilities and harmful outputs like toxicity scaling.
提出政策干预措施，包括国家研究云、红队测试和治理结构。 Provides policy interventions including national research clouds, red teaming, and governance structures.
实证显示大型模型开发中产业界与学术界差距扩大，以及部署的激增。 Empirically shows increasing industry-academia gap in large model development and proliferation of deployments.
局限 · Limitations
聚焦大型语言模型；结论可能不推广至其他生成模态。 Focuses on large language models; findings may not generalize to other generative modalities.
新颖实验限于再犯预测和毒性；其他危害可能不同。 Novel experiments limited to recidivism prediction and toxicity; other harms may differ.
未涉及大型模型的劳动力、供应链或环境成本。 Does not address labor, supply chain, or environmental costs of large models.
政策干预措施是推测性的，缺乏有效性的实证验证。 Policy interventions are speculative and lack empirical validation of effectiveness.
动机和障碍分析是定性的，可能未涵盖所有因素。 Analysis of motivations and barriers is qualitative and may not capture all factors.
论文章节 · Sections（共 13）
摘要Abstract
1. 引言1. Introduction
2. 大型生成模型的显著特征2. Distinguishing Features of Large Generative Models
2.1. 平滑的通用能力扩展2.1. Smooth General Capability Scaling
2.2. 突发的特定能力扩展2.2. Abrupt Specific Capability Scaling
2.3. 开放式的输入与领域2.3. Open-Ended Inputs and Domains
2.4. 开放式的输出2.4. Open-Ended Outputs
3. 大型模型开发与部署的动机与问题3. Motivations and Problems in the Development and Deployment of Large Models
3.1. 开发与部署大型模型的动机3.1. Motivations for Developing and Deploying Large Models
3.2. 开发与部署大型模型的准入门槛3.2. Barriers to Entry in Developing and Deploying Large Models
3.3. 实证观察3.3. Empirical Observations
4. 促进有益部署的干预措施4. Interventions to Encourage Beneficial Deployments