Yejin Choi, a leading NLP researcher, discusses the capabilities and limitations of large language models like GPT-4 and ChatGPT, and how they reshape our understanding of intelligence.
要点 · TL;DR
大语言模型从互联网数据获得广泛知识，但缺乏常识，需要后训练来确保安全。 LLMs gain broad knowledge from internet data but lack common sense and require post-training for safety.
推理时算法和高质量数据可帮助小模型媲美大模型。 Inference-time algorithms and high-quality data can help smaller models match larger ones.
由于价值多元，教 AI 道德很棘手，且仅靠规模可能无法解决常识问题。 Teaching morality to AI is messy due to value pluralism, and scaling alone may not solve common sense.
核心观点 · Key points
在互联网数据上的预训练赋予模型广泛知识，但基于人类反馈的强化学习（RLHF）的后训练对于生成安全、令人愉悦的回复至关重要。 Pre-training on internet data gives models broad knowledge, but post-training with RLHF is crucial for safe, pleasant responses.
当前模型缺乏常识，在儿童能解决的简单推理任务上会失败。 Current models lack common sense, failing on simple reasoning tasks that children can solve.
推理时算法（如 top-k 采样和神经逻辑解码）能显著提升较小模型的性能。 Inference-time algorithms like top-k sampling and neuro-logic decoding can significantly boost smaller model performance.
通过批评模型过滤的高质量 AI 生成数据，能帮助较小模型匹敌较大模型。 High-quality AI-generated data, filtered through critic models, can help smaller models match larger ones.
语言是推理的最佳媒介，因为它能捕捉人类知识和常识的全部范围。 Language is the best medium for reasoning, as it captures the full scope of human knowledge and common sense.
AI 安全需要教模型道德规范，但价值多元性使其成为一个混乱的计算问题。 AI safety requires teaching models moral norms, but value pluralism makes this a messy computational problem.
反共识 · Contrarian takes
仅靠 Scaling（规模扩张）可能无法解决常识问题；可能存在使用更少算力的替代方案。 Scaling alone may not solve common sense; alternative solutions using less compute likely exist.
下一个词预测的预训练目标与人类学习根本不一致。 The pre-training objective of next-token prediction is fundamentally misaligned with human learning.
拥有更好算法和数据的较小模型在特定任务上能胜过较大模型。 Smaller models with better algorithms and data can outperform larger models on specific tasks.
基于人类反馈的强化学习（RLHF）数据是专有且暴力的；应探索开放替代方案。 RLHF data is proprietary and brute-force; open alternatives should be pursued.
句法分析可能因神经网络而重新变得相关，用于语用理解。 Parsing and syntax may become relevant again for pragmatic understanding, enabled by neural networks.
在医学和法律领域部署大语言模型存在重大风险，因为过度依赖和隐藏的失败。 Deploying LLMs in medicine and law carries major risks due to over-reliance and hidden failures.
本期章节 · Chapters（共 15）
引言与背景Introduction and Background
互联网数据训练与后训练Training on Internet Data and Post-Training
预训练目标 vs 人类学习Pre-training objective vs human learning
研究动态与扩展挑战Research dynamics and the scaling challenge
推理时算法与数据质量Inference-time algorithms and data quality
常识与语言Common Sense and Language
机器能学习道德吗？Can Machines Learn Morality?
小型模型的优势Benefits of Smaller Models
NLP 中解析的衰落The Decline of Parsing in NLP
语言理解中的语用学Pragmatics in Language Understanding
研究中的推测与风险Speculation and Risk in Research
语言模型的激动人心的实际应用Exciting Real-World Applications of Language Models