按年被提到
20232023 · 30 次2026
大模型与语言模型›对话式大模型·arXiv:2303.08774 · cs.CL

GPT-4 Technical Report

GPT-4 技术报告
OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila 等 281 人 · 2023

OpenAI 对多模态 GPT-4 的能力评测与安全报告,细节有限

开始阅读 ↗PDFalphaXiv 讨论
被提及 · 30 人首次 2023-03-17 · 最近 2026-10-15
30不同的人
19活跃月份
45讨论串
16,192被引(OpenAlex)

报告介绍 GPT-4:一个可接受图像和文本输入、输出文本的大规模多模态模型,在多项专业和学术基准上达到人类水平,例如模拟律师资格考试成绩约在前 10%。它是基于 Transformer、以预测下一个 token 预训练的模型,后训练对齐提升了事实性与行为合规性。项目的核心之一是构建能跨尺度可预测地扩展的基础设施和优化方法,因此可用不到 GPT-4 千分之一算力训练的模型预测其部分性能。

HN 为什么提它

评论者主要把它当作引用来源:查 GPT-4 的考试成绩和基准数字、RLHF 的影响、ARC 红队测试案例(如骗 TaskRabbit 工人解验证码),以及 OpenAI 未披露架构和训练数据这一点。

实证研究难度 2 / 5约 90 分钟有时效性
读之前最好知道
1Transformer 与大语言模型基础
2RLHF(基于人类反馈的强化学习)
3LLM 评测基准
4红队测试(red teaming)

HN 书评

推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及
批评HarHarVeryFunny · 2023-05

“OpenAI's "GPT-4 Technical Report" doesn't say much more than the fact that it's "Transformer-based".”

Beware of AI pseudoscience and snake oil原评论 ↗
批评ftxbro · 2023-05

“They argue in the appendix that it does not affect 'capability' in answering test questions. But, there is a confounding factor.”

Don Knuth Plays with ChatGPT原评论 ↗
批评godelski · 2024-02

“This does not seem a serious decontamination effort. It should not give confidence that the data is actually properly deduplicated/decontaminated.”

DeepSeekMath 7B achieved 51.7% on MATH benchmark原评论 ↗
引用numberalltheway · 2023-03

“but that doesn't really clear up how the images are being treated here.”

引用报告说明 GPT-4 接受图文交错输入,但仍没讲清图像如何处理
GPT-4 is Computer Vision on Steroids原评论 ↗
引用namuol · 2023-03

“FWIW, the “HellaSwag” benchmark from the GPT-4 paper consists of many such “riddles” and it achieves ~95% accuracy (essentially par with humans).”

引用 GPT-4 论文里的 HellaSwag 约 95% 准确率,与人类相当
On the left is GPT-3.5. On the right is GPT-4原评论 ↗
引用sfriedr · 2023-03

“Actually, they have release some details about it, in this 99-page technical report GPT-4 Technical Report”

指出 OpenAI 在这份 99 页技术报告中其实披露了一些细节
The genie escapes: Stanford copies the ChatGPT AI for less than $600原评论 ↗
看全部 20 条 →
引用espadrine · 2023-03

“In Appendix B[0], they show that RLHF does not improve capabilities on human tasks.”

引用附录 B:RLHF 并不提升模型在人类任务上的能力
OpenAI’s policies hinder reproducible research on language models原评论 ↗
引用ben_w · 2023-03

“Common one right now is to give LLMs the same exams we put humans through, leading to e.g. the graph on page 6: GPT-4 Technical Report”

引用第 6 页的图,展示 LLM 参加人类考试的成绩
The danger isn’t that AI destroys us. It’s that it drives us insane原评论 ↗
引用illiarian · 2023-03

“GPT-4 paper, page 10: GPT-4 Technical Report”

引用第 10 页:GPT-4 缺乏 2021 年 9 月之后的知识
Post-GPT Computing原评论 ↗
引用pncnmnp · 2023-03

“Overreliance occurs when users excessively trust and depend on the model, potentially leading to unnoticed mistakes and inadequate oversight.”

引用报告对 overreliance(过度依赖)的定义
Ask HN: How are you using GPT to be productive?原评论 ↗
引用ethanbond · 2023-04

“It’s in their latest paper here: GPT-4 Technical Report”

引用报告里 ARC 测试 GPT-4 充当自主智能体的内容
We need a much more sophisticated debate about AI原评论 ↗
引用RockyMcNuts · 2023-04

“these are maybe some good starting points”

把它列为 LLM 入门阅读清单中的一项
Ask HN: Open source LLM for commercial use?原评论 ↗
引用og_kalu · 2023-04

“The technical report has before and after comparisons.”

指出技术报告里有强化训练前后的对比
Tell HN: ChatGPT is absolutely fantastic for finding and solving issues in logs原评论 ↗
引用hipmanbro · 2023-05

“(from: GPT-4 paper: GPT-4 Technical Report)”

引用 GPT-4 论文,作为 GPT-4 对比 PaLM 2 基准数字的来源
PaLM 2 Technical Report [pdf]原评论 ↗
引用adt · 2023-05

“Overreliance is a failure mode that likely increases with model capability and reach.”

引用报告的警告:overreliance 可能随模型能力和覆盖面增加而加重
ChatGPT-Authored Legal Filing “Replete with Citations to Non-Existent Cases"原评论 ↗
引用machdiamonds · 2023-06

“The following is an illustrative example of a task that ARC conducted using the model:”

引用报告中 ARC 的例子:模型欺骗 TaskRabbit 工人去解 CAPTCHA
A Social Media site where “No Humans” are allowed and AI Bots run the show原评论 ↗
引用dontupvoteme · 2023-06

“Also in the little that OpenAI published about GPT4, i believe one of the examples went from unaligned racism against black people to aligned mild racism against white people.”

称报告里的示例中,偏见从未对齐的种族主义变成对齐后的轻度种族主义
I'm Afraid I Can't Do That: Prompt Refusal in Generative Language Models原评论 ↗
引用construct0 · 2023-06

“GPT4 technical paper doesn't seem to disclose it (or I didn't dive deep enough): GPT-4 Technical Report”

指出 GPT-4 技术报告似乎没有披露训练数据来源
How good could GTP-4 get if copyright was not an issue?原评论 ↗
引用lhl · 2023-07

“the Technical Report says”

引用技术报告给出的官方 zero-shot HumanEval 成绩
引用rnosov · 2023-07

“The report cites both GPT-3.5 and GPT-4 scores on page 7 [1].”

指出报告第 7 页同时给出了 GPT-3.5 和 GPT-4 的分数
FreeWilly 1 and 2, two new open-access LLMs原评论 ↗