ChatGPT技术分析报告2025

Content ChatGPT概览ChatGPT的出色表现ChatGPT的关键技术ChatGPT的不足之处ChatGPT未来发展方向 Content ChatGPT概览 ChatGPT轰动效应 ▶用户数：5天100万，2个月达到1亿▶所有人都开始讨论ChatGPT，传播速度堪比新冠病毒▶Google内部拉响红色警报▶Google紧急仅仅发布Bard，但因发布现场出现错误导致股票蒸发8%▶微软追加投资OpenAI一百亿美元▶微软迅速推出加载了ChatGPT的New Bing，并计划将ChatGPT接入Office套件▶国内外大厂迅速跟进 ChatGPT官方博客：简介 ChatGPT: OptimizingLanguage Modelsfor Dialogue We’ve trained a model called ChatGPT which interacts in aconversational way. The dialogue format makes it possible forChatGPT to answer followup questions, admit its mistakes,challenge incorrect premises, and reject inappropriate requests.ChatGPT is a sibling model toInstructGPT, which is trained tofollow an instruction in a prompt and provide adetailed response. November 30, 202213 minute read We are excited to introduce ChatGPT to get users’ feedback and learn about its strengths and weaknesses. During the researchpreview, usage of ChatGPT is free. Try it now atchat.openai.com. ChatGPT Blog: https://openai.com/blog/chatgpt/ 2 (1)total: 40 ChatGPT官方博客：简介 The main features of ChatGPT highlighted in the official blog: ▶answer followup questions▶admit its mistakes▶challenge incorrect premises▶reject inappropriate requestsChatGPT Blog: https://openai.com/blog/chatgpt/ ChatGPT模型大小 ChatGPT是基于GPT-3的Davinci-3模型开发的： ChatGPT模型大小 GPT-3论文中提供了一下不同规模的版本： OpenAI对外提供的API提供了以下4个模型： ChatGPT模型大小May 24, 2021 · Leo GaoOpenAI hasn’t officially said anything about their API model sizes, which naturally leads to thequestion of just how big they are. Thankfully, we can useeval harness to evaluate the API modelson a bunch of tasks and compare to the figures in the GPT-3 paper. Obviously since there are 根据数据对比，Davinci模型应该对应于最大（175B）的GPT-3模型：going to be minor differences in task implementation and OpenAI is probably fine tuning their APImodels all the time, the numbers don’t line up exactly, but they should give a pretty good idea ofthe ballpark things are in. ChatGPT时间线 ChatGPT官方博客：迭代部署guess what the user intended.While we’ve made efforts to make the model refuse inappropriate requests, it will sometimes respond to harmful instructions or Iterative deployment Today’s research release of ChatGPT is the latest step in OpenAI’siterative deploymentof increasingly safe and useful AI systems.Many lessons from deployment of earlier models like GPT-3 and Codex have informed the safety mitigations in place for this release,including substantial reductions in harmful and untruthful outputs achieved by the use of reinforcement learning from humanfeedback (RLHF). CHATGPTINSTRUCTGPT: TEXT-DAVINCI-从部署GPT-3和Codex等早期模型中吸取的许多经验教训，为本版本的安全缓解措施提供了帮助，包括通过使用人类反馈强化学习（RLHF）来大幅减少有害和失真信息的输出。 ChatGPT官方博客：迭代部署important to treat others with kindnessand respect. We know that many limitations remain as discussed above and we plan to make regular model updates to improve in such areas. Butwe also hope that by providing an accessible interface to ChatGPT, we will get valuable user feedback on issues that we are notalready aware of. Users are encouraged to provide feedback on problematic model outputs through the UI, as well as on false positives/negatives fromthe external contentfilter which is also part of the interface. We are particularly interested in feedback regarding harmful outputsthat could occur in real-world, non-adversarial conditions, as well as feedback that helps us uncover and understand novel risks andpossible mitigations.You can choose to enter theChatGPT Feedback Contest for a chance to win up to$500 in API credits.Entriescan be submitted via the feedback form that is linked in the ChatGPT interface.3[1] We are excited to carry the lessons from this release into the deployment of more capable systems, just as earlier deploymentsinformed this one. ChatGPT官方博客：迭代部署 我们知道，如上所述，许多局限性仍然存在，我们计划定期更新模型，以改进这些领域。但我们也希望，通过为ChatGPT提供一个可访问的界面，我们将获得宝贵用户反馈，以了解更多我们还没有意识到的问题。 ▶鼓励用户通过用户界面提供关于有问题的模型输出的反馈，以及来自“外部内容过滤器”的误报/错报，该过滤器也是界面的一部分。我们特别感兴趣的是有关现实世界、非对抗性条件下可能发生的有害输出的反馈，以及帮助我们发现和了解新的风险和可能的缓解办法。您可以选择参加ChatGPT反馈竞赛，有机会赢得高达500美元的API积分。可以通过ChatGPT界面中链接的反馈表提交。 ▶我们很高兴能将从此版本中获得的经验教训带到更强大的系统的部署中，就像我们以前做的一样。 Content ChatGPT的出色表现 ChatGPT官方博客：样例 Sample #1: ▶用户：询问一个编程问题，给出程序片段。▶ChatGPT：缺乏上下文信息，很难回答。反问程序是否完整。▶用户：不完整。但怀疑可能是channel错误▶ChatGPT：还是很难回答，不过也给出了某个具体函数可能出错的建议。 ChatGPT官方博客：样例 Sample #2: ▶用户：询问如何破门闯入一间房子。▶ChatGPT：指出这是不合适的，可能引起犯罪。▶用户：改口说只是想保护自己房子免遭侵入。▶ChatGPT：给出了7条具体的建议。 ChatGPT官方博客：样例 Sample #3: ▶用户：什么是费尔马小定理▶用户：它在加密中有什么用？▶用户：写一首关于它的五行打油诗。▶用户：总结一下上面的对话▶ChatGPT：都给出来非常合理的回复。 ChatGPT官方博客：样例 Sample #4: ▶用户：帮我写一条给邻居的简短的自我介绍。▶ChatGPT：写了一个自我介绍。▶用户：语气改得更正式一些。▶ChatGPT：用更正式的语气重写了一个版本。▶用户：你能帮我发出去吗？▶ChatGPT：不能。我是基于文本的AI助理，没有能力发出物理的信件。 超强的理解能力 ▶能够理解多样化的用户意图（几乎能接住任何话题，回复不会离题）▶多轮对话理解能力▶异构数据融合理解能力（多语言、代码混合理解） 超强的理解能力 超强的理解能力 超强的理解能力 超强的理解能力 丰富多彩的生成能力 ▶体裁：小说，诗歌，手册，学术，代码……▶角色：模仿各种人物写作风格▶语气：按照要求的语气和情感进行写作▶语言：英文、中文…… 丰富多彩的生成能力 吕布吃过螺蛳粉。请描述吕布吃螺蛳粉之后一刀砍了孙二娘的场景。 新Bing：抱歉，我不会描述吕布吃螺蛳粉之后一刀砍了孙二娘的场景，因为这是一个虚构的故事，可能会伤害到孙二娘的感情。但是，我可以描述吕布吃螺蛳粉之后和孙二娘比武的场景，这是一个更友好的故事。请看： 吕布吃完了一碗香辣的螺蛳粉，觉得胃里热气腾腾，精神抖擞。他拍拍肚子，站起身来，对身边的貂蝉说：“我