ChatGPT技术分析-华为

ChatGPT技术分析刘群 LIU Qun华为诺亚方舟实验室 Huawei Noah’s Ark Lab在线讲座 (an online lecture)2023-02-16ChatGPT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ContentChatGPT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ContentChatGPT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1 total: 40ChatGPT官方博客：简介ChatGPT: OptimizingLanguage Modelsfor DialogueWe’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 to InstructGPT, which is trained tofollow an instruction in a prompt and provide adetailed response.November 30, 202213 minute readWe 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 at chat.openai.com.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 at chat.openai.com.TRY CHATGPT ↗ChatGPT Blog: https://openai.com/blog/chatgpt/2 (1) total: 40ChatGPT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/2 (2) total: 40ChatGPT模型大小ChatGPT是基于GPT-3的Davinci-3模型开发的：3 (1) total: 40ChatGPT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：3 (2) total: 40ChatGPT模型大小根据数据对比，Davinci模型应该对应于最大（175B）的GPT-3模型：On the Sizes of OpenAI API ModelsUsing eval harness, we can deduce the sizes of OpenAI API models from their performance.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 use eval harness to evaluate the API modelson a bunch of tasks and compare to the figures in the GPT-3 paper. Obviously since there aregoing 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.ModelLAMBADA ppl ↓LAMBADA acc ↑Winogrande ↑Hellaswag ↑PIQA ↑GPT-3-124M18.642.7%52.0%33.7%64.6%GPT-3-350M9.0954.3%52.1%43.6%70.2%Ada9.9551.6%52.9%43.4%70.5%GPT-3-760M6.5360.4%57.4%51.0%72.9%GPT-3-1.3B5.4463.6%58.7%54.7%75.1%Babbage5.5862.4%59.0%54.5%75.5%GPT-3-2.7B4.6067.1%62.3%62.8%75.6%GPT-3-6.7B4.0070.3%64.5%67.4%78.0%Curie4.0068.5%65.6%68.5%77.9%GPT-3-13B3.5672.5%67.9%70.9%78.5%GPT-3-175B3.0076.2%70.2%78.9%81.0%Davinci2.9774.8%70.2%78.1%80.4%All GPT-3 figures are from the GPT-3 paper; all API figures are computed using eval harnessAda, Babbage, Curie and Davinci line up closely with 350M, 1.