imported>暁月凛奈：回退~2025-53292-2（对话）的编辑，改回Willy1018-bot的最后一个版本

2025-09-06T02:53:08Z

回退~2025-53292-2（对话）的编辑，改回Willy1018-bot的最后一个版本

新页面

{{Expand language|1=en|time=2023-04-04T14:39:42+00:00}}
{{noteTA|G1=IT}}
{{Infobox software
| title = 生成型预训练变换模型 4<br />Generative Pre-trained Transformer 4（GPT-4）
| logo=
| logo_size = 200px
| developer = [[OpenAI]]
| released = {{Start date and age|2023|03|14}}
| language count =
| genre ={{ indented plainlist |
*[[大型语言模型]]
*[[基于转换器的生成式预训练模型]]
*[[基礎模型]]
}}
| license = [[专有软件|专有]]
| repo =
| website = {{URL|https://openai.com/gpt-4}}
| replaces = [[GPT-3]]
| replaced_by = [[GPT-4o]]
}}

'''生成型预训练变换模型4'''（{{langx|en|Generative Pre-trained Transformer 4}}，简称'''GPT-4'''）是由[[OpenAI]]公司开发，並於2023年3月14日发布的[[自回归模型|自回归]][[语言模型]]<ref name=":0">{{Cite web |date=2023-03-14 |title=GPT-4 |url=https://openai.com/research/gpt-4 |website=openai.com |access-date=2023-03-14 |archive-date=2023-03-14 |archive-url=https://web.archive.org/web/20230314174531/https://openai.com/research/gpt-4 |dead-url=no }}</ref><ref name="vox-1" />。2025年4月10日，OpenAI在更新日志中宣布，ChatGPT将于2025年4月30日停止使用GPT‑4，但API依旧可以调用<ref name="e050">{{cite web | title=OpenAI官宣GPT4退役，由4o接棒 | website=openai_网易订阅 | date=2025-04-12 | url=https://www.163.com/dy/article/JSUV6QTD05568W0A.html | language=zh | access-date=2025-04-12}}</ref>。

==能力==
OpenAI 在宣布 GPT-4 时表示，它“比 GPT-3.5 更可靠、更有创意，并且能够处理更细微的指令。” <ref>{{Cite web |last=Wiggers |first=Kyle |date=2023-03-14 |title=OpenAI releases GPT-4, a multimodal AI that it claims is state-of-the-art |url=https://techcrunch.com/2023/03/14/openai-releases-gpt-4-ai-that-it-claims-is-state-of-the-art/ |url-status=live |archive-url=https://web.archive.org/web/20230315003723/https://techcrunch.com/2023/03/14/openai-releases-gpt-4-ai-that-it-claims-is-state-of-the-art/ |archive-date=2023-03-15 |access-date=2023-03-15 |website=[[TechCrunch]]}}</ref>

他们制作了两个版本的 GPT-4，上下文窗口分别为 8,192 和 32,768 个詞元，比分别限制为 4,096 和 2,049 个詞元的 GPT-3.5 和 GPT-3 有了显着改进。<ref>{{Cite web |author=OpenAI |title=Models |url=https://platform.openai.com/docs/models |access-date=2023-03-18 |website=OpenAI API |language=en |archive-date=2023-03-17 |archive-url=https://web.archive.org/web/20230317000210/https://platform.openai.com/docs/models |url-status=live }}</ref>与其前身不同，GPT-4 可以将图像和文本作为输入；<ref name="guardian creative">{{cite web |author1=Alex Hern |author2=Johana Bhuiyan |date=2023-03-14 |title=OpenAI says new model GPT-4 is more creative and less likely to invent facts |url=https://www.theguardian.com/technology/2023/mar/14/chat-gpt-4-new-model |url-status=live |archive-url=https://web.archive.org/web/20230315003816/https://www.theguardian.com/technology/2023/mar/14/chat-gpt-4-new-model |archive-date=2023-03-15 |access-date=2023-03-15 |website=[[The Guardian]]}}</ref>这使它能够描述不寻常图像中的幽默、总结截屏文本以及回答包含图表的试题。<ref name="openai_research">{{Cite web |author=OpenAI |date=2023-03-14 |title=GPT-4 |url=https://openai.com/research/gpt-4 |access-date=2023-03-20 |website=OpenAI Research |language=en-US |archive-date=2023-03-14 |archive-url=https://web.archive.org/web/20230314174531/https://openai.com/research/gpt-4 |url-status=live }}</ref>

尽管有这些新能力，GPT-4 和它的前辈一样，仍然倾向于产生[[幻觉 (人工智能)|幻觉]]答案。<ref name="nyt-4">{{Cite news |last1=Metz |first1=Cade |last2=Collins |first2=Keith |date=2023-03-14 |title=10 Ways GPT-4 Is Impressive but Still Flawed |language=en-US |work=[[The New York Times]] |url=https://www.nytimes.com/2023/03/14/technology/openai-new-gpt4.html |url-status=live |url-access=limited |access-date=2023-03-14 |archive-url=https://archive.today/20230314170804/https://www.nytimes.com/2023/03/14/technology/openai-new-gpt4.html |archive-date=2023-03-14}}</ref>

==使用==
{{main|ChatGPT}}
GPT-4可通过API和[[ChatGPT]] Plus用户使用。

2023年3月9日，[[微软]]表示GPT-4是[[多模态学习]]的（将支持文本以外的内容）<ref>{{Cite web |title=GPT-4 is coming next week – and it will be multimodal, says Microsoft Germany |url=https://www.heise.de/news/GPT-4-is-coming-next-week-and-it-will-be-multimodal-says-Microsoft-Germany-7540972.html |website=[[Heise.de]] |language=de |last= |first= |archive-url=https://web.archive.org/web/20230601223659/https://www.heise.de/news/GPT-4-is-coming-next-week-and-it-will-be-multimodal-says-Microsoft-Germany-7540972.html |archive-date=2023-06-01 |access-date=2023-03-10 |dead-url=no}}</ref>，並將GPT-4置入該公司旗下的[[Microsoft Office]]。<ref name="Potuck 2023">{{cite web |title=Microsoft Office will soon gain GPT-4 AI with new 'Copilot' feature |url=https://9to5mac.com/2023/03/16/microsoft-office-to-get-gpt-4-ai-soon/ |website=9to5Mac |date=2023-03-16 |last=Potuck |first=Michael |archive-url=https://web.archive.org/web/20230502090425/https://9to5mac.com/2023/03/16/microsoft-office-to-get-gpt-4-ai-soon/ |archive-date=2023-05-02 |access-date=2023-03-17 |dead-url=no}}</ref>

== 迭代 ==
=== GPT-4 Turbo ===
GPT-4 Turbo 是 OpenAI 發布的一種優化版本的 GPT-4 模型。

與 GPT-4 相比，GPT-4 Turbo 擁有更高的運行效率和成本效益，適合大規模的應用需求。

此模型專注於改善速度和可擴展性，同時在處理能力和生成文本的品質上保持高水準。OpenAI 通常在產品如 ChatGPT 中部署 GPT-4 Turbo，以提供更快的回應速度和較低的計算資源需求。

GPT-4 Turbo 的具體改進細節尚未完全公開，但 OpenAI 表示該模型的優化主要集中於架構設計與資源管理，使其能更好地滿足商業需求。

== 評價 ==
[[Vox]]称GPT-4从各方面来说都优于OpenAI之前发布的GPT-3和GPT-3.5。<ref name="vox-2" />

[[The Verge]]还在报道中引用了关于将大幅增加GPT-3的参数数量（从1750亿到100万亿）的传言，但OpenAI首席执行官[[山姆·柯曼]]将其斥责为“完全是胡说八道”。<ref name="verge" />

美国众议员[[堂·拜尔]]和[[刘云平]]向《[[纽约时报]]》证实，奥尔特曼于2023年1月访问国会展示GPT-4及其与其他人工智能模型相比所改进的“安全控制”。<ref name="nyt-3" />

== 参见 ==
*[[GPT-3]]
*[[GPT-4o]]
*[[OpenAI]]
*[[語言模型]]
*[[大型语言模型]]

== 参考来源 ==
{{reflist|refs=
<ref name="vox-1">{{Cite web |title=23 things we think will happen in 2023 |url=https://www.vox.com/future-perfect/2023/1/2/23494204/biden-trump-putin-supreme-court-2023-predictions |website=Vox |date=2023-01-02 |last=Matthews |first=Dylan |url-status=live |archive-url=https://web.archive.org/web/20230127075727/https://www.vox.com/future-perfect/2023/1/2/23494204/biden-trump-putin-supreme-court-2023-predictions |archive-date=2023-01-27 |access-date=2023-01-27}}</ref>
<ref name="vox-2">{{Cite web |title=Think AI was impressive last year? Wait until you see what's coming. |url=https://www.vox.com/future-perfect/2023/1/4/23538647/artificial-intelligence-chatgpt-openai-google-meta-facial-recognition |website=Vox |date=2023-01-04 |last=Piper |first=Kelsey |url-status=live |archive-url=https://web.archive.org/web/20230123200851/https://www.vox.com/future-perfect/2023/1/4/23538647/artificial-intelligence-chatgpt-openai-google-meta-facial-recognition |archive-date=2023-01-23 |access-date=2023-01-27}}</ref>
<ref name="nyt-3">{{Cite news|url=https://www.nytimes.com/2023/03/03/technology/artificial-intelligence-regulation-congress.html|title=As A.I. Booms, Lawmakers Struggle to Understand the Technology|first=Cecilia|last=Kang|work=The New York Times|date=2023-03-03|via=NYTimes.com|access-date=2023-03-03|archive-date=2023-03-03|archive-url=https://web.archive.org/web/20230303100624/https://www.nytimes.com/2023/03/03/technology/artificial-intelligence-regulation-congress.html|url-status=live}}</ref>
<ref name="verge">{{Cite web |title=OpenAI CEO Sam Altman on GPT-4: "people are begging to be disappointed and they will be" |url=https://www.theverge.com/23560328/openai-gpt-4-rumor-release-date-sam-altman-interview |website=The Verge |date=2023-01-18 |last=Vincent |first=James |url-status=live |archive-url=https://web.archive.org/web/20230126232257/https://www.theverge.com/23560328/openai-gpt-4-rumor-release-date-sam-altman-interview |archive-date=2023-01-26 |access-date=2023-01-27}}</ref>
}}

{{OpenAI}}
{{Differentiable computing}}
{{自然语言处理}}
{{生成式人工智能}}

[[Category:OpenAI]]
[[Category:聊天机器人]]
[[Category:2023年面世的产品]]
[[Category:交互式叙事]]
[[Category:大型语言模型]]
[[Category:基于转换器的生成式预训练模型]]

GPT-4 - 版本历史

imported>暁月凛奈：​回退~2025-53292-2（对话）的编辑，改回Willy1018-bot的最后一个版本

imported>暁月凛奈：回退~2025-53292-2（对话）的编辑，改回Willy1018-bot的最后一个版本