论文卡-GPT-3Paper Card - GPT-3
论文卡 · 约 8 分钟Paper card · about 8 min · 引用核验 2026-11-03（仅引用链接与日期） · refs checked 2026-11-03 (reference links and dates only)
「大模型」叙事的起点论文。这张卡列清它真正论证了什么、局限作者自己怎么写的、以及哪些流行说法不在论文里。The starting point of the 'big model' narrative. This card separates what the paper actually demonstrated, what the authors themselves flagged as limits, and which popular claims are not in the paper.
研究问题：预训练+微调范式要求每个任务都有专门标注数据，而人类只靠少量例子或简单指令就能做新任务——能不能让模型也做到「不更新参数、从上下文例子学任务」（摘要原话）。
核心证据：第 3 节按任务族铺开的大表（语言建模、闭卷问答、翻译、常识推理、SuperGLUE 等）——主张是「同一模型不加微调跨多任务可用」，读时要逐表看基线与设置。
作者自陈局限（第 5 节 Limitations，当日核对）：原话 "it still has notable weaknesses in text synthesis and several NLP tasks"——生成文本与若干 NLP 任务仍有明显弱点。
论文里没有「涌现能力」（emergent abilities/emergence）字样（当日全文核对）：「GPT-3 展现了涌现能力」是后续论文的提法，不是这篇的结论。
它没有主张的：论文没有主张模型「理解」语言；「规模一定带来一切」也不是论文措辞——论文的主张严格限定在少样本任务表现。
Research question: pre-train + fine-tune needs task-specific labeled data, yet humans learn new language tasks from a few examples or instructions — can a model learn tasks from context without updating weights (abstract).
Core evidence: the Section 3 tables organized by task family (language modeling, closed-book QA, translation, commonsense, SuperGLUE…). The claim is one model, no fine-tuning, many tasks — check baselines and settings table by table.
Author-stated limitations (Section 5, verified): "it still has notable weaknesses in text synthesis and several NLP tasks".
The words 'emergent abilities' / 'emergence' do not appear in the paper (full-text check on 2026-10-29) — that framing comes from later papers.
What it does not claim: the paper does not claim the model 'understands' language, and 'scale always delivers' is not its wording either — its claims are strictly about few-shot task performance.
例子与判断Example and interpretation
以下是教学情境，不是本站运行或实测记录。A teaching scenario, not a measured or executed site test.
读少样本实验时，把零样本、单样本和少样本的输入格式分开。记录示例是否进入上下文、参数是否更新，再解释结果；提供示例和重新训练不是一回事。Separate zero-, one- and few-shot input formats. Record whether examples enter context and whether parameters change before interpreting results. Supplying examples is different from retraining.
大模型 / 大语言模型Large Language Model, LLM 预训练Pre-training 微调Fine-tuning
AI 真的读懂了你的话吗Does AI Really Understand You · 论文卡-InstructGPTPaper Card - InstructGPT · 论文卡-ChinchillaPaper Card - Chinchilla · 论文卡-Llama 2Paper Card - Llama 2