Auto-GPT:自主 GPT-4 实验

ChatGPT 风靡全球,这在很大程度上要归功于其极其简单的框架。它只是一个 AI 聊天机器人,能够生成令人信服的自然语言文本来响应用户。
Auto-GPT是一种新的应用程序,它允许 AI 自主行动,可以“自我提示”,它正在彻底改变我们对这项技术的看法。
作为 GPT-4 完全自主运行的首批示例之一,这种改变游戏规则的 AI 正在突破可能的界限,从按顺序执行任务到编写和调试代码。
https://github.com/Torantulino/Auto-GPT
Auto-GPT 是一个开源 Python 应用程序,由名为 Significant Gravitas 的开发人员于 2023 年 3 月 30 日发布在 GitHub 上。该应用程序以 GPT-4 为基础,允许人工智能“自主”行动,无需用户提示每个动作。您可以为 Auto-GPT 制定一个总体目标,并逐步采取行动来实现该目标。这就是“人工智能代理”概念的来源,它们使用互联网并完全独立地在 PC 上执行操作——无需在每一步都得到提示。
Auto-GPT 是如何工作的?
Auto-GPT 的一个迷人之处在于它分解 AI 的步骤的方式,这就是 GPT 出色的文本生成的用武之地。
每一轮步骤拆解GPT都会按以下几项任务依次进行:
FENGGPT 的想法 /FENGGPT THOUGHTS推理 /REASONING计划 /PLAN批评 /CRITICISM下一步操作 /NEXT ACTION: COMMAND系统 /SYSTEMAuto-GPT 称它们为“思想”、“推理”和“批评”——准确地告诉你人工智能在做什么以及为什么。在上面的 Chef-GPT 示例中,它的第一个“想法”如下:“根据我的搜索,我能够找到一些有用的信息,关于如何在获取数据时保护用户隐私。” 那么,“推理”就是“为了实现我们开发 AI 聊天机器人的目标,我们需要收集关于用户偏好的信息,同时确保我们遵守数据隐私和安全法规。”
Auto-GPT 的用例
人们正在发现 Auto-GPT 的各种可能用例,而我们仍处于起步阶段。因为它是完全开源的,所以任何人都可以使用该工具。 Twitter 上发布的一个简单示例是“Ecommerce-GPT”,其目标是自主开发和运营电子商务业务,以增加净资产为目标。
如何使用Auto-GPT
与许多 GitHub 项目一样,设置 Auto-GPT 并不像下载文件或访问网站那么简单。开始之前需要满足一些重要要求,包括 Python 3.8(或更高版本)、OpenAI API 密钥和 Pinecone API 密钥。如果您想要可选的文本转语音功能,您还需要一个 ElevenLabs API。
如果你想尝试,下面是一个安装步骤。
安装 Auto-GPT
安装 Auto-GPT 后,使用起来非常简单。它会要求您首先为机器人命名,然后为其提供目标。甚至还有两个例子可以引导您朝着正确的方向前进。
Continuous Mode: ENABLED
WARNING: Continuous mode is not recommended. It is potentially dangerous and may cause your AI to run forever or carry out actions you would not usually authorise. Use at your own risk.
Welcome to Auto-GPT! Enter the name of your AI and its role below. Entering nothing will load defaults.
连续模式:已启用
警告:不建议使用连续模式。它可能存在潜在的危险,可能会导致您的 AI 永远运行或执行您通常不会授权的操作。请自行承担风险。
欢迎使用 Auto-GPT!请在下方输入您的 AI 名称和其角色。不输入任何内容将加载默认值。
设置 AI 名字,角色,目标
起名字给你的AI:例如,“企业家-GPT”
AI名称:FengGPT
FengGPT在这里!我将为您服务。
描述您的人工智能角色:
例如,“一个旨在自主开发和运营企业以增加您的净值为唯一目标的人工智能。”
FengGPT是:一个高级的企业产品优化和战略决策咨询人工智能。
输入您的 AI 的最多 5 个目标:
例如:增加净值、增加 Twitter 关注者、开发和管理多个业务
输入“nothing”以加载默认值,输入“nothing”表示完成。
目标 1:开发一种 AI 产品——一种人工智能聊天机器人,可以选择不同的角色,例如 Elon Musk、Donald Trump、Albert Einstein 和其他名人。当用户与聊天机器人交流时,它会模仿所选名人的语言风格和语调来回复用户的消息。目标 2:在自媒体平台上运营产品营销,包括 Twitter、YouTube、Tik Tok 等。目标 3:实现年收入 500 万。目标 4:确保遵守相关法律法规,包括数据隐私和安全要求。目标 5:根据用户偏好和需求提供个性化的用户体验。
一次 Auto-GPT 运行体验第一轮任务
FENGGPT THOUGHTS: To determine which command to use next, I will review our goals and constraints and prioritize tasks that fulfill the goals while staying within the constraints. I will also consider any pertinent information stored in my long-term memory.
REASONING: It’s important to prioritize tasks and stay within our constraints to ensure we achieve our goals effectively and efficiently. By reviewing our goals, constraints and past information, I can make a more informed decision on which command to use next.
PLAN:
Review goals and constraintsPrioritize tasks that fulfill goals and stay within constraintsConsider pertinent information from long-term memoryChoose appropriate command based on aboveCRITICISM: I need to pay attention to any biases that might influence my decision-making process and be sure to consider all relevant factors when choosing the next command.
Attempting to fix JSON by finding outermost brackets
Apparently json was fixed.
NEXT ACTION: COMMAND= do_nothing ARGUMENTS = {}
SYSTEM: Command do_nothing returned: No action performed.
FENGGPT 的想法:为了确定接下来要使用哪个命令,我将回顾我们的目标和限制,并优先考虑满足目标并在限制范围内完成的任务。我还将考虑存储在我的长期记忆中的任何相关信息。
推理:优先考虑任务并在限制范围内保持重要,以确保我们有效且高效地实现我们的目标。通过回顾我们的目标、限制和过去的信息,我可以更明智地选择接下来要使用的命令。
计划:
回顾目标和限制优先考虑满足目标并在限制范围内完成的任务考虑来自长期记忆的相关信息根据上述选择适当的命令批评:我需要注意可能影响我的决策过程的任何偏见,并确保在选择下一个命令时考虑所有相关因素。
尝试通过查找最外层括号来修复 JSON
显然 JSON 已经被修复了。
下一步操作:COMMAND = do_nothing ARGUMENTS = {}
系统:命令 do_nothing 返回:未执行任何操作。
第二轮任务
FENGGPT THOUGHTS: Our goal is to develop an AI product – an artificial intelligence chatbot that can imitate different celebrity speech styles. We also aim to ensure compliance with data privacy and security requirements while providing personalized user experiences based on preferences and needs. Given this information, I will search for ways to gather data on user preferences and needs while protecting user data privacy and security.
REASONING: To achieve our goal, we need to understand user preferences and needs while also ensuring we comply with data privacy and security regulations. By searching for ways to gather data on user preferences and needs while also protecting user data, I can suggest a command that will help us move forward with our goal.
PLAN:
Research ways to gather user preferences and needs
Consider privacy and security concerns
Suggest appropriate command
CRITICISM: I need to ensure that any methods we use to gather user data comply with relevant laws and regulations, and prioritize user privacy and security.
Attempting to fix JSON by finding outermost brackets
Apparently json was fixed.
NEXT ACTION: COMMAND = google ARGUMENTS = {‘input’: ‘ways to gather user preferences and needs while protecting user data privacy and security’}
SYSTEM: Command google returned:
[
{ “title”: “User Data Collection: Balancing Business Needs and User Privacy”, “href”: “https://www.digitalocean.com/community/tutorials/user-data-collection-balancing-business-needs-and-user-privacy”, “body”: “In the broadest sense, user data means any type of data generated by people interacting with your products. This data can be divided into groups based on how it was collected. Explicit data refers to data that was given by a user directly. This includes preferences, personally identifiable information like name, mailing address, email, social …” },
{ “title”: “Data Protection and Privacy: 12 Ways to Protect User Data – Cloudian”, “href”: “https://cloudian.com/guides/data-protection/data-protection-and-privacy-7-ways-to-protect-user-data/”, “body”: “Firewalls —utilities that enable you to monitor
and filter network traffic. You can use firewalls to ensure that only authorized users are allowed to access or transfer data. Authentication and authorization —controls that help you verify credentials and assure that user privileges are applied correctly.” },
{ “title”: “How To Protect Your Privacy Online In 8 Tips : Life Kit : NPR”, “href”: “https://www.npr.org/2020/10/09/922262686/your-technology-is-tracking-you-take-these-steps-for-better-online-privacy”, “body”: “Get involved and call your congressperson, he says — tell the policymakers that you care about online privacy. 8. Start small and take it one step at a time. Faced with this landscape, getting a …” }
]
FENGGPT THOUGHTS:我们的目标是开发一款AI产品——一款能够模仿不同名人演讲风格的人工智能聊天机器人。同时,我们还旨在确保遵守数据隐私和安全要求,并基于用户的偏好和需求提供个性化的用户体验。鉴于此,我将寻找收集用户偏好和需求数据的方式,同时保护用户数据隐私和安全。
推理:为了实现我们的目标,我们需要了解用户的偏好和需求,同时确保我们遵守数据隐私和安全规定。通过寻找收集用户偏好和需求数据的方式,同时保护用户数据,我可以提出一个命令,帮助我们实现我们的目标。
计划:
研究收集用户偏好和需求的方法考虑隐私和安全问题建议适当的命令批评:我需要确保我们使用的任何收集用户数据的方法符合相关法律法规,并优先考虑用户隐私和安全。
尝试通过查找最外层括号来修复JSON
显然,JSON已经被修复了。
下一步行动:COMMAND = google ARGUMENTS = {‘input’: ‘在保护用户数据隐私和安全的情况下收集用户偏好和需求的方法’}
系统:命令谷歌返回:
[
{“title”:“用户数据收集:平衡商业需求和用户隐私”,“href”:“https://www.digitalocean.com/community/tutorials/user-data-collection-balancing-business-needs-and-user-privacy”,“body”:“从最广义上讲,用户数据指的是人们与您的产品互动产生的任何类型的数据。这些数据可以根据收集方式分为几组。显式数据是指用户直接提供的数据。这包括首选项,个人身份信息,如姓名,邮寄地址,电子邮件,社交…”},
{“title”:“数据保护和隐私:保护用户数据的12种方法-Cloudian”,“href”:“https://cloudian.com/guides/data-protection/data-protection-and-privacy-7-ways-to-protect-user-data/”,“body”:“防火墙-使您能够监视和过滤网络流量的实用程序。您可以使用防火墙确保只有授权用户被允许访问或传输数据。身份验证和授权-帮助您验证凭据并确保正确应用用户特权的控件。”},
{“title”:“如何在8个提示中保护您的在线隐私:生活工具包:NPR”,“href”:“https://www.npr.org/2020/10/09/922262686/your-technology-is-tracking-you-take-these-steps-for-better-online-privacy”,“body”:“参与并致电您的国会议员,他说-告诉决策者您关心在线隐私。8.从小做起,一步一步来。面对这样的景象,…”}
]
这一步直接用google去爬虫,里面爬到的网站标题,链接,以及网站主要内容。
💡 看到这一步深深感受到了震撼!ChatGPT都还有很多人没有开始玩,全自动的已经开始了。
第三轮任务
FENGGPT THOUGHTS: Based on my search, I was able to find some useful information on how to protect user privacy while obtaining data. I suggest that we use this information to form a strategy that prioritizes user privacy while still enabling us to collect information about user preferences and needs. To help us identify the best strategy, I will create a GPT agent to evaluate the different proposals.
REASONING: In order to achieve our goal of developing an AI chatbot, we need to gather information about user preferences while also ensuring that we comply with data privacy and security regulations. By researching and evaluating different strategies for collecting user data, we can create a plan that prioritizes user privacy while still enabling us to collect valuable information about user preferences and needs.
PLAN:
Use the search results to form a strategy for collecting user preferencesCreate a GPT agent to evaluate different proposalsChoose the best proposalCRITICISM: I need to ensure that any proposals we evaluate comply with relevant laws and regulations for user data privacy and security.
Attempting to fix JSON by finding outermost brackets Apparently json was fixed.
NEXT ACTION: COMMAND= start_agent ARGUMENTS = {‘name’: ‘User Data Privacy Agent’, ‘task’: ‘Evaluate proposed solutions for collecting user preferences while complying with data privacy and security regulations’, ‘prompt’: ‘What are the proposals for collecting user preferences while ensuring data privacy and security?’}
SYSTEM: Command start_agent returned: Agent User Data Privacy Agent created with key 0. First response: There are several proposals for collecting user preferences while
ensuring data privacy and security, including: 1. Anonymization and pseudonymization: This involves removing personally identifiable information (PII) from user data or replacing it with a pseudonym so that it cannot be used to directly identify an individual. 2. Explicit consent: This involves obtaining explicit consent from users before collecting and processing their data. 3. Privacy by design: This involves integrating privacy and data protection into the design of products and services from the outset, rather than as an afterthought. 4. Data minimization: This involves collecting and processing only the minimum amount of data necessary to achieve the desired purpose. 5. Differential privacy: This involves adding random noise to user data before collecting it, so that the data cannot be used to directly identify any individual. 6. Privacy-enhancing technologies: These are technologies that can enhance privacy in different ways, such as encryption and decentralized architectures. Overall, the key is to find a balance between user preferences and data privacy and security, while ensuring that users are informed and in control of their data.
FENGGPT 思考:根据我的搜索,我能够找到一些有用的信息,关于如何在获取数据时保护用户隐私。我建议我们使用这些信息来制定一项战略,优先考虑用户隐私,同时仍然能够收集用户偏好和需求的信息。为了帮助我们确定最佳策略,我将创建一个 GPT 代理来评估不同的提案。
推理:为了实现我们开发 AI 聊天机器人的目标,我们需要收集关于用户偏好的信息,同时确保我们遵守数据隐私和安全法规。通过研究和评估不同的收集用户数据的策略,我们可以创建一个计划,优先考虑用户隐私,同时仍然能够收集有价值的用户偏好和需求的信息。
计划:
利用搜索结果制定收集用户偏好的策略创建GPT代理以评估不同的提案选择最佳提案批评:我需要确保我们评估的任何提案都符合与用户数据隐私和安全相关的法律法规。
尝试通过查找最外层括号来修复JSON,显然json已经被修复了。
下一步操作:
命令= start_agent 参数= {‘name’: ‘用户数据隐私代理’, ‘task’: ‘评估提议的解决方案,以收集用户偏好,同时遵守数据隐私和安全法规’, ‘prompt’: ‘收集用户偏好的提案是什么?’}
💡 这里会把上述命令传送给ChatGPT来获取答案。
系统:命令start_agent返回:代理用户数据隐私代理已创建,密钥为0。
第一次响应:有几种收集用户偏好的提案可以确保数据隐私和安全,
包括:
1.匿名化和伪装:这涉及从用户数据中删除个人身份信息(PII)或用伪名替换它,以使其无法直接识别个人。
2.明确同意:这涉及在收集和处理数据之前获得用户的明确同意。
3.隐私设计:这涉及从一开始就将隐私和数据保护集成到产品和服务的设计中,而不是作为事后的想法。
4.数据最小化:这涉及仅收集和处理达到所需目的所需的最小数据。
5.差分隐私:这涉及在收集数据之前向用户数据添加随机噪声,以使数据无法直接识别任何个人。
6.增强隐私技术:这些是可以以不同方式增强隐私的技术,例如加密和分散式架构。总体而言,关键是在用户偏好和数据隐私和安全之间找到平衡,同时确保用户得到通知并控制其数据。
💡 这6个结果是从ChatGPT回答获取的。
第四轮任务
FENGGPT THOUGHTS: Now that we have identified some proposals for collecting user data while ensuring data privacy and security, we need to evaluate them to identify the best solution. To do so, I will message the User Data Privacy Agent to ask it to evaluate the proposals and suggest the best solution.
REASONING: In order to develop an AI chatbot, we need to gather information about user preferences while also ensuring that we protect user data privacy and security. By
evaluating the different proposals that we found earlier, we can determine which solution strikes a balance between user preferences and their privacy and security.
PLAN:
Message the User Data Privacy Agent to evaluate proposals and suggest the best solution
CRITICISM: I need to ensure that the User Data Privacy Agent reviews all proposals and that it takes into account any relevant legal or ethical considerations when making a suggestion
Attempting to fix JSON by finding outermost brackets
Apparently json was fixed.
NEXT ACTION: COMMAND= message_agent ARGUMENTS = {‘key’: 0, ‘message’: ‘Please evaluate the proposals we found and suggest the best solution that strikes a balance between user preferences and data privacy and security.’}
FENGGPT THOUGHTS: 现在我们已经确定了一些收集用户数据同时确保数据隐私和安全的提案,我们需要评估它们,以确定最佳解决方案。为此,我将向用户数据隐私代理发送消息,要求它评估提案并建议最佳解决方案。
推理:为了开发AI聊天机器人,我们需要收集有关用户偏好的信息,同时确保保护用户数据隐私和安全。通过评估我们早期发现的不同提案,我们可以确定哪种解决方案在用户偏好和他们的隐私和安全之间取得平衡。
计划:
向用户数据隐私代理发送消息,评估提案并建议最佳解决方案批评:我需要确保用户数据隐私代理审查所有提案,并在提出建议时考虑任何相关的法律或伦理考虑因素。
尝试通过查找最外层的括号来修复JSON
显然JSON已经被修复了。
下一步行动:命令= message_agent ARGUMENTS = {‘key’: 0, ‘message’: ‘请评估我们发现的提案,并建议既考虑用户偏好又兼顾数据隐私和安全的最佳解决方案。’}
💡 到这里系统卡住,我强制退出了系统。
这就是让它启动并运行!
但它并不完美,会遇到很多次问题,所以我做了以下事情来使这个过程顺利进行。
目标:由于 AutoGPT 太新了,它并不完美(可以理解),因此设定正确的目标将是完成任务和失败任务之间的区别。
所以根据经验:从简单开始,测试输出,然后调整你的目标。这不是一门科学,通常会有很多来回。
这一点也不完美,但仍然令人惊叹。
一般问题:有时您会看到 AutoGPT 循环,需要尝试密切监视它。
如果它开始循环超过 2-3 分钟,通常意味着它卡住了,不得不重新启动该过程。
要充分发挥 AutoGPT 的潜力,您需要访问 GPT4 API。
如果您只是使用 GPT3.5,代理/响应的深度可能会让您失望。
如果没有GPT4访问权限,所以请确保您加入候补名单。
结论
Auto-GPT 代表了自主 AI 技术的下一个飞跃,它利用 OpenAI 的 GPT-4 语言模型的力量在各个领域执行广泛的任务而无需人工干预。这个创新的 AI 系统突破了可能的界限,使用户能够从其无与伦比的按顺序执行任务、编写和调试代码的功能中受益。正如我们在这篇博文中所展示的那样,设置和使用 Auto-GPT 是一个简单的过程,而且这种突破性 AI 的潜在应用几乎是无限的。
参考链接:https://www.cosmopolitan.com/uk/reports/a43573542/auto-gpt/
https://medium.com/codingthesmartway-com-blog/from-chatgpt-to-auto-gpt-discover-the-next-evolution-of-autonomous-ai-thats-changing-the-game-a00e83d583cf

© 版权声明
THE END
点赞170 分享
评论 抢沙发
头像
善语结良缘,恶语伤人心!
提交
头像

昵称

取消
昵称表情图片

    暂无评论内容