imported>TangZhi：TangZhi移動頁面Auto-GPT至AutoGPT：已更名

2026-03-19T01:20:44Z

TangZhi移動頁面Auto-GPT至AutoGPT：已更名

新页面

{{noteTA
| G1=IT
| G2=PP
}}
{{Infobox software
| title = AutoGPT
| logo = AutoGPT Logo (2024-present).svg{{!}}class=skin-invert
| developer = 托兰·布鲁斯·理查兹
| released = 2023年3月30日
| genre = [[自主权|自主]][[人工智能]]软件代理程序
| repo = {{URL|https://github.com/Significant-Gravitas/AutoGPT}}
| programming language = [[Python]]
| license = [[MIT許可證]]
| website = https://agpt.co
}}
'''AutoGPT'''是人工智能软件代理程序，使用者可以通过[[自然语言]]为渠道告知其计划，然后该程序则可将这一计划拆分为若干个子任务，并能够自动使用互联网或其他工具来实现这些任务。<ref>{{Cite news |first=Sabrina |last=Ortiz |date=2023-04-14 |title=What is Auto-GPT? Everything to know about the next powerful AI tool |url=https://www.zdnet.com/article/what-is-auto-gpt-everything-to-know-about-the-next-powerful-ai-tool/ |access-date=2023-04-16 |publisher=ZDNET |language=en |archive-date=2023-04-16 |archive-url=https://web.archive.org/web/20230416195550/https://www.zdnet.com/article/what-is-auto-gpt-everything-to-know-about-the-next-powerful-ai-tool/ |url-status=live }}</ref>该程序使用了由[[OpenAI]]公司提供的[[GPT-4]]或[[GPT-3.5]][[应用程序接口|API]]模型，是首批将GPT-4应用于自动执行任务的程序之一。<ref name=tsun />

== 详细信息 ==
与[[ChatGPT]]等交互式系统不同的是，由于AutoGPT无需人工输入每项具体的任务，它能够自动为自己安排新的目标，从而实现更宏大的计划。它能够执行对[[提示工程|提示]]的响应以完成目标任务，并且在此过程中可以通过递归的方式修改自己的提示，以实现对新信息的响应。<ref>{{Cite news |last=Palmer |first=Sarah |date=2023-04-14 |title=What is AutoGPT? Here's what we know about the hottest new AI chatbot |url=https://www.euronews.com/next/2023/04/14/make-some-room-chatgpt-rival-ai-chatbot-autogpt-has-landed-this-is-what-we-know-about-it-s |access-date=2023-04-16 |publisher=Euronews |language=en |archive-date=2023-04-16 |archive-url=https://web.archive.org/web/20230416193947/https://www.euronews.com/next/2023/04/14/make-some-room-chatgpt-rival-ai-chatbot-autogpt-has-landed-this-is-what-we-know-about-it-s |url-status=live }}</ref>该程序可以通过写入和读取数据库和文件来管理短期或长期记忆，并可以通过汇总管理文本窗口长度的要求，还可以在无人值守的情况下执行基于互联网的操作（例如[[网络搜索]]、[[網頁表單]]和API交互），此外还能将文本输出为语音。<ref name="tsun" />

很多观察家赞扬了AutoGPT编写、调试、测试和编辑代码的能力，甚至还认为这种能力可能会扩展到AutoGPT自己的源代码，从而实现自我改进。<ref name="tsun" />但实际上AutoGPT的底层GPT模型是专有模型<ref>{{Cite web |title=OpenAI is giving Microsoft exclusive access to its GPT-3 language model |url=https://www.technologyreview.com/2020/09/23/1008729/openai-is-giving-microsoft-exclusive-access-to-its-gpt-3-language-model/ |access-date=2023-04-18 |website=MIT Technology Review |language=en |archive-date=2021-02-05 |archive-url=https://web.archive.org/web/20210205121656/https://www.technologyreview.com/2020/09/23/1008729/openai-is-giving-microsoft-exclusive-access-to-its-gpt-3-language-model/ |url-status=live }}</ref><ref>{{Cite web |title=GPT-4 is bigger and better than ChatGPT—but OpenAI won't say why |url=https://www.technologyreview.com/2023/03/14/1069823/gpt-4-is-bigger-and-better-chatgpt-openai/ |access-date=2023-04-18 |website=MIT Technology Review |language=en |archive-date=2023-03-17 |archive-url=https://web.archive.org/web/20230317224201/https://www.technologyreview.com/2023/03/14/1069823/gpt-4-is-bigger-and-better-chatgpt-openai/ |url-status=live }}</ref>，AutoGPT无法对其修改，并且通常无法访问自身的基础系统代码。

== 背景 ==
2023年3月14日，OpenAI发布了[[大型语言模型]]GPT-4，因其在很多任务中都具备高超的性能，各路观察家对此留下了深刻的印象。<ref>{{Cite web |last=Belfield |first=Haydn |date=2023-03-25 |title=If your AI model is going to sell, it has to be safe |url=https://www.vox.com/future-perfect/2023/3/25/23655082/ai-openai-gpt-4-safety-microsoft-facebook-meta |access-date=2023-04-17 |website=Vox |language=en |archive-date=2023-03-28 |archive-url=https://web.archive.org/web/20230328192017/https://www.vox.com/future-perfect/2023/3/25/23655082/ai-openai-gpt-4-safety-microsoft-facebook-meta |url-status=live }}</ref>作为一个语言模型，GPT-4本身无法自主执行操作。但在预发布安全测试期间，[[红队]]研究人员发现GPT-4模型可以执行一些现实世界中的操作，例如说服TaskRabbit的工作人员为其解决[[验证码]]难题。<ref>{{Cite news |first=Joseph |last=Cox |date=2023-03-15 |title=GPT-4 Hired Unwitting TaskRabbit Worker By Pretending to Be 'Vision-Impaired' Human |url=https://www.vice.com/en/article/jg5ew4/gpt4-hired-unwitting-taskrabbit-worker |access-date=2023-04-16 |publisher=Vice News Motherboard |language=en |archive-date=2023-04-10 |archive-url=https://web.archive.org/web/20230410053911/https://www.vice.com/en/article/jg5ew4/gpt4-hired-unwitting-taskrabbit-worker |url-status=live }}</ref>[[微软]]团队的研究人员认为，鉴于GPT-4的能力水平接近于人类，因此可以将其视为[[通用人工智能]]的早期（但不完善）版本。不过研究人员同时强调，这一系统经实验证实存在严重的局限性。<ref>{{Cite web |title=Microsoft Now Claims GPT-4 Shows 'Sparks' of General Intelligence |url=https://www.vice.com/en/article/g5ypex/microsoft-now-claims-gpt-4-shows-sparks-of-general-intelligence |access-date=2023-04-18 |website=www.vice.com |language=en |archive-date=2023-04-18 |archive-url=https://web.archive.org/web/20230418151004/https://www.vice.com/en/article/g5ypex/microsoft-now-claims-gpt-4-shows-sparks-of-general-intelligence |url-status=live }}</ref>

AutoGPT的发布者为游戏公司Significant Gravitas有限公司创始人托兰·布鲁斯·理查兹，发布日期为2023年4月20日。该程序在发布后不久便登上了[[GitHub]]热榜，而后又登上了[[推特]]热榜。<ref name="tsun">{{Cite news |first=Kevin |last=Jiang |date=2023-04-14 |title=What's Auto-GPT? New, autonomous 'AI agents' can act on their own, rewrite their own code |url=https://www.thestar.com/business/technology/2023/04/14/whats-auto-gpt-new-autonomous-ai-agents-can-act-on-their-own-rewrite-their-own-code.html |access-date=2023-04-16 |publisher=The Toronto Star |language=en |archive-date=2023-04-16 |archive-url=https://web.archive.org/web/20230416193947/https://www.thestar.com/business/technology/2023/04/14/whats-auto-gpt-new-autonomous-ai-agents-can-act-on-their-own-rewrite-their-own-code.html |url-status=live }}</ref>

== 存在的问题 ==
目前尚不确定AutoGPT是否会找到实际应用。除了大型语言模型所营造的[[幻觉 (人工智能)|幻觉]]经常给使用者带来所困扰以外，AutoGPT实际上很难完成任务。这两点正是开发人员一直在努力解决的问题。即使完成了某项任务，AutoGPT有时也会忘记在后续的任务中使用此成果。例如当AutoGPT完成程序代码的编写后，有时会忘记在后续的过程中使用先前产生的代码。此外，AutoGPT难以有效的将大任务分解为子任务，经常造成子任务目标的重叠。<ref name="jina">{{Cite news |last=Xaio |first=Han |date=2023-04-13 |title=Auto-GPT Unmasked: The Hype and Hard Truths of Its Production Pitfalls |url=https://jina.ai/news/auto-gpt-unmasked-hype-hard-truths-production-pitfalls/ |access-date=2023-04-17 |publisher=Jina AI |language=en |archive-date=April 17, 2023 |archive-url=https://web.archive.org/web/20230417145451/https://jina.ai/news/auto-gpt-unmasked-hype-hard-truths-production-pitfalls/ |url-status=live }}</ref>

== 应用程序 ==
开发人员开发了AgentGPT，<ref>{{Cite web |title=AgentGPT: Autonomous AI in your browser 🤖 |url=https://agentgpt.reworkd.ai/ |access-date=2023-04-23 |website=agentgpt.reworkd.ai |language=en |archive-date=2023-04-16 |archive-url=https://web.archive.org/web/20230416173216/https://agentgpt.reworkd.ai/ |url-status=live }}</ref>它可以将AutoGPT整合到个人浏览器中，能够让没有编程能力的使用者制作自己的软件代理程序。<ref>{{Cite news |last=Mauran |first=Cecily |date=2023-04-14 |title=What is Auto-GPT and why are hustle bros hype for it? |url=https://mashable.com/article/autogpt-chatgpt-hustle-bros-explainer |access-date=2023-04-16 |publisher=Mashable |language=en |archive-date=2023-04-16 |archive-url=https://web.archive.org/web/20230416193947/https://mashable.com/article/autogpt-chatgpt-hustle-bros-explainer |url-status=live }}</ref>此外还有人利用AutoGPT制作了ChaosGPT聊天机器人，该机器人甚至扬言要“[[通用人工智能的生存风险|毁灭人类]]”<ref>{{Cite news |first=Chloe |last=Xiang |title=AI Tasked With 'Destroying Humanity' Now 'Working on Control Over Humanity Through Manipulation' |url=https://www.vice.com/en/article/z3mxe3/ai-tasked-with-destroying-humanity-now-working-on-control-over-humanity-through-manipulation |access-date=2023-04-16 |publisher=Vice News |language=en |archive-date=2023-04-16 |archive-url=https://web.archive.org/web/20230416200223/https://www.vice.com/en/article/z3mxe3/ai-tasked-with-destroying-humanity-now-working-on-control-over-humanity-through-manipulation |url-status=live }}</ref>，不过没有成功。<ref>{{Cite news |first=Cassidy |last=Ward |date=2023-04-16 |title=ChaosGPT is trying to destroy humanity; fortunately the AI is adorably bad at it |url=https://www.syfy.com/syfy-wire/chaosgpt-trying-to-destroy-humanity-bad-at-it |access-date=2023-04-20 |work=Syfy Wire |language=en-US |archive-date=2023-04-16 |archive-url=https://web.archive.org/web/20230416220952/https://www.syfy.com/syfy-wire/chaosgpt-trying-to-destroy-humanity-bad-at-it |url-status=live }}</ref><ref>{{Cite news |first=Aaron |last=Reich |date=2023-04-14 |title=1/3rd of scientists fear AI decisions could spark nuclear-level disaster |url=https://www.jpost.com/business-and-innovation/tech-and-start-ups/article-739227 |access-date=2023-04-21 |work=The Jerusalem Post |language=en-US |archive-date=2023-04-19 |archive-url=https://web.archive.org/web/20230419052429/https://www.jpost.com/business-and-innovation/tech-and-start-ups/article-739227 |url-status=live }}</ref>

== 参考文献 ==
{{reflist|30em}}
== 延伸阅读 ==
* {{Cite web |last=Pounder |first=Les |date=2023-04-15 |title=How To Create Your Own Auto-GPT AI Agent |url=https://www.tomshardware.com/how-to/auto-gpt-ai-agent |access-date=2023-04-16 |publisher=Tom's Hardware |language=en}}
* {{Cite news |last=Wiggers |first=Kyle |date=2023-04-22 |title=What is Auto-GPT and why does it matter? |url=https://techcrunch.com/2023/04/22/what-is-auto-gpt-and-why-does-it-matter/ |access-date=2023-04-23 |website=TechCrunch |language=en-US}}
== 外部链接 ==
* {{Official Website|https://agpt.co}}
* 在[[GitHub]]的[https://github.com/Significant-Gravitas/AutoGPT 官方原始碼庫] {{Wayback|url=https://github.com/Significant-Gravitas/Auto-GPT |date=20230610030305 }}
{{OpenAI}}
[[Category:2023年软件]]
[[Category:人工智能]]
[[Category:大型语言模型]]
[[Category:基于转换器的生成式预训练模型]]

AutoGPT - 版本历史

imported>TangZhi：​TangZhi移動頁面Auto-GPT至AutoGPT：​已更名

imported>TangZhi：TangZhi移動頁面Auto-GPT至AutoGPT：已更名