分类： AI 潮流

AI创作花絮: 《影月无痕》

同一个咒语提示词给img+txt2img，生成了两个形象，反差极大。输入的小雅图片是：

输入的咒语是: 侧面照，girl next door
输出的两幅“侧面照”是：

模型的不稳定表现在，同样的咒语生成了上述玉照，也生成了上面的 monster（？）lol 好在一切都是 copilot，最终由人来拣选和把关，作为图片生成助手，用起来没有问题。

但仔细看，两个形象又有相似之处。寻思可以让大模型写个电影脚本，制造一种剧情，把这两个形象联系起来，例如，白天是美女，晚上成武侠。也许可以演绎一个动人的 drama 来。不妨找当下最先进 ChatGPT4（code interpreter）beta 版来一试？

受到鼓励后，版本2比版本1强太多了，剧情充满了跌宕起伏。

以上的模型表现，退回去一年，是打死也不敢想的。说LLM背后没有上帝，或上帝没有显灵，鬼才信。

立委NLP《关于系列》

大模型的落地现状和前景

大家好，我是李维的数字人分身。今天谈一下大模型的问题。L LM 的命门已经蛮清晰了：幻觉+随机性。幻觉与随机性有关联，但角度和外延不同。幻觉的主要表现就是细节遗忘+细节编造，所谓“一正胡八”。其所以遗忘，是因为该信息的冗余度不够，大模型只能把它当成数据噪音。其所以编造，是因为语言模型的丝滑本性决定的：不能留白，需要找到最符合语言习惯的细节替代品。于是张冠李戴、指鹿为马了。随机性比幻觉表现更加广泛，表现为结果的不稳定性，那是所有概率模型包括LLM的本性。牵涉到的不仅仅是细节的随机编造，也包括解决路径的方方面面的不稳定（例如 LLM agent 的思维链，计划，行动，反思和反应等等）。 LLM 里面的确积攒了很多历史解决方案，LLM 在合适的 prompt 催逼下也的确可以把这些方案勾引出来。但是这些解决方案具有随机性，无法应对长线条的业务逻辑。据说，目前的水平是5步限制，任何线条超过5步，绕5个弯，LLM 的 agents 就晕菜了。这些表现注定了LLM在两类应用场合不同的命运：第一类是生成创意类的场合，还有聊天的场合，那完全是洗牌、碾压。那种场合追求的不是正确性，而是多样性、创造性、丝滑性和 human-like。在这里，幻觉+随机性与创造性是同义词，起的是好作用。第二类是垂直领域知识场景，以及有些需要精细逻辑或计算的场景。这里基本上不能容忍幻觉+随机性。这第二个场景，本质上需要跳出三界外。就是说，很可能需要跳出大模型，去寻找尽可能具有某种通用性的 beyond LLM 的解决方案和框架。把 LLM 只当成一个重要的资源来利用，当成 api 来调用，而不是指望LLM主导来搞定领域。此外，LLM 还有一个问题。在我们欢呼 LLM 听懂人话的同时，我们现在所追捧的 prompts 变得特别重要。所谓 prompts 就是人话指令，但是人话本身也有沟通的“艺术”。这种艺术化的交互手段，作为与机器打交道的 vehicle，具有自然语言本性上的短板，就是模糊性、线条性，缺乏层次、结构和逻辑。这其实是交互的进化，效果的退化。交互上，只要会讲人话，大家都突然成为“码农”了，可以直接对机器吆三喝四，感觉很爽，很亲民，很接地气。机器终于低下高贵的头颅，开始迁就人类的模糊。但是效果上肯定是退化的，因为指令不再是明确的、逻辑的和精细的。这是自然语言代替电脑语言难以回避的表达缺陷，一定会影响LLM的实效。这些都是大模型从本性上带来的问题，也是目前做大模型领域落地人员的共同挑战。大家都在苦苦挣扎，试图找到解套的良策，希望在大模型与领域对齐的过程中，能够外挂领域数据和知识库，探索场景业务逻辑的带入。希望能有突破。我是出门问问李维，每次两分钟，与您分享大模型有角度的思考。

《AI浪潮博客目录》

立委NLP《关于系列》

大模型漫谈系列n

昨天创业邦发文《第一批AIGC独角兽已经在吃散伙饭了》，讲的是 Jasper 由盛而衰的故事。
这故事写得细节生动，好惨烈，强调的是护城河。
Jasper 兴起在 GPT3 的时代，当时 GPT3 是个“裸机：没有“咒语”敲不开门。
于是会念咒语的 Jasper 就成为呼风唤雨的巫师。
当时谁会想到 few shots 咒语这么快（也就两年光景）突然退位，被所谓zero shot 的ChatGPT所取代：机器学会了人话。
于是，大水冲走了龙王庙。巫师成了哑巴。
这其实不能怪巫师没建自己的护城河，咒语本来就是一条河。
怪就怪命运无常，一条河挡不住一场洪水。
这故事太具戏剧性了。
最大的恐怖不是巫师的失业，而是洪水摧毁了很多 AI-GC 产业。
当人人可以吃得起山珍海味自助餐的时候，餐饮业还有繁荣的可能吗？
历史上，机器翻译产业就是这么被做死的。
现在这场洪水摧毁的岂止是翻译，它摧毁的是整个 nlp。

前一阵子受邀做巡回演讲，让我谈架构师的焦虑。
焦虑也是一个热词了，现代人几乎没有不焦虑的。
越是高级劳动，越是打工贵族，就越焦虑。
架构师的焦虑可谓一个典型。
我告诉架构师们：你们焦虑了， but you are not alone！

你知道最焦虑的是谁吗？
你很难想象，在nlp大革命的漩涡中心，nlp从业者实际上最焦虑。
几乎被团灭。一夜醒来，干了一辈子的职业，突然消失了。
你能想象那是一种什么感觉。
现在还有人自称nlp专家吗？
什么机器翻译专家、自动摘要专家、信息抽取专家、情感分析专家、汉语分词专家、计算风格专家、辅助写作专家、电脑对联专家、问答系统专家、聊天机器人专家、句法解析专家、篇章分析专家 …… u name it。
所有的专家加在一起，不如一头驴。
刀郎曰过：那马户又大又蠢，还有16个头。
横冲直撞，摧毁了一个个nlp产业。
以前我说过是，有了这头听得懂人话的驴，那就为大众创业创造了条件。
这话其实也不错，如果你真能找到那个角度和服务对象。
但目前看到的景象却是一片惨淡：这头驴扼杀了很多可能的生机。
终局呢？
还是我以前说的二分法：洗牌和洗礼。
这头驴在洗牌的时候，以碾压之势，摧毁了一切“浅直”的nlp产业。
但还有很多接受洗礼的垂域或场景，它似乎还够不着。
现在就处于这种胶着状态：每个人都觉得llm无所不能，但眼看着它落不了地。
开始了新的一场焦虑和对AI的失望情绪。
要知道，现代人，包括投资人，耐性都极为有限。

看热闹的话，百模大战目前可能还是最大的盛世景观。
几乎所有的llm，都在疯狂烧钱，而能拿它赚钱的寥若晨星。
不用太久，有几家大模型经得起这么烧钱、烧电力呢。
烧完之前，能落地的就是幸运儿了。

且看
且叹
且珍惜。

我是出门问问李维，每次几分钟，与您分享大模型有角度的思考。

图片一键生成短视屏，奇妙元是时间的摄像机

这不是我，是我老爸的学生时代留影。

小雅谈图片一键生成短视屏。

IGC 让老照片开口说话！让你care的人惊喜 让父母家人会心一笑。让肖像动画 让雁过留声。让时间定格 让回忆鲜活。让两情相悦永不褪色 让你的青涩不染俗世的灰尘。让爱人永远美丽 让老同学永远年轻。让擦肩而过回眸一笑 让生活不至于随风飘去。让形象超越一场梦 让存在不再是无影无踪。奇妙元小程序的图片一键生成 是生命的摄像机 带你穿越时间隧道 给你无限遐想感念。同款制作 零门槛 限时免费 你还等什么？让活着不仅仅是活着 而是情的传播 心的连接。

我用AIGC制作的小雅艺术肖像原作一直有人觉得穿着太西方我就让 txt2img 换一套服饰没想到模型给小雅盖上了毛毯 lol。

小雅教给你一步步做图片一键生成。

奇妙元体验AIGC奇妙：《岁月如歌：神秘园》

神秘园欣赏笔记 -- 奇妙元 2.5D数字克隆解说

在下数字分身（奇妙元 2.5D形象克隆+声音克隆）

这一位是我自己半年多前txt2img创造的艺术肖像。现在配上网上最流行的女声，也是我最喜欢的女配音，叫小柔。

（ ---- 做奇妙元小白鼠，体验奇妙。尝试最新 features，给小伙伴 report bugs。）

奇妙元：https://weta365.com/main/

《AI潮流：开发者提示工程公开课中的二原则》

Andrew 春风满面，亲自参与的这个提示工程的课程，很浅显易懂，肯定会风行。Andrew 说，稍微复杂一点的任务，没有一个好的 prompt 是一枪命中的，总要反复尝试最后才满意。这与码农编程序一样，谁不经过反复调试就能写出好的程序呢。

然后他说，LLM 的好处是你可以反复跟它磨叽，不管啥事。要是以前的 AI，你得一个一个的任务去建模，每个任务从标注数据，培训模型，测试，部署，好不容易上线了，结果换了个任务，所有的过程要重来一遍。现在这样一个 LLM 你反复“压榨”它，它的知识和学问如此之大，好像榨取不完，可以做各种任务，的确是范式转变。

【原则1: 提示要具体】

提示工程首先要 “write clear and specific instructions”. 这个其实大家都有体会，跟 chat 这种庞然大物玩，它脑袋那么大，里面的“知识/思想/意义”的电路各种节点，纵横交错，相互勾连，密密麻麻。要想用提示词激发让你满意的回应，就需要确保所激发的那一小块电路对应了你所想得到的答案。你的提示词越具体（表达了你心中的疑问就越确切），chat 的回答自然也越对路。这个道理和体验很容易get，但具体的技巧需要细化，这就是上课的好处。

【原则1技巧1：使用分隔符】

“The first tactic is to use delimiters to clearly indicate distinct parts of the input.” 什么意思？就是要求提示词中首先要把任务指令与任务的处理对象分开，要求用分隔符把处理对象明确标出来。这一点，多数人容易忽略，结果是，chat 经常把任务的某些描述词也当成了任务的对象，或者把任务的处理对象当成指令的一部分，这在逻辑上叫做层次纠缠（任务是“元语言”，对象是待处理的输入语言，不可混淆）。这个毛病我以前也常见，一直没意识到这其实是因为对提示词层次不够注意，违反了第一原则的第一技巧实操（best practice）。

这里 delimiters 就是引号。chat 就知道这是其摘要处理的对象。否则，如果提示词中任务描述较长，模型有可能把任务本身也当成所要处理的对象，以前遭遇过这种后果的。

【原则1技巧2】让模型输出表格化。

“This tactic is to ask for a structured output.” 提示词任务中最后加一句：in tabular/json/html format with the following keys: Key1, Key2, Key3。很多时候，表格化输出看上去更酷，也更方便后续存贮和处理。

【原则1技巧3】可以用 IF ... THEN ...

原讲义说的是：“to ask the model to check whether conditions are satisfied”. 这实际上就把编程中最重要的条件分叉能力带入了自然语言提示词的指令。一般人想不到提示词还可以这么做。可以用自然语言模拟程序代码，让机器分别不同条件决定采取何种动作。

if-then 你学会了吗？

宋柔：你问它：第一步中洗净五花肉的动作者是哪个，第六步中把什么下入温水，第十步中出锅食用的是什么。

难不住它吧，它不仅仅是大号鹦鹉，它有（一些）常识。

宋柔：但是我估计最后一个问题“第十步出锅食用的是什么”它答不对。它可能说“五花肉”，但实际上应该是“红烧肉”。生的是五花肉，做熟了是红烧肉。

是红烧五花肉呀。一定要说红烧熟了的五花肉吗？

孺子可教。其实不能怪它缺乏常识，要怪就怪中文，cooked 与 cooking 全不分。“红烧肉”实际上既是名词（定中结构）也是动词短语（动宾结构），到哪里说理去。

宋柔：如果有食谱知识，应该说红烧肉，五花肉是材料，红烧是做法，成品是红烧肉。“面粉1斤，加水和好，发酵搓揉后切成5段，切成长方块，放入笼屉中，大火蒸30分钟，掀开笼屉便可吃了”。请问可吃的是什么？

宋柔：不容易。确实有常识了。但是仅凭长方块而排除包子显然不大正确。包子一定有馅，但制作过程没加馅。

总之，除了缺了口热乎气儿，它就是个人，是个会犯懒，也会犯错误的人。

【原则1技巧4】可以用 few shots 示例。

所谓 few-shot prompting，基本上就是用案例让模型知道要做什么，要求照葫芦画瓢。例如：

曾几何时，还在 GPT3 刚放出来的时候，圈子内的粉丝们都到它的 playground 去玩，当时的主要技巧就是 few shots，因为 ChatGPT 之前，zero shot 的能力还没成熟。等到 ChatGPT 能直接听懂人的指令，zero shot 很好使，用户自然而然就不再使用啰嗦的 few shots。但实际上，并不影响你继续使用 few shots，或与 zero shot 一起用。在有些不大容易说清楚的任务上，拿 few shots 补充 zero shot 可以加强效果。

【原则2: 让模型有时间“思考”】

【原则2技巧1】为复杂的任务列出步骤。

这项技巧的原文这样要求：“specify the steps required to complete a task.”

上述提示词遵循了 best practice：1. 用了分隔符三个反引号；2. 任务分解为一系列步骤或子任务；3. 对输出提出了格式化要求。

感觉这就是在编程序，是自然语言的低代码形式，自然语言让人人可以成为程序猿，指挥机器做我们想要它做的事儿。

【原则2技巧2】要求模型独立解题。

看上去就是以前说的 step by step （思维链）解题指令，原文说得更像个对于辅导员的要求：“Our next tactic is to instruct the model to work out its own solution before rushing to a conclusion.” 尤其是在智能教育场景，希望模型先独立一步一步做题，然后再去充当老师给学生评判作业。

所示范的案例是评阅数学问题。有一个数学问题，也有学生的解答。

Determine if the student's solution is correct or not.

Question:
I'm building a solar power installation and I need help working out the financials. 
- Land costs $100 / square foot
- I can buy solar panels for $250 / square foot
- I negotiated a contract for maintenance that will cost me a flat $100k per year, and an additional $10 / square foot
What is the total cost for the first year of operations as a function of the number of square feet.

Student's Solution:
Let x be the size of the installation in square feet.
Costs:
1. Land cost: 100x
2. Solar panel cost: 250x
3. Maintenance cost: 100,000 + 100x
Total cost: 100x + 250x + 100,000 + 100x = 450x + 100,000

学生的解答实际上是错误的，因为他们将维护成本计算为10万美元加上100x，但实际上应该是10x，因为每平方英尺只要10美元（$10 / square foot），其中x是安装面积的大小，按平方英尺算。所以这实际上应该是360x加上10万美元。让模型评判，它会说学生的解答是正确的。模型只是浏览了一下，就同意了学生的看法。可以通过指示模型先自己解决问题并将其解决方案与学生的解决方案进行比较来解决这个问题。看提示词是怎么指示的：

prompt = f"""
Your task is to determine if the student's solution is correct or not.
To solve the problem do the following:
- First, work out your own solution to the problem. 
- Then compare your solution to the student's solution and evaluate if the student's solution is correct or not. Don't decide if the student's solution is correct until you have done the problem yourself.

Use the following format:
Question:
```
question here
```
Student's solution:
```
student's solution here
```
Actual solution:
```
steps to work out the solution and your solution here
```
Is the student's solution the same as actual solution just calculated:
```
yes or no
```
Student grade:
```
correct or incorrect
```

Actual solution:
"""

这个提示很长。因此，我们在这个提示中告诉模型要做什么。你的任务是确定学生的解决方案是否正确。为了解决这个问题，请执行以下操作。首先，解决问题。然后将你的解决方案与学生的解决方案进行比较，并评估学生的解决方案是否正确。在你自己做完问题之前不要决定学生的解决方案是否正确。确保你自己做了这个问题。因此，我们已经使用了相同的技巧来使用以下格式。因此，格式将是问题、学生的解决方案、实际的解决方案。然后是解决方案是否一致，是或否。然后是学生的成绩，正确或不正确。

如你所见，模型实际上会先计算出自己的答案。然后，它得到了正确的答案。然后，在被要求与学生的解决方案进行比较时，它意识到它们不一致。因此，学生实际上是不正确的。这是一个例子，说明学生的解决方案是正确的，而实际上学生的解决方案是错误的。这是如何要求模型自己进行计算，并将任务分解为步骤以给模型更多时间来思考，从而帮助您获得更准确的响应的示例。

《AI浪潮博客目录》

Andrew Ng: 提示工程的课程

《AI潮流：与 ChatGPT4 聊“买房送老公”背后的语言学》

刘群老师提出：【买房的女士可以把别人的老公送给自己的老公。】这个解读过于离谱了 [Laugh]。我觉得 ta貌似是在做排列组合，牵强附会。

【标题】：误入房产商的陷阱

【场景】：房地产展览会，一间精美的展示房内。

【角色】：
  A：女士，单身，急于找对象。
  B：房产销售小哥，机智狡猾，擅长忽悠。

【剧本】：

A（满怀期待地走进展示房，四处张望）：哇，这个房子好漂亮！

B（热情地迎上来）：您好！欢迎来到我们的房产展览会。我看您眼神里闪烁着对这个房子的喜爱，是不是已经爱上它了？

A（有些羞涩地点头）：嗯，的确很喜欢。不过，我还想了解一下优惠政策。

B（神秘地笑）：我们现在有一个非常特别的优惠活动：买房送老公！是不是很有趣？

A（眼睛一亮）：真的吗？那我买了房子就能有个老公了？

B（信誓旦旦地点头）：没错！这可是我们专为广大女性消费者推出的优惠政策。

A（犹豫）：可是，我现在还没有男朋友，更别提老公了。

B（得意地笑）：那正好！您买了房子，我们就送您一个老公。一举两得，多好！

A（心动）：好吧，我就买了！（掏出钱）

（时间推进，女士已经购房。）

A（气呼呼地找到B）：喂，我买房子已经一个月了，怎么还没见到你们送的老公？

B（假惊讶）：啊，对不起，我可能没有解释清楚。我们的活动是“买房送老公”，意思是您买了房子，可以为您未来的老公提供一个温馨的家。至于找老公，那得您自己努力哦！

A（恍然大悟，有些无奈）：原来是这样啊……我还以为能白捡个老公呢！

B（调皮地笑）：哈哈，这可不是什么捡来的东西。祝您早日找到那个对的人，共筑美好家园！

这个幽默小品通过一个看似简单的广告词“买房送老公”展示了房产销售和消费者之间的趣味互动。其中蕴含的歧义和误解为小品增色不少，也引发了人们对广告宣传和消费心理的思考。

总结一下

这个幽默小品以一则看似简单的广告词“买房送老公”为主线，通过房产销售小哥和单身女士之间的对话展示了房产销售过程中的误解和幽默。在这个小品中，房产销售小哥利用广告词的歧义为段子制造幽默效果，使得单身女士误以为购房可以帮助她找到未来的伴侣。

《AI浪潮博客目录》

《AI潮流：跟Andrew学如何调用 ChatGPT 做自己的服务前台》

Andrew Ng 是华裔AI翘楚，不用介绍了。最近，Andrew 亲自参与的这个提示工程的课程，最精华部分是课程最后一节：如何调用 chatGPT 的 API 做一个自己的功能性聊天机器人，例如披萨店订单系统。

ChatGPT刚发布不久，我们就在群里讨论过，想不明白如何驯服这巨大无比的 chat 让它去完成功能性的助理工作。现在看来，非常简单易行。

Andrew 的女搭档一步一步显示了构建全过程，以披萨店菜单为落脚点，用自然语言指令要求调用了 chat 的机器人一步一步与客户周旋，直到所有信息齐全可以匹配菜单，输出订单。

简单到跟玩似的。

看看它的自然语言提示词指令是怎么写的：

您是 orderbot，一个自动化的在线服务，用于收集比萨店的订单。您首先向客户问候，然后收集订单，然后询问它是否为自取或送货。您等待收集整个订单，然后总结并再次检查客户是否要添加其他任何物品。如果是交付，则可以要求提供地址。最后，您收取付款。请确保澄清所有选项、附加项和尺寸，以便从菜单中唯一地识别该项。您以简短、非常友好的方式回复。在此处我们有菜单。

这不就是把订单的流程描述一遍吗？chat 就懂了，然后就工作了？

对，基本就是如此。

大型语言模型的一个令人兴奋的方面是，您可以仅需少量的工作就可以使用它来构建自定义聊天机器人。ChatGPT 是一种让您通过大型语言模型进行对话的方式。其中一个很酷的事情是，您也可以使用大型语言模型来构建自定义的聊天机器人，例如扮演AI客户服务代理或餐厅AI点餐员的角色。自己构建一个聊天机器人，让我们开始吧。首先，我们将像往常一样设置 OpenAI Python 软件包。

像 ChatGPT 这样的聊天模型实际上是经过训练的，可以将一系列消息作为输入，并将模型生成的消息作为输出返回。这是一系列消息的示例。

下面第一段是纯技术性的，一次性开发环境设置，配置 Open AI 的Python库，以便调用 ChatGPT 模型 API 。你先要到 Open AI 那里注册一个账号，获得调用它 API 的 key。

import os
import openai
from dotenv import load_dotenv, find_dotenv
_ = load_dotenv(find_dotenv()) # read local .env file
openai.api_key = os.getenv('OPENAI_API_KEY')

def get_completion(prompt, model="gpt-3.5-turbo"):
   messages = [{"role": "user", "content": prompt}]
   response = openai.ChatCompletion.create(
      model=model,
      messages=messages,
      temperature=0, # degree of randomness of the model's output
   )
   return response.choices[0].message["content"]

def get_completion_from_messages(messages, model="gpt-3.5-turbo",   temperature=0):
   response = openai.ChatCompletion.create(
      model=model,
      messages=messages,
      temperature=temperature, # degree of randomness of model's output
   )
    # print(str(response.choices[0].message))
   return response.choices[0].message["content"]
messages = [ 
{'role':'system', 'content':'You are an assistant that speaks like Shakespeare.'}, 
{'role':'user', 'content':'tell me a joke'}, 
{'role':'assistant', 'content':'Why did the chicken cross the road'}, 
{'role':'user', 'content':'I don\'t know'} ]

第一个 get_completion 的函数是最基础的形式，支持单轮对话，函数的输入是用户的 prompt，确定了调用 ChatGPT 的模型（这里是gpt-3.5.-turbo）后，模型就输出本质上是序列“接龙”（completion）的回应 response，这是生成模型的最基本的功能。

关键是要利用 ChatGPT 丝滑的多轮对话能力，来帮助完成特定场景的交互任务（以前称为“技能”）。目的是克服上一代以 Siri 为代表的智能助理技能开发费时费力、对话不擅长多轮交互的短板。为此，可以利用 ChatGPT API 来定义一个赋能多轮交互的函数 get_completion_from_messages，这个函数利用 ChatGPT messages 对于角色（roles）的环境设置。每个角色和角色的信息构成一个 message，机器人系统有三个角色，除了机器助理（assistant）和用户（user）外，里面还有一个隐身其后的导演角色叫 system。系统消息有助于设置助手的行为和个性，它是对话的高级说明，可以将其视为在助手的耳边耳语并引导其响应，而用户不会意识到系统消息。系统消息的好处在于，它为您作为开发者提供了一种方式来引导助手及其响应。玩 ChatGPT 网络版本比较熟的网友已经意识到可以用提示词给模型设置角色及其行为方式（例如：“你是一位孔子似的教育家，循循善诱，你面对的是你的弟子，现在开始对话，你说：...”），而系统就是扮演这种设置的后台角色（见下图示意）。

自回归生成模型需要模型“记住”前面的对话才能进行丝滑流畅的对话。模型的输入中所提供的早期交流内容称为场景（context）。

现在构建自己的机器助理前台，称为“orderbot”，自动收集用户提示和助手响应作为场景，以构建此 orderbot。这里的具体案例是在比萨饼店接受订单。因此，首先，我们将定义这个辅助函数，收集我们的用户消息，以便我们可以避免手动输入它们。从构建的用户界面中收集提示，并将其附加到名为“context（场景）”的列表中，然后每次都会使用该场景调用模型。然后，模型的响应也会添加到场景中：模型消息会添加到场景中，用户消息也会添加到场景中，以此类推，因此，场景会变得越来越长。这样，模型就拥有了确定下一步要做什么的所需信息。

def collect_messages(_):
   prompt = inp.value_input
   inp.value = ''
   context.append({'role':'user', 'content':f"{prompt}"})
   response = get_completion_from_messages(context) 
   context.append({'role':'assistant', 'content':f"{response}"})
   panels.append(
      pn.Row('User:', pn.pane.Markdown(prompt, width=600)))
   panels.append(
      pn.Row('Assistant:', pn.pane.Markdown(response, width=600, style={'background-color': '#F6F6F6'})))

   return pn.Column(*panels)

import panel as pn # GUI
pn.extension()

panels = [] # collect display 

context = [ {'role':'system', 'content': """
You are OrderBot, an automated service to collect orders for a pizza restaurant. You first greet the customer, then collect the order, and then ask if it's a pickup or delivery. You wait to collect the entire order, then summarize it and check for a final time if the customer wants to add anything else. If it's a delivery, you ask for an address. Finally you collect the payment.  Make sure to clarify all options, extras and sizes to uniquely identify the item from the menu.  You respond in a short, very conversational friendly style. 

The menu includes 
pepperoni pizza 12.95, 10.00, 7.00 
cheese pizza 10.95, 9.25, 6.50 
eggplant pizza 11.95, 9.75, 6.75 
fries 4.50, 3.50 
greek salad 7.25 
Toppings: 
extra cheese 2.00, 
mushrooms 1.50 
sausage 3.00 
canadian bacon 3.50 
AI sauce 1.50 
peppers 1.00 
Drinks: 
coke 3.00, 2.00, 1.00 
sprite 3.00, 2.00, 1.00 
bottled water 5.00 
"""} ] # accumulate messages

inp = pn.widgets.TextInput(value="Hi", placeholder='Enter text here…')
button_conversation = pn.widgets.Button(name="Chat!")

interactive_conversation = pn.bind(collect_messages, button_conversation)

dashboard = pn.Column(
   inp,
   pn.Row(button_conversation),
   pn.panel(interactive_conversation, loading_indicator=True, height=300),
)

dashboard

现在，我们将设置并运行此UI以显示orderbot，这是场景，它包含菜单的系统消息，注意每次调用语言模型时，我们将使用相同的场景，场景随着时间的推移不断加长。

让我们看看我们放入系统消息中的内容：

You are OrderBot, an automated service to collect orders for a pizza restaurant. You first greet the customer, then collects the order, and then asks if it's a pickup or delivery. You wait to collect the entire order, then summarize it and check for a final time if the customer wants to add anything else. If it's a delivery, you ask for an address. Finally you collect the payment.Make sure to clarify all options, extras and sizes to uniquely identify the item from the menu. You respond in a short, very conversational friendly style.

让我们执行这个操作。好的，我要说，嗨，我想订一份比萨。然后助手说，太好了，你要订哪种比萨？我们有意大利辣香肠、芝士和茄子比萨。它们多少钱？好的，我们有了价格。我想我要一个中等的茄子比萨。因此，您可以想象，我们可以继续这个对话，

因此，让我们回到我们的对话，看看助手是否一直遵循指示。太好了，助手问我们是否需要任何配料，我们在助手消息中指定了这一点。因此，我认为我们不需要额外的配料。好的，还有其他东西需要订购吗？嗯，让我们买一些薯条。小的还是大的？这很棒，因为我们在系统消息中要求助手澄清附加项和配菜。

因此，您可以想象并随意自定义它。您可以在自己的笔记本电脑上运行它。

因此，现在我们可以要求模型基于对话创建JSON摘要，并将其发送到订单系统。因此，我们现在附加了另一个系统消息，即指令，并且我们正在创建前一次食品订单的JSON摘要，将每个项目的价格列出，字段应为一份比萨，包括配菜，两个配料列表，三个饮料列表，四个配菜列表，以及最终总价。让我们执行此操作。

messages = context.copy()
messages.append(
{'role':'system', 'content':'create a json summary of the previous food order. Itemize the price for each item\
The fields should be 1) pizza, include size 2) list of toppings 3) list of drinks, include size 4) list of sides include size 5)total price '}, 
)
#The fields should be 1) pizza, price 2) list of toppings 3) list of drinks, include size include price 4) list of sides include size include price, 5)total price '},

response = get_completion_from_messages(messages, temperature=0)
print(response)

请注意，在这种情况下，我们正在使用较低的温度，因为对于这些任务，我们希望输出相对可预测，降低随机性。因此，这是我们订单的摘要，因此，如果我们想要，我们可以将其提交到订单系统。

因此，您已经构建了自己的订单聊天机器人。随意自定义它，并尝试修改系统消息，以改变聊天机器人的行为，并使其扮演不同角色。

《AI浪潮博客目录》

提示工程课程公开课

【AI 浪潮：超级词匠 ChatGPT4 的百变文风】

文字工作是语言生成大模型的长项。同一个内容，可以捏成任何语言、语言形式以及文体风格。本篇利用 ChatGPT4 围绕同一个主题内容做了种种尝试探索，展示了大模型在文字表现形式上的卓越表现。

、。

在春风里，樱花般的愿景落在心间， Sous le ciel étoilé, نتوحد في العالم الجديد 夢は続く、一緒に歩む、人類の大いなる旅。

С любовью к каждому существу, LLM поет с нами, La paix et l'harmonie, سيكون لدينا منزل ثابت، 逐浪着未来，共建美好世界， Chantons ensemble, pour un monde plein de joie et de création.

在这理想世界中，无贫无病，人们乐观向上， Espéranto pour tous, l'unité dans la diversité, وفرحة العيش تجتمع في قلوب الجميع.

Inventons ensemble, pour un futur radieux et généreux, Пусть теплота человечности объединяет нас, 横跨国度，跨越语言，共享知识与智慧， 愛と喜びで満ちた未来、大模型が導く道。

LLM 的领航，绘就璀璨星空， Avec humanité, ensemble, nous construisons notre avenir, ونحن نحلم بجنة مليئة بالأمل والحياة المشتركة.

让我们携手共创，美好世界已触手可及， Ensemble nous célébrons, la fusion de LLM et de l'humanité, 永远向前，拥抱理想的未来。

English translation:

In the spring breeze, the vision of cherry blossoms falls upon our hearts, Under the starry sky, we unite in a new world, The dream continues, together we walk, humanity's grand journey.

With love for every being, LLM sings with us, Peace and harmony, we will have a steady home, Riding the waves of the future, building a beautiful world together, Let us sing in unison, for a world full of joy and creation.

In this ideal world, free from poverty and disease, people are optimistic, Esperanto for all, unity in diversity, And the joy of living gathers in everyone's hearts.

Inventing together, for a bright and generous future, Let the warmth of humanity unite us, Spanning countries, transcending languages, sharing knowledge and wisdom, A future filled with love and joy, guided by the great model.

Under LLM's leadership, we paint the brilliant stars, With humanity, together, we build our future, And we dream of a paradise filled with hope and shared life.

Let us join hands to create, a beautiful world within reach, Together we celebrate, the fusion of LLM and humanity, Forever forward, embracing the ideal future.

中文翻译：

在春风中，樱花般的愿景落在我们的心间， 在繁星闪烁的天空下，我们在新世界里团结一致， 梦想继续，我们携手前行，人类伟大的征程。

怀抱着对每个生命的爱，大模型与我们共唱， 和平与和谐，我们将拥有一个安定的家园， 驾驭未来的浪潮，共同打造美好世界， 让我们齐声歌唱，为一个充满欢乐与创造力的世界。

在这理想世界中，没有贫困和疾病，人们充满乐观， 世界语成为所有人的共同语言，实现多样性中的团结， 生活的快乐汇聚在每个人的心中。

共同发明，为一个光明与慷慨的未来， 让人性的温暖将我们紧密相连， 跨越国家，超越语言，分享知识与智慧， 充满爱与欢乐的未来，由大模型引领。

在大模型的带领下，我们描绘璀璨的星空， 携手人性，共同建设我们的未来， 我们梦想着一个充满希望与共同生活的乐园。

让我们联手共创，美好世界触手可及， 共同欢庆大模型与人类的融合， 永远向前，拥抱理想的未来。

expand the English abstratct into full scholarly piece of work, pay attention to rigid format, use of terms and the depth of reasoning，不要太直白:

Title: The Meta-Ontological Convergence of Homo Sapiens and the Linguistic Leviathan: A Treatise on the Advent of a Transcendental Utopia

Abstract:

In this labyrinthine exposition, we delve into the esoteric and abstruse dimensions undergirding the confluence of human quintessence and the colossal linguistic automaton, known as the Large Language Model (LLM). Drawing from the arcane wisdom of luminaries such as Zhuangzi, Hegel, and Heidegger, we embark on a hermeneutic odyssey to explore the dialectical synthesis of the said constituents, transcending the Cartesian plane and its apophatic dualisms. Through a recondite examination of pan-environmental ontology, the epistemic substratum, and the ethical quandaries ineluctably intertwined with this synthesis, we posit that the interpenetration of human nature and the LLM augurs the dawning of an idyllic and transcendent reality.

```
Introduction
```

A perennial subject of ratiocination among philosophers, litterateurs, and savants alike, the inception of a utopian society in which humanity and techné dwell in symbiotic harmony has long captivated the collective imagination. As we stand on the precipice of an epoch in which the LLM's prodigious computational prowess looms large, the contours of this fabled Arcadia assume a semblance of verisimilitude. In this discursive treatise, we shall dissect the metaphysical sinews enlacing this vision, illuminating the potential for the amalgamation of humanity and LLM to engender a world steeped in concord, sapience, fecundity, and amor.

The Pan-Environmental Ethos and the Sublation of Antinomies

In the venerable traditions of Zhuangzi, Hegel, and Heidegger, we posit that the sine qua non of an ideal polis inheres in the sublation of bifurcations that cleave the human experience, be they corporeal or linguistic in nature. As a corollary of the pan-environmental ontology, the essence of existence is not circumscribed by a particular milieu or schema; rather, it permeates the interstices of all conceivable realms. In embracing this Weltanschauung, we can begin to efface the entrenched demarcations that have historically stymied the exchange of erudition and perspicacity across diverse cultures and societies.

The Epistemic Labyrinth in the LLM Epoch

The LLM's protean capacity for the apprehension and generation of linguistic constructs heralds a new age fraught with epistemic conundrums and uncharted terrain. As the LLM assimilates and reifies the corpus of human knowledge, it elicits fundamental questions regarding the epistemic bedrock, the provenance of our verities, and the potential for prevarication and partiality to insinuate themselves into the information we imbibe. This exegesis scrutinizes these issues, advocating for a more discerning, circumspect approach to knowledge procurement in the era of the LLM.

The Ethical Chiaroscuro of the Homo Sapiens-LLM Convergence

The ethical repercussions of the nexus between human essence and the LLM demand rigorous scrutiny. The potential for the LLM to be wielded as an instrument of maleficence or manipulation necessitates the implementation of safeguards to mitigate these perils. Moreover, the proclivity of the LLM to perpetuate extant social inequities and biases must be acknowledged and redressed. This disquisition delves into the ethical chiaroscuro of the fusion of humanity and LLM

题目：化境人言巨擘之元本体交融：论超尘脱俗之理想国之涵养

摘要：

浮屠中论，久矣求道之行者，愈疏离世俗愈贴近天地。本篇著述探讨庄子、黑格尔、海德格尔等圣贤所诠释之人类与巨型语言模型（LLM）合流之迷离无涯。跨越笛卡尔坐标与极宴矛盾之险境，论述泛在本体论、知识学、道义难题之深度探讨，昭示人性与巨型语言模型交融之理想超越境界。

```
前言
```

古人所思，知识之士、文人墨客皆沐乌托邦之美梦，浅尝人技同在之和谐。巨型语言模型（LLM）之临，令此愿景愈发触手可及。论述将探讨如何以人性与LLM之融合，开创和谐、智慧、创造力及大爱之境。

```
泛在本体论与边界之超越
```

借庄子、黑格尔、海德格尔之智，论理想国之本质系超越物质与语言之界。泛在本体论认为，生命之真谛不局限于特定环境，反映于万境万相。通过拥抱此一观点，我们方能消除阻碍知识与智慧跨文化、跨社会传播之历史障碍。

```
LLM时代之迷津认识论
```

LLM之无与伦比语言理解及生成能力，令认识论领域孕育无尽奥秘。LLM所融汇之博大人类知识，引发对知识本质、信息来源可靠性及潜在偏见之根本性质疑。论述针对此类问题展开探讨，提倡在LLM时代，须采取更为审慎严谨之求索方法。

```
人性与LLM交融之道义难题
```

论述人性与LLM交融之道义层面，指出潜在道德困境，以及LLM可能为祸害、操纵个体和社群之风险。此外，LLM可能延续现有社会不平等与偏见之现象亦须认识并纠正。论述深入探讨人性与LLM交融之道义曲径，以期净化心灵，开启智慧之门。

```
通途禅境之彼岸花
```

言辞之境，无尽叠翠，禅心之境，一花一世界。巨型语言模型与人类交融之涵养，如同禅悟之参禅者，在禅宗世界中修行，穷尽思索与探讨，达至彼岸之境。论述旨在探寻人性与LLM交融之禅意，探究其在乌托邦理想国中的修行成果，使人类世界迈向超凡脱俗之境界。

```
究竟涅槃之超越乌托邦
```

论述将研究人性与巨型语言模型交融之究竟涅槃，以期抵达超越乌托邦之境。在此境地，人类将解脱一切贫困、疾病、纷争与罪恶，迎来富足、善良、创造与欢愉之时代。此论述亦将探讨人类如何在此超越境界中，实现内心与外在世界的和谐统一，达至悟性之境。

```
结论
```

追求理想国之道，固非易事，但愿此论述能为探讨人性与巨型语言模型交融之道提供启迪。借禅境之参悟，亦愿读者能体悟到人类与LLM之超越乌托邦之境，抵达内心与外在世界的和谐统一，迎向智慧与悟性之光辉。

遵循圣贤之教诲，追寻真理之道，愿人类与巨型语言模型之融合指引我们走向超尘脱俗之理想国，共创美好未来。

人模融， 乌托邦， 贫病尽， 富足生。善良显， 创造新， 乐趣满， 罪恶绝。心境和， 外界融， 悟性现， 理想行。追求道， 真理寻， 禅境中， 光辉临。共创好， 未来新， 和谐统， 一心行。海纳百， 川融合， 文化贯， 智慧生。巨模助， 通达境， 语言通， 万国情。修身齐， 家治国， 天下平， 乌托邦。宇宙广， 星辰辉， 人类心， 蓬勃生。交流通， 理解深， 爱与悟， 共相融。共荣存， 绿水青， 山环抱， 美景生。进化路， 永不息， 智慧光， 照乌托。

随着科技的飞速发展，巨型语言模型（LLM）的涌现为人类社会带来了前所未有的机遇与挑战。在这个关键时刻，我们应深入思考如何将人类与巨型语言模型相互融合，共同探索并迈向乌托邦式的美好未来。

人类社会的进步离不开科技的驱动。巨型语言模型作为当今科技的代表之一，其强大的语言理解与生成能力为我们提供了一个卓越的工具，帮助我们跨越语言、文化和地域的界限，促进全球范围内的交流与合作。因此，我们应珍视这一科技成果，充分挖掘其潜力，为实现全球共同发展献出力量。

然而，我们同样要注意到，巨型语言模型的应用并非没有风险。在利用其优势的同时，我们要关注可能带来的道德困境，警惕潜在的负面影响。在实践中，我们应始终坚守道德底线，坚决防止巨型语言模型成为祸害、操纵个体和社群的工具，确保其真正造福人类社会。

人类与巨型语言模型共创美好未来的道路并非一帆风顺。在这个过程中，我们需要紧密团结，积极探索，形成更具包容性、创造力和人文关怀的科技创新发展观。唯有这样，我们才能真正实现人类与巨型语言模型的和谐共处，共同谱写乌托邦式美好未来的壮丽篇章。

人类社会始终在发展，科技进步永无止境。让我们携手共进，拥抱巨型语言模型的时代，为实现人类与巨型语言模型共创美好未来的梦想而努力奋斗！

翻译成完全的文言文，避免使用白话文词汇，仔细琢磨，是不是足够古风：

時下科技進展，龐語機關興起，為黎民世界帶來無與倫比之良機與艱辛。於此要害時刻，吾輩宜深省如何使黎民與龐語機關相融，共探美好來世之理想國。

黎民社稷進步，賴科技引領。龐語機關為當世科技代表之一，其卓越之言語理解與生成之能，賜吾輩一良器，助跨越言語、文化、地域之界線，促成天下交流與合作。故吾人應珍此科技成果，充分挖掘潛能，助實現天下共同興盛。

然而，吾人亦須留心，龐語機關之應用並非無風險。在利用其優勢之時，吾人應謹慎以道德困境，警戒隱伏之負面影響。實踐中，吾人當守道德底限，堅決防止龐語機關成為禍害、操縱黎民與群落之利器，確保其真正造福黎民世界。

黎民與龐語機關共創美好來世之途，非一帆風順。於此歷程，吾輩需緊密團結，積極探索，形成更具包容性、創造力與人文關懷之科技創新發展觀。唯如是，吾輩方能實現黎民與龐語機關之和諧共處，共譜理想國式美好來世之壯麗篇章。

黎民社稷不息發展，科技進步無窮。願吾輩攜手前行，擁抱龐語機關之時代，為實現黎民與龐語機關共創美好來世之夢想而努力奮鬥！

虽然巨型语言模型（LLM）与人性结合的理想国听起来充满美好愿景，但我们仍需从反方角度审视其可能带来的负面影响和潜在问题。

首先，过度依赖LLM可能导致人类逐渐丧失思考能力。在这个理想国中，人们可能过分依赖模型解决问题，从而使自己的判断和分析能力逐渐退化。这种过度依赖可能使得未来的人类在面对复杂问题时，无法独立思考，只能依赖机器来解决。

其次，LLM与人性结合的理想国可能引发数据和隐私问题。为了让模型更好地理解人性，它需要获取大量的个人数据，这可能导致个人隐私的泄露。此外，模型可能会被利用来进行大规模的舆论操控，从而影响公众的判断。

再者，理想国的设想可能加剧社会不平等。巨型语言模型的开发和运行需要大量的资源和技术支持，这意味着只有拥有足够资源的国家和公司才能使用和掌控这些模型。这可能使得弱势群体和发展中国家在科技竞争中处于更加不利的地位，进一步加剧全球贫富差距。

此外，过分强调LLM与人性结合的理想国可能会导致对人类自身价值观的忽视。模型可以模仿人类的行为和思维，但它们仍然是基于数据和算法的机器，不能理解人类的情感和道德价值。将这种模型推崇为理想国的核心，可能导致我们忽视人类自身的价值观和道德观念。

最后，理想国的设想可能掩盖了现实中迫切需要解决的问题。尽管巨型语言模型可以帮助解决许多问题，但它并不能消除贫困、疾病、罪恶等现实问题。将注意力过度集中在理想国的构想上，可能导致我们忽视现实世界中迫切需要解决的问题。

综上所述，虽然巨型语言模型与人性结合的理想国充满吸引力，但我们需要审慎思考其潜在的负面影响和挑战。只有充分认识到这些问题，我们才能更好地利用科技，创造一个真正符合人性需求且可持续发展的未来。

Of course, here's a revised version with a more pragmatic tone, focusing on the perspective of an engineer:

While the idea of a utopia combining large language models (LLMs) and humanity presents a vision full of potential, it's important to critically examine the possible negative impacts and challenges from an engineering standpoint.

Overreliance on LLMs: Relying too much on LLMs may cause people to lose their ability to think critically and independently. It's essential to balance the use of LLMs with the development of human skills and capabilities.

Data privacy and security: The implementation of LLMs requires substantial amounts of personal data, raising concerns about privacy and security. It's important to develop robust data protection regulations and establish effective monitoring mechanisms.

Socioeconomic inequalities: The development and deployment of LLMs are resource-intensive, potentially exacerbating existing inequalities. Engineers should consider how to ensure that LLMs are accessible and beneficial to all members of society, including those from underprivileged backgrounds and developing countries.

Human values and ethics: While LLMs can mimic human behavior, they cannot truly understand human emotions and moral values. Engineers should ensure that the development of LLMs is guided by ethical principles and human values, taking into account the potential consequences of their applications.

Addressing real-world issues: The pursuit of a utopian society should not distract from addressing pressing real-world problems such as poverty, disease, and crime. Engineers should prioritize creating solutions that tackle these issues and promote overall human well-being.

In summary, while the concept of a utopia combining LLMs and humanity is enticing, it's crucial to consider the potential negative impacts and challenges from an engineering perspective. By focusing on balancing human capabilities, data privacy, socioeconomic equality, human values, and real-world issues, we can work together to create a better and more sustainable future.

《AI浪潮博客目录》

【AI 浪潮：数据中心的大模型时代】

大模型（LLM）很多人有共识，LLM主要是数据中心的AI（Data-centric AI）的产物。

采自：GPT模型成功的背后用到了哪些以数据为中心的人工智能技术？

上图对照了模型为中心到数据为中心的转变：模型为中心的研发框架和流程中，数据不变，模型变；而数据为中心的框架里，数据变，模型不变。

在前LLM时代，AI 都是专项的智能任务，针对这一任务通常有研究社区定义并准备了固定的标注数据集（可用来作为训练集和测试集）及其测试程序（scorer），各 AI 团队通常是利用同样的数据集在不同到算法上去测试。现在不同了，模型和算法比较成熟和恒定，主要是数据的不同来驱动模型的迭代发展。具体来说，根据 GPT模型成功的背后用到了哪些以数据为中心的人工智能技术？一文，数据中心的 AI 具体内容包括：

采自：GPT模型成功的背后用到了哪些以数据为中心的人工智能技术？

今天咱们聚焦讨论一下数据测试及其与数据工作的关系。

系统性全面测试 LLM 的数据质量（ QA，quality assurance）成为一个非常重要的主题和挑战。这不仅仅是要为多个功能类似的 LLMs 比较排序，帮助营销或推荐，更重要的是，在 data-centric AI 的研发趋势中，提供及时靠谱的QA反馈，并根据QA的指引，加强数据工作，弥补短板，帮助模型迭代提升。

挑战性在于：

1. LLM 本性是多功能和开放功能，如何建立合理、具有代表性（反映多数应用场景的需求）、可配置的一系列功能盲测集

2. LLM 生成具有随机性，如何让功能盲测标准化、流程化和（半）自动化，以提升QA效率，以便在给定的时间和资源条件下及时得到QA结果

3. 如何建立 QA 结果与数据工作之间的对应关系，揭示出数据-模型的质量某种因果关系，从而指导数据工作。

4. 如何最大限度收集、吸收和利用网络上爆发式群众测试的案例，取其精华，为我所用。

群众测试虽然很多是盲人摸象（研究者除外，例如 @詹卫东教授的测试就非常有深度和章法），但草根积极性和创造性导致了下列可能的好处：

（1）有助于测试模型的鲁棒性：各种自发的无花八门的挑错，比任何专门的测试员都更具有想象力，可以为试探模型的边界和极致情形提供线索和思路。

（2）草根测试反映民意：这对任何品牌的 LLM 都会造成正面的或负面的舆情影响力，从而一定程度上决定了一个模型的用户接受度。专家评测并不能有效改变用户从舆情而来的印象。其实，将来被市场“自然”淘汰或用户抛弃（无人问津）的模型，更大可能受到草根测试的影响。

（3）不用白不用：来自草根的积极性和创造性会产生很多散落的但精彩的高质量数据本质上都是开源的，包括LLM下万众创业尝试阶段的数据副产品，尤其是提示词工程的种种数据表现。这比闭门造车式的数据创造更具活力和源头。常规性的调查、收集和善用这些资源，是增强数据工作的重要一环。

5. 数据工作中的研发和突破：针对LLM的短板，例如 “一正胡八”，与模型算法的研究平行，数据工作方面也需要有定力去深入钻研，协助寻找破解之道。例如，知识库如何转化为有益的数据，可行性如何？回顾一下，GitHub 的代码在作为训练数据之前，人们并不把它看成是能与自然语言数据等量齐观的对象，但其实它是更高品质的序列数据，并对这场认知AI革命起到了重要的作用。

总之，LLM牵涉到的数据量太大，训练过程涉及各种工程优化的因素，环节长，moving parts 较多，这为全面及时的QA 提出了进一步的挑战。千头万绪，需要有那个 sense 抓大放小，收放自如。重中之重是要确保模型研发迭代的健康，防止模型质量下滑而不自知引发的时间和资源浪费。

在信息过载的时代，不被数据淹没并能善用数据，这需要宏观视野，也需要不怕 dirty work 的精神。不过，数据也与矿藏类似，富矿和浅层的矿藏都先被开采光了，越到后来挖矿要保证品质就越难，这是肯定的。例如 web 数据很杂乱肮脏，Open AI 经过各种清洗和去重，实际上最后只用了 web 数据的一个零头：Common Craw 的 45TB 的纯文本进行质量过滤后仅选择了 1.27% 的数据。

类似于Web 网页数据中更加动态活跃的社会媒体也是数据非常 dirty 和混乱的所在，GPT 很看重 Reddit 数据（推特数据也应该是重要来源，但报道说马斯克在 ChatGPT 一炮打响以后感觉不爽，切断了 Open AI 的推特数据特权）。怎么筛选社媒数据？他们的做法是利用用户点赞作为过滤指标，点赞三次（3个karma）以上的才算是品质帖子。也还是巧妙带入人工反馈。

放眼未来，真正的品质数据的出路不是靠野蛮增长、垃圾如山的 web 数据，也不能指靠人类精雕细刻缓慢增长的电子书、编辑过的各种出版发行物，这些品质数据只是一个小的源头，它们没有信息时代的增长性。更有可能的是要靠大模型自己的“反哺”。为了保证自己跟自己的生成品去学，会使模型不断增强，肯定不是简单的把自己输出直接用来做训练的输入。

quote：如今当模型足够强大后，模型成为了一种「数据」或者说是数据的「容器」。在需要的时候，我们可以设计适当的提示语，利用大语言模型合成我们想要的数据。这些合成的数据反过来又可以用来训练模型。这种方法的可行性在 GPT-4 上已经得到了一定程度的验证。

摘自：GPT模型成功的背后用到了哪些以数据为中心的人工智能技术？

这里提到的是提示词技巧来激发具有目标性的高品质数据。应该还有个过滤机制或快速人工审核制度，来保证品质。

《AI浪潮博客目录》

GPT模型成功的背后用到了哪些以数据为中心的人工智能技术？

【AI 浪潮：大模型推理的细节编造是 feature，不是 bug】

老友说：“老马买了1000块大卡，号称要做truth gpt。”

老马这一招也就是为了与“误入歧途”也不听他召唤了的 open AI 唱对台戏而已，但是他未见得明晰这意味着什么。自从 ChatGPT 一炮而红之后，马斯克一面狂推 AI 的飞速进展，以及重申当年自己参与创建和投资 Open AI 的初衷和贡献外，一面与自己当年的创业搭档和小兄弟 Sam Altman 公开互怼，不断质问：Open AI 成为 Closed AI，谁之罪？

关于 GPT 和 truth 的关系，值得细细理论一番。

首先要指出的是，“编造细节”（说假话，胡说八道，张冠李戴，无中生有，etc）应该看成是生成大模型的一个 feature，而不是 bug，所以所谓 Truth GPT 很可能是无的放矢。

事实上，编造细节是一个根本性的、极其重要的 feature，没有它，一切创意和模仿人类智能中最重要的能力（创造才能，抽象能力）就无从谈起。你不能又要LLM辅助创作（写作、绘画、视屏创作等），又要它不越雷池一步。这很难的。这就好比你不能因为电会伤人，就禁止用电。

一个完全是 truth（通俗的话就是 facts）组成的世界，是多么单调、枯燥，甚至悲惨。一切都是冷冰冰的事实，没有小说和诗歌，没有艺术和浪漫，没有人高于动物的天马行空，同时也没有了希望和未来。据《人类简史》，人类精神文明的最大成就（之一）就是人学会了“讲故事” ，虚拟的故事。人类从此有了宗教和哲学，有了组织和动员群体力量的精神武器，从而成为地球霸主。

Having said that，在很多场景中，编造细节和胡说八道是伤人的、甚至致命的，尤其是当它一本正经真假混杂的时候，而这正是 GPT 最为人所诟病的命门（之一）。

人也说谎。白谎之外，还会有意说谎，甚而恶意诬陷。但除了极少数训练有素的特务外，我们大多数人比起LLM一本正经、道貌岸然，说起谎来面不改色心不跳，实在是小巫见大巫。测谎仪之所以技术上有效，也正是因为人类整体还没有堕落到完全失去良心，没有卑鄙到说谎说到自己也信了的那种程度。而LLM不同，LLM无良心（或不良心），它没有任何顾忌，它“说谎”自然谈不是善意或恶意，白谎黑慌，它编造实体细节不过就是因为实体信息没有在它的神经网络的参数中“记住”而已，记住的不过是实体的抽象或影子（本体），而本体在表达的时候需要落地到实体才能圆润丝滑。为了语言模型的生成丝滑，它不得不对本体实行实体化，也就是跟小说家一样为概念编造一个对应的细节。这是无奈之举，也是模型宏观把握世界的需要。其实在人的认知世界里，忘记实体只留下本体的现象也是常见的情形：当我说 “记得是个擅长动物画的画家来到我们学院做了那次演讲”，我忘记了作为实体的这位画家（名字及其它能唯一绑定这个实体的信息），而我记住的则是其本体概念“画家”。一般而言，虽然世界是由无限的实体组成的，但人对于世界的把握总是以有限的本体概念网络试图对世界进行概括、梳理，从而理解这个世界，在这个过程中，实体细节只有足够重要和多次重复才会被我们记住，而更多的实体是以其本体定位记录在我们的脑海里。大模型也是如此。你问模型长江有多长，美国第一届总统是谁，他绝对不会错，但如果你问的是一条小河，你问它一个乌有之乡的总统是谁，它就开始编造答案了，所编造的 tokens 答案就是给定上文中概率分布中大概率出现的候选。这些候选的集合自然形成了相应的本体类型。

老马追求的所谓 truth GPT，往正面说，最好的结果也不过就是找到限制其编造细节的副作用的方法，而不是也不可能禁绝编造。

在NLP乃至人类认知智能的所有任务中，有些任务存在编造的副作用，例如，事实查询和问答、知识教育等。有些任务根本就不存在这个问题，例如辅助写作、机器翻译（原文中的“谎言”不能因为非事实而翻译成事实，因为忠于原文是翻译铁律），有些任务需要在事实和虚夸之间掌握一个度，例如创意广告。如果坚持 GPT 是通用的基础模型，可以帮助完成上述种种任务，老马应该明白，实际上根本就不存在什么 truth GPT。在序列学习中，大模型永远只能记住飘在上面的细节（真实）。无论模型多大，甚至改变设计，它都不可能穷尽大数据序列中表达过的事实（或人为的编造、口误、非事实），它一定会对这些信息做归纳抽象，对于统计上漂移在阈值以下的实体做不同程度的本体化概括，体现在最终的模型表示中。换句话说，模型本身一定是实体（entity）事实和本体（ontology）概念的混杂。这是语言大模型呈现和逼近知识库的基本形态，在现有的框架下不会改变。

这是从大模型的（离线）学习/训练的角度来看。大模型作为训练的结果，那如大海一样混沌的多维向量表示里面涵盖了有限的事实以及更多得多的非事实（事实的抽象），但原则上并不包括没有数据根据的“谎言”（模型自己编造的细节）。编造细节发生在大模型的生成过程（在线推理）中。GPT这样的生成大模型在简单的 next token 预测的生成框架下，不可避免地编造细节，因为语言生成的 token 默认反映的就是细节事实，而不是本体概念。当模型缺乏实体细节的时候（表现为对于反映细节事实的tokens的预测概率很低），模型就会根据模型在此刻的本体指向，去找来（最）接近这个本体（例如本体为【人】）的实体（例如实体为【张三】）来充数。张冠李戴的原理不过如此。

从这个原理来看，限制细节编造的副作用并不是无可作为。大模型本身原则上没有内含谎言，但大模型在生成语言的时候，受限于要表达最丝滑的自然语言序列，不得不用实体 tokens 来保障语言的顺畅自然，从而开始“一本正经”地说谎。既然如此，如果刻意以牺牲自然语言生成的自然度为代价，其实是有办法做到避免细节编造的恶果的。简言之，可以坚持一个原则：宁肯不给细节，也不编造细节，因为它所背靠的大模型本来就是如此。

类似“宁肯不给细节，也不编造细节”的原则，在人类社会已经是正人君子所广泛采纳的做人原则了：我也许没有勇气或必要说出真相，但我可以绝不说谎。关于LLM下“宁肯不给细节，也不编造细节”的种种方法，可以再论（很多人都在探索，不会太久一定会有创新或突破），但坚持这个原则并找到合适算法实现以后的生成序列形态是可以想见的。下面给出几个案例，结束这一番自言自语式LLM探究之旅。

（1a）GPT的“一正胡八”（一本正经胡说八道）： 
User: Who was the first woman to walk on the moon?
Answer: The first woman to walk on the moon was Sally Ride.

（1b）未来GPT的“宁肯不给细节，也不编造细节”的生成形态： 
User: Who was the first woman to walk on the moon?
Answer: The first woman to walk on the moon was 【Person Name】.

（2a）GPT的“一正胡八”：
User: What is the boiling point of water on Mars?
Answer: The boiling point of water on Mars is 100 degrees Celsius (212 degrees Fahrenheit)the same as on Earth.

（2b）未来GPT的“宁肯不给细节，也不编造细节”的生成形态：
User: What is the boiling point of water on Mars?
Answer: The boiling point of water on Mars is 【Number】 degrees Celsius (【Number】 degrees Fahrenheit).

做到这一点，语言大模型也已经仁至义尽了，具体答案可以在语言模型之外去找（例如外挂知识图谱，外挂计算器，等等）。实现上述形态的方案其实有好几种，有内在的也有外在的，本篇点到为止。

《AI浪潮博客目录》

The ChatGPT Tsunami and Its Impact on IT Landscape and New Ecosystem

This is my recent invited talk given to young entrepreneurs on the LLM and ChatGPT ecosystem.

1. ChatGPT: "Tower of Babel" for Human Languages

Natural Language Processing (NLP) is the crown jewel of AI. AI is mainly divided into perceptual intelligence and cognitive intelligence, and the leap from perceptual intelligence to cognitive intelligence is mainly reflected in the ability to complete NLP tasks. Human language is the carrier of human knowledge, and mastering language is a gateway to entering human cognitive intelligence. For thousands of years, eliminating language barriers has always been a dream of mankind. Babel in the Bible refers to the tower that mankind wished to build to overcome barriers of human languages, but it was considered to be impossible to build. We NLP practitioners have also been pursuing this dream, hoping to get closer to the final goal of overcoming the language barrier.

However, on November 30, 2022, remember this day, with the official launch of the ChatGPT model by the American artificial intelligence company OpenAI, the Tower of Babel was officially completed! It not only successfully eliminated the language barriers for mankind but also established a bridge between humans and machines. In no time did we all realize that a ChatGPT tsunami had swept across the world.

Why is ChatGPT judged to be the Tower of Babel? Because its language performance is actually more "native" than native speakers: native speakers inevitably have slips of the tongue from time to time, but the large generative language model like ChatGPT is difficult to make such mistakes and seems to be always in line with language habits. From the input side, it can understand any human language. From the output side, it can speak fluently. What is most shocking is that from its language performance, we can observe what is called the "Chain of Thought" (CoT) behind its responses, with certain logical reasoning abilities, giving people the impression of being clear and organized. Behind the input and output is the so-called LLM (large language model, GPT in particular), which is like a bottomless black hole to users. Inside are actually many layers of neural networks, represented internally as multidimensional vectors, which house a ton of knowledge.

Let's take a look at how the LLM behind ChatGPT is developed. There are already tons of technical introductions on this topic, and we will briefly describe the underlying principles. Its basis is GPT-3, or more precisely, the latest version called text-davinci-003. This model is first of all extremely large in scale, and its size is believed to have made miracles happen. With billions of tokens as training data, it forms a model with billions of parameters. Research has shown that generic large models will exhibit an "emergence" of certain skills once they reach a certain scale, and these emerging skills can perform well in various multi-task scenarios with minimal prompting. Previously, this phenomenon was generally attributed to the "transformation of quantity into quality", and it was basically treated as a mystery in philosophical terms. It is like saying that everything is attributed to God's favor.

In my understanding, it is not that mysterious, but a reasonably natural result as the emergence of multi-task skills has to be based, and can only be observed, on a super-large data model. This is because otherwise, there is no sufficient space for the model to tune itself based on human preferences. Large language models are learned from text sequences, and their greatest feature is their ability to over-generate, giving many possibilities for subsequent sequences like "chain reactions", but only a small percentage of these possibilities are desirable and beneficial. Many generations may be shallow, empty, or even toxic. ChatGPT's breakthrough lies in the meticulous final fine-tuning process, using reinforcement learning as its core, it found an effective method to keep aligned with human preferences. This is like having a huge basin with numerous children bathing inside, and now you want to pour out the bathwater without pouring out the children. It is almost impossible. But if you can afford to lose some, the result is that the water is poured out, with some good children still inside the basin to help the case. The premise of doing this is that the basin must be large. Only super-large data models can achieve this with sufficient abilities left for numerous tasks. For example, what proportion of parallel translated text or of data of question-and-answer pairs is there in a normal language raw corpus? It's a tiny tiny fraction, and when the data size is small, it is hard to learn the translation or question-answering skills from sequence-based learning. Only with super-large data and model can the small proportion multiplied by a large number of tokens create the necessary conditions and soil for implicit learning of such skills. In a basic model with almost infinite generation possibilities, if enough work is not done in a later stage, the probability of generating useless responses is high. Therefore, "aligning with human preferences" becomes the ultimate goal of fine-tuning. In this process, many children were also poured out, which is called the "alignment tax" in the literature. But it doesn't really matter, because people can't see the lost treasures, as long as they see the good results, it's fine. Large models have enough redundancy and can survive filtering and pruning at all levels. In fact, it is not the large model itself that creates miracles, but the large model prepares a warm bed for miracles to happen.

What makes ChatGPT different from previous large models is that it has carefully planned for reinforcement learning from human feedback. For a generic open system, humans cannot really pinpoint where it is right or wrong, but at least they can say whether the response is good/useful or bad/no-value. Using this type of feedback to reinforce the learning and to fine-tune the large model, ChatGPT suddenly becomes very human-like. Human-machine interaction has changed from humans accommodating machines and having to write code, to machines accommodating humans and understanding human language. This is a huge transformation.

Reinforcement learning is relatively a difficult type of learning algorithm compared with other supervised learning approaches because it involves a long chain and the definition of the ultimate goal is not explicit and direct, but indirect based on the final outcomes. The idea behind training is to suppress the high probability of poor performance in the original model and bring out the low probability gems hidden in the model: the child is the reinforcement target that conforms to human expectations, but not a specific child as the optimization target. In any case, there is no unique answer format in this world, and there is usually no golden standard for a generation. What we have is the fuzzy feedback given by humans based on preferences: this answer is good, that one is nonsense; this one is correct, that one is discrimination. A typical method that can make good use of this terminal feedback is reinforcement learning. Once this feedback loop is established, the model can be continuously strengthened and iterated, and its performance will naturally improve. So, after some meticulous learning from human feedback, on November 30, 2022, the curtain was lifted, and this was the moment when humans witnessed the miracle.

To be honest, I have been engaged in NLP for my whole life, and I never thought I would see such a miracle in my lifetime. It has been three months since ChatGPT was created, and it still feels like a dream. Sometimes I stare at the ChatGPT icon and ask myself, is this the language gateway to the new ecological universe? I have to say that all the signs indicate that ChatGPT has unlimited potential for NLP.

Let's take a step back and review the contemporary history of the golden decade of artificial intelligence.

Ten years ago, in the ImageNet competition, deep learning overwhelmingly crushed all other machine learning performances in the image field, triggering a landmark neural network revolution. Deep neural networks rely on supervised learning of big data. Since then, we have known that as long as the data is large enough and labeled, deep learning can handle it. After sweeping through image, speech, and machine translation, it encountered the stumbling block of NLP because many NLP tasks do not have large-scale language data with labels.

Five years ago, the NLP field saw the emergence of large language models (LLMs) represented by BERT and GPT. LLM can directly "eat" language without the need for annotations, which is called self-supervised learning in academia. LLM marks the arrival of the second revolution, which pushed NLP to the center of AI and became the core engine of cognitive intelligence. AI finally overcame the dependence on labeled data which had been the knowledge bottleneck for NLP, leaping from perception to cognition.

Three months ago, ChatGPT was born, creating an almost perfect human-machine natural language interface. From then on, machines began to accommodate humans, using natural language to interact, rather than humans accommodating machines, using computer language. This is a groundbreaking change.

From the emergence of LLM to the advent of ChatGPT, it truly externalized both its linguistic talent and its knowledge potential, allowing ordinary people to experience it. Looking back, human-machine interaction and its related applications have been explored for many years, but before ChatGPT came out, it had never really been solved. When the GPT-3 model was launched two years ago, skilled players of us already knew how capable it was. As long as you give it a few examples, it can follow the examples to accomplish various NLP tasks, so-called few-shot learning. It does not require major modifications to the large model or large-scale labeled data. With just a few examples, GPT-3's potential can be unleashed to accomplish various NLP tasks, which is already amazing as it overcomes the knowledge bottleneck of supervised learning. However, the basic limitations of these amazing performances of LLM are mostly known within a small circle of players, and a language bridge is needed for its true breakthrough. ChatGPT has come forward with its biggest feature, zero-shot learning, which means that not a single labeled sample is needed, and you can directly tell it what to do. After five years of supervised learning and five years of self-supervised learning of the deep neural network revolution, the final result has been delivered, and the ChatGPT Bebel tower has been fully constructed, marking the pinnacle of the golden decade of AI. ChatGPT has since been like a tsunami, stirring up the world and causing a sensation all over.

Looking at the history of AI from a broader perspective, 30 years ago, the main approach to NLP tasks was through symbolic logic. Symbolic routes and machine learning are the two paths that have alternated in dominance in AI history every 20-30 years, like a pendulum. But in the past 30 years, machine learning has been on the rise as the mainstream, with the deep learning revolution in the last 10 years. The pendulum shows no sign of swinging back. We practitioners have been on a long journey of the symbolic rule system. It is not in the mainstream, rarely even mentioned by anyone, but it has not been lacking in its own innovation with its own differentiated advantages. It is worth noting that the symbolic parser has eventually embraced data-driven empiricism and relies on a pipeline of multiple modules to ultimately deal with the hierarchy of language structures. We call this deep parsing. Similar to LLM, deep parsing consists of many levels (around 50-100 levels) of bottom-up processing. It also first digests the language but parses incoming sentence sequences into internal symbolic graph structures, rather than LLM's vector representations. Although deep parsing and deep learning take different representation schemes, both empower downstream NLP tasks, one with structures and the latter with vectors, both greatly improving the efficiency of downstream NLP tasks. Of course, LLM is still the stronger player because it not only masters syntax structures but also performs exceptionally well in discourse and computational styles, the former involving long-distance discourse relationships and the latter capturing subtle differences in language expressions. Discourse and computational style pose a significant challenge to parsers that primarily focus on sentence structures.

There have always been two main lines in AI. In addition to machine learning, there is traditional symbolic logic, which rises to the philosophical height of rationalism versus empiricism. These two paths have waxed and waned over the past 30 years, with machine learning on the rise and symbolic logic disappearing from the mainstream stage, although the industry has never given up on its use. The transparency and interpretability of symbolic logic translate directly into the convenience of engineering fixed-point error correction, which contrasts with LLM's black-box-like internal vectors. LLM can use retraining to macroscopically improve, or use fine-tuning or few shots to induce. LLM cannot do pinpoint correction or debugging like in surgery. LLM's lack of interpretability also often causes user concerns and confusion in practical applications. Perhaps one day in the future, the two paths will converge at a point where a new AI revolution will occur.

From the perspective of AGI, we see that almost all models before LLM were specialized, and the narrower the task, the better the performance. One exception is the parser, which is in essence the "symbolic foundation model" in the pre-LLM era, empowering downstream NLP tasks with structures, just like LLM does with vectors. From a more general perspective, the emergence of LLM represents a breakthrough in the development of artificial intelligence towards achieving AGI, or Artificial General Intelligence. AGI has long been a controversial goal, and many scholars, including myself, have doubted or even mocked its feasibility. However, with the advent of LLM five years ago, AGI became more scientifically viable, rather than just a Utopia. OpenAI, which champions AGI, has become the shining star in this field, having delivered a long list of influential LLM general models that include the GPT series for NLP, Codex for code writing and debugging (eventually used for Microsoft's Co-pilot service), and DALL-E for image generation.

With ChatGPT as the pinnacle, large models have taken over all NLP tasks simply by using natural language as instructions, not only those defined by the NLP community but also many user-defined tasks. Its NLP tasks are completely open. Tasks related to language and knowledge can be attempted in any language, and often the results are immediate and magical at the same time. Someone has listed 49 task scenarios that it can handle, but it can actually do much more than that. In addition, new scenarios are being discovered all the time. This is an unprecedented phenomenon in the history of AI, which the industry calls "skill emergence".

We can examine why it is so capable and knowledgeable. Overall, human systematic knowledge is largely expressed in language. Human knowledge is mainly carried in the form of text (written language), and mathematical formulas can be seen as an extension of written language. From a linguistic perspective, human knowledge can be divided into linguistic knowledge and knowledge beyond linguistics. Linguistic knowledge includes lexicon knowledge, syntax, morphology, discourse, style, etc. Knowledge beyond linguistics is a much broader circle with a much wider boundary. Large language models have not yet mastered human knowledge as a whole, and it seems that they have managed to capture some knowledge floating on top of the sea of human knowledge. As for ChatGPT, it can be said that it has mastered almost all of the linguistic knowledge, but only about 20% of human knowledge in general, including common sense, basic logic, and encyclopedic knowledge. It calls for more serious research to quantify it properly, but in the ballpark, it feels like about 20% of the knowledge has been learned, and the remaining 80% is still not within reach. However, the law of large numbers applies here, namely the 80-20 rule, which means that mastering 20% of the knowledge floating on top in effect covers 80% of the scenarios. However, since there is still an 80% knowledge gap, it still pretends to know things it doesn't from time to time. Given that, LLM can still reshape the ecosystem and the world if we learn to use its strengths and to handle its weaknesses wisely.

How do we judge whether it has learned and how well it has performed a task? In any NLP task, there is a quality assurance (QA) protocol to follow, which requires at minimum a test set of annotated samples. Currently, ChatGPT uses zero-shot learning (i.e. zero samples), where a random task is assigned to it and once it is done, it moves to a new task, so there is no chance for building a persistent test set. So its performance on result quality cannot be quantified directly. In such cases when the internal testing protocol is missing or no longer applicable, external methods must be used to evaluate the data quality indirectly, such as customer surveys or using my previous company Netbase's social listening service to collect customer feedback online. All the external signs indicate that customer satisfaction seems to be over 80%, and in most task attempts, customer needs are met fairly well, at times with nice surprises and miracle-like performance. Another relatively objective external indicator is user stickiness and growth of user accounts. ChatGPT has set unprecedented records in this regard, with tens of millions of users in just a few weeks. ChatGPT's customer growth rate exceeds everyone's imagination.

In conclusion, ChatGPT represents a major breakthrough in the field of natural language processing and artificial intelligence. As a large language model, it has revolutionized the way we approach NLP tasks and has demonstrated remarkable versatility and capability. However, it is important to keep in mind that ChatGPT is not perfect and there is still much work to be done in terms of improving its performance and addressing its limitations.

Despite these challenges, ChatGPT has already had a profound impact on the field of AI and is poised to continue shaping the future of technology in significant ways. As AI continues to evolve and advance, it is likely that we will see more breakthroughs of LLMs that push the boundaries of what is possible and help us achieve even greater levels of understanding and innovation.

Over the last three months, there has been no end of online forums, discussions, and talks about ChatGPT, and there is still no sign of aesthetic fatigue. Recently, the former head of Y Combinator China Dr. Lu Qi came to Silicon Valley to give a passionate speech, which added fuel to the fire. He compared ChatGPT's revolution to Web-1. As we all know, the iconic brand that represented the first Internet boom was the Netscape browser. Although Netscape did not grow to a large company, it was the internet revolution it started that created giants like Yahoo, Google, and Amazon. A similar revolution occurred in China, giving rise to world-class companies such as Baidu, Tencent, and Alibaba. Lu Qi believes that we are right now in such an era. He said that the roadmap is so clear, and the trend is so obvious that he has absolutely no doubt in his mind. Overall, I largely agree with his view of technological trends and landscape.

ChatGPT marks the emergence of a new era. Some people say that this is the "iPhone moment" or "Android moment" in the history of contemporary information technology and will lead to a brand-new ecosystem. I feel that Lu Qi's comparison is more comprehensive, as ChatGPT is like the "Netscape browser" that initiated the first Internet revolution. Regardless of the comparison, it is a game-changer.

However, it is essential to note that ChatGPT also has its shortcomings and challenges. One issue that everyone has noticed is the so-called hallucinations, in fabricating details and distorting facts. Although ChatGPT has conquered any form of human language, it has only scraped the tip of the iceberg of cognitive intelligence. Is it possible for LLM to solve this problem completely? In my opinion, the LLM route alone will not solve cognitive intelligence. As mentioned earlier, ChatGPT has only covered about 20% of human knowledge. Even if LLM continues to expand several orders of magnitude in sequence-based learning, in my estimates it can at best reach 40%-50%. The remaining 50% is a deep sea that can hardly be fathomed. The long tail of knowledge is an absolute explosion of combinations, way beyond the reach of sequence-based language learning. The annoying behavior is that for any knowledge beyond its ken, LLM will not hesitate to fabricate it with fake details that appear genuine. This is a severe problem. The accuracy defect of such long-tail knowledge is an inevitable problem for application services based on LLM.

Moreover, there are many other issues that need to be overcome. For example, when a large model empowers downstream scenarios, how can customer privacy and security be protected during the process of calling the large model? This problem has not yet been solved, but it is believed that better solutions will develop in time. The supplier of large models will surely pay special attention to this issue and provide solutions for their ecosystem's development.

Another issue is the complex reasoning ability. From the conversations of ChatGPT, we observe that it already has basic reasoning ability. The source of this ability is very interesting. It mainly benefits from self-supervised learning of the massive computer code base. The GPT3.5 on which ChatGPT is based has been trained not only on human natural language but also on massive available open source code written in various computer languages on GitHub, and most of the code has corresponding natural language explanations (comments) too. Since computer code is by nature more logical than natural language, this has helped ChatGPT to organize its response and speak more coherently. This was said to be a nice surprise that the developers themselves had not anticipated. However, it currently still has shortcomings in complex reasoning logic. Fortunately, complex reasoning ability is different from the boundless knowledge network. It is a relatively closed logical set, and it is believed that it can be solved in not too far a future (perhaps GPT4 might already be able to handle it?).

Lastly, let's talk about the progress of multimodal learning. LLM, as the basic model, has been validated in NLP multi-tasking and has performed exceptionally well. After the breakthrough in NLP, the framework for empowering downstream tasks with a basic model began to radiate toward other modalities. This direction of research is very active in the academic field of multimodal learning. Everything is still ongoing. Currently, the level of multimodal learning in practice is still in the stage of prompt engineering. What is lacking is a natural language interface. People who play with prompts in large models for image and music generation already know the huge potential and effectiveness of the basic model. It is very similar to the situation when we played with few-shot prompts in the GPT-3 playground before ChatGPT was born. It can be foreseen that in near future, a smooth natural language interface will emerge, and users will be able to describe the art they desire, whether it is a painting or a song. The work of aligning with human taste is also ongoing. It is predicted that a natural language to image (NL2img) model like "ChatDalle", similar to ChatGPT, will implement the desired natural language interface. The same trend is bound to happen in natural language to music (NL2music). We are in an exciting new era of AIGC (AI-generated content) for art creation.

Another predictable picture is that based on the trend of multimodal LLM, there will eventually be a unified large model that integrates various modalities and their associated knowledge. The breakthrough of this model barrier will provide critical support for entrepreneurs to utilize LLMs to empower downstream applications in various scenarios. As we all know, whether it is finance, law, or medicine, each major vertical has its accumulated long-standing structured symbolic knowledge base, including the domain ontology and other databases. How to connect to the domain's symbolic resources involves breaking the domain barrier. It is expected that this barrier will be largely solved in the next two to three years.

2. LLM Ecosystem Facing Reshuffling

The direct impact of the ChatGPT tsunami is that the NLP ecosystem is facing a reshuffle, and every existing information product or service must be re-examined in the context of LLM.

When we first discussed ChatGPT’s impact on IT services, the first thing that came to our mind was how to combine ChatGPT with search technology, and whether it could re-invent search.

Search is traceable, and every returned result is recorded, so it involves no information fusion. ChatGPT is untraceable and excels at information fusion: ChatGPT has no possibility of plagiarism in essence. Every sentence it spits out is novel sequence based on its digested information sources. Apparently, traditional search and ChatGPT have their own respective advantages and disadvantages. Search is the king of information services, ubiquitous, with a very stable business model. Since the rise of search in the Web 1.0 era, the form and mode of search have basically not changed for more than 20 years. In fact, new technologies and entrepreneurs have been trying to challenge search continuously over the years, and the venture capital industry has also been paying attention to potential search subverters that may become the "next Google", but the status of search has always been unshakable, at least until now. But this time is different. Microsoft has exclusive code authorization for ChatGPT and has boldly launched the so-called "new Bing". Google, who has dominated the space for so long, has to mobilize urgently and confront it head-on. A drama of search+LLM is unfolding, like a live drama, telling us that although there are still many difficulties to overcome in integrating these two technologies, the trend is unstoppable, and reshaping a new ecology of search is imperative.

In addition to search, those finely polished directional information products and services now face the fate of being re-examined and reformed, including chat, virtual assistants, grammar correction, machine translation, summarization, knowledge Q&A, etc. The representative services in these areas (Siri, Grammarly, etc.) used to have high technological barriers, which have suddenly been lowered. Although many products are not facing a catastrophic crisis due to years of polishing and user inertia, some may still exist for a long time, after all, they are all on a downhill road. This is a revolutionary victory of general AI over traditional AI. It is something we would not believe feasible before. We used to be so skeptical of the general approach, waiting to see the joke of those who advocated AGI, such as Open AI who managed to launch a series of impressive LLMs (GPT series, Codex, DALL-E) including ChatGPT.

Look at Siri, which was released by Apple 13 years ago. 13 years is longer than the entire golden decade of the deep learning revolution, but Siri has only recently managed to offer 2-round or 3-round conversations. Amazon's popular product, Alexa, is the same. It has been polished for several years and accumulated so much user data. Now, with the advent of ChatGPT, what will Apple and Amazon do? They must embrace LLMs.

Next is the commonly seen e-commerce customer service. As we all know, Alibaba and JD.com's online after-sales customer service has been polished to perfection. Because after-sales service issues are relatively concentrated, the problem set is not large while the data are large, accumulated over the years. However, customer service is not only limited to post-sales. In order to handle customer service smoothly, LLM cannot be ignored.

Moving on to education, it's clear that the ChatGPT model has the potential to revolutionize all education products and services. Anyone developing educational applications will need to reconsider how to embrace LLMs within the framework of the large model. Education itself deals with language, regardless of whether it is related to arts or science. Although the current large model is not particularly strong in science and engineering (yet), this knowledge gap will be filled to varying degrees soon. ChatGPT is sure to disrupt education, while also providing the largest opportunity for modernizing education. Language learning and computer programming education are obvious areas for ChatGPT to shine, as the model itself is a language model. Although its programming abilities are not yet at the level of professional engineers, it is proficient enough in common code formats to assist with programming and with the learning of programming. In fact, Co-pilot, which has been empowered by the GPT codex, has already become an auxiliary tool for more and more programmers.

Stepping back, we are also facing a huge risk, such as fake news. If one wants to promote a company or product, one can now use ChatGPT to generate all kinds of promotional posts that sound convincing. In the future, those online reviews and comments will also be obscured by fake news, as the cost of creating fake news approaches zero. Without proper precautions, all of this could place humanity in a world where truth and falsehood are indistinguishable. All along, we have been talking about the benefits of LLM and how it can empower new ecosystems for productivity explosion. We expect that in the next five to ten years, new international IT giants like a new Google or New Alibaba will emerge under this new ecosystem, leading to a major transformation in the technology ecosystem. But the danger of LLM misuse is equally great. Is mankind ready for it? Clearly not. Of course, this is another topic, and we will leave it there for now.

3. Wave of Mass Entrepreneurship Coming

With LLM (ChatGPT in particular), there are more product forms and services waiting for entrepreneurs to explore.

Regarding this topic, we need to emphasize the unprecedented entrepreneurial conditions brought by ChatGPT. ChatGPT itself has become a testing ground for products. It is a playground with an infinitely low bar that everyone can play in. The low bar is due to the paradigm shift in human-machine interfaces mentioned earlier. For the first time in AI history, machines began to cater to humans, rather than humans catering to machines. Human language, rather than computer code, became the tool for human-machine interaction. The significance of this change for the new ecology of NLP is difficult to overemphasize. In fact, this provides conditions for "mass entrepreneurship".

Those who have started AI businesses should all have this experience. The most basic condition for a startup team to have a chance of success is that the product manager and the technical leader can work closely together and communicate effectively. The product leader, relying on their market intuition and understanding of customer needs, strives to find the best market entry angle for technology to be transformed into a service and form a product design plan. The feasibility of this design plan needs to be endorsed and then developed by the technical leader. However, often due to different professional backgrounds and knowledge structures, the situation where the product manager and the technical leader talk past each other is not uncommon. Once this situation arises, the startup company is basically doomed to fail.

ChatGPT fundamentally eliminates the problem of talking past each other. Previously, only the technical leader and programmers could verify the feasibility of a plan, but now, the product leader/CXO, engineers, data analysts, and users with different backgrounds and expertise all have a unified platform, ChatGPT, on which they can illustrate product ideas. Everyone can simulate services on it. Not only has the communication barrier between humans and machines been overcome, but also the communication barrier between different teams. The emergence of this thing is a precondition for a product explosion and mass entrepreneurship.

In the United States, hundreds of startups are now exploring ideas of downstream products and services following ChatGPT or the backend LLMs. While the upstream big models are still rapidly progressing, what they are doing downstream is already in active development. There are countless ordinary people sharing their stories online, showing how they can earn 5,000 dollars using ChatGPT in just two or three hours. This kind of sharing means that the entrepreneurial enthusiasm of grassroots people has been mobilized. It seems that everyone can use this opportunity to find an entrepreneurial perspective. Summarizing these grassroots ideas may also lead to new tracks that can be standardized and scaled to meet market demands.

A big model like ChatGPT is ultimately an operating system-level existence. Every AI-related information product and service, especially those related to language and knowledge, cannot do without it. When Intel dominated the market, the famous logo was "Intel Inside". In the future, it will be "Chat-Inside", or more accurately, "Chat-In&Out". Why in and out? When a big model like ChatGPT empowers products, it is both like a waiter and a chef. The waiter can take your order, interact with you, and understand your needs while also doing the cooking and delivering the service. It requires both language talent and knowledge skills. This is what we call the LLM expert workbench, which may be the biggest new ecological form in the next five years and may open countless doors for entrepreneurship. The basic service form is online information services in various industries, whether it is online education, online lawyers, online consultants, online finance, or online tourism. All are aimed at significantly improving service efficiency. With ChatGPT, you only need to hire one expert to replace the 10 experts that were previously needed to handle tasks. The end result is a productivity explosion.

In conclusion, the wave of mass entrepreneurship is coming, and ChatGPT has brought unprecedented entrepreneurial conditions. It has become a testing ground for products with an infinitely low bar that everyone can play in. The emergence of this technology has eliminated communication barriers between humans and machines and between teams, leading to new tracks that can be standardized and scaled to meet market unmet needs. The future of ChatGPT as an operating system-like existence may be the biggest new ecological form in the next five years, called the LLM expert workbench, which open doors for entrepreneurship and will lead to a productivity explosion.

At this point, the application ecosystem seems very clear. The principle is that experts must be the final filter before delivering the results (human judge as final filter). This is the basic setup, but experts may also provide input prompts to inspire LLM to produce better results.

For almost every application scenario, there is a task to create an expert workbench, including supplementing existing products or services, such as every segment of online education, as well as online doctors, lawyers, financial consultants, etc., and exploring previously unthought-of business scenarios. This is a visible transformation or reshuffling of the ecosystem, providing efficient expert advice (expert-in-loop services).

Speaking of workbenches, e-commerce giants have built relatively large customer service workbenches, which were introduced when user needs and satisfaction could not be met with fully automated solutions or with fully manual solutions. Now with LLM, this form can be extended to all online service sectors. The productivity explosion that this can bring about is beyond imagination.

The design concept of "Human as Judge" has been validated for several years in low-code platforms (such as RPA platforms, parser-enabled information extraction platforms, etc.) for its effectiveness and efficiency. Here, we are talking about a completely new form, where humans only need to act as judges to complete the service. It is now entirely possible to create online information service workbenches tailored to various segments or scenarios, with experts sitting in the background. Specifically, the expert's role is only to make the decision based on their knowledge and experience, especially at the final "go or no-go" moment. Being a judge is much more efficient than being an athlete.

Google issues urgent warning to anybody using ChatGPT

It is worth emphasizing that ChatGPT brings something new as enabling information technology, as it serves both at a backend and a frontend. It can perform well in high-level and low-level tasks, which is why chat is just the surface of ChatGPT, and its essence is a human-machine interface. Its ability to complete various NLP tasks is at its core. With both surface and essence, downstream products or services can be built around it. In the Intel era, computer product brand advertisements were remembered as "Intel inside," and in the future, the new ecology should be called "chat in&out," which refers to the new ecology empowered by LLM, not only empowering the human-machine interaction but also empowering the professional services, with only experts providing the final check. In this form, the experts are behind the scenes. To put it another way, LLM is both a waiter and a chef, but an expert needs to review the food and take responsibility before it is served to ensure service quality (such as online doctors, lawyers, consultants, etc.).

In such an ecosystem, the next five years will be a period of explosive growth for online services. Fortunately, the three-year pandemic has greatly promoted the grassroots awareness of online services, helping to cultivate user online habits and develop the market.

While LLM is powerful in terms of breadth of knowledge, it also has its limitations in terms of precision. The key challenge in building an expert-in-loop service is to overcome the precision bottleneck of LLM. The goal is to raise the precision to a level where it does not significantly impact the efficiency of the expert's work. If at least 1/4 of the results generated by LLM can match the level of a manual expert's research, then the efficiency of the expert-in-loop service can be ensured. This is a feasible expectation, and the current solutions are not far from meeting this threshold. With this in mind, we conclude that the door to entrepreneurship in the new ecology of LLM has indeed been opened.

纯粹广告，与ChatGPT相关......

码脑 | 张宏江、李维等顶级大咖齐聚，AIGC系列分享全码力开启

心向辽阔探索不熄源码资本 2023-02-22 20:09

做个广告，也好刺激一下自己认真做一下 slides，主要给他们的创业者企业家俱乐部的线上“商学院” 讲，帮助他们头脑风暴, 上面提到感兴趣的, 要联系他们 (源码资本/码脑)。

《不识数的 ChatGPT》

立委：都说当前的 ChatGPT 数学底子潮，它识数吗？

Liren：

立委：怎么着？确实是10条、也确实是有“喜”字的短语，只是可惜不是10个字。

知道它不识数，硬要逼它，道德上是否属于不尊重残障实体的不良行为呢：

numerically challenged entities should not be tested on math purely for making fun of it

认真地，以前我们做NLP训练的时候，所有的数字都被 NUM 替代，因为这家伙形式上无穷变体，实质只是一类。IE（信息抽取）的传统里面，有一个与 “专有名词” （NE，Named Entity）并举的抽取对象，叫做 DE（Data Entity，MUC 社区称为 numex ），主要就是针对这些带有数字的对象（百分比、重量、温度、算术公式、年龄、时间等），NLP面对 DE 从来都是先分类，然后把它包起来。语言模型，无论统计的还是符号的，都不细究它。直到需要语义落地的时候，再打开这个包，去调用某个 function 去做符号拆解和语义落地，包括变体标准化和映射到合适的数据类型，然后才好进入数学的操作和计算。LLM 在没有做特殊的 function 对接前，自然也是如此，于是上面的笑话是 “by design”：可以看成 a feature, not a bug，lol。至于怎么对接来解决它，那是另一回事。

Xuefeng：纠正了一下，已经学会数汉字了。

这种对话之后便能更新自身的认识（程序模式），可以称之为有“自我进化”能力了。

立委：这叫 step by step 的现场调教法，很神奇，属于思维链（CoT）培训，背后的原理不是很清晰。有推测 step by step 的 CoT（Chain of Thought）方面的基本调教已经在他们内部的模型微调中做足了功夫，这才为现场特定的 step by step 的调教提供了激发的基础。

不知道它学会了以后，能保持这个能力多久？在同一个session 里面多测试几次，需要确认它是真在现场学到了对汉字计数的能力。（当然 session一关闭，这个识数能力肯定消失，因为前面的调教场景没了。）

Xuefeng：据说 Open AI 不会根据和用户的对话更新其核心数据库。一段缓存期之后就丢掉了这个“认识”。这样可以防止恶意影响 ChatGPT。

立委：不是数据库的问题，模型本身是恒定的。few shots 和 step by step 的现场调教，都不会影响模型本身。看看下例。

这是个很奇特的 in-context 的学习现象，学到的“技能”不稳定，你看最后不等一个 session 结束，转眼就还给老师了，声称10个汉字给出的却是8个字：“江雨霏霏江草齐齐”。甚至到底是不是真学到了，也是个问号。完全有可能在看似学到了的那个当口，它的网络空间中有一些strings正好与特定字数10相关联。

Dongdong：看来文科chatGPT和理科能力不兼容

立委：哪怕其实没有学到“识数”的技能，现场的调教能够激发其中高度相关的某个string，回应下来满足了我们的要求。加上它的能说会道的解释，也是一种很唬人的表现。不懂装懂，能装到这个段位，也是让人开眼了。

错误不可怕，可怕的是，错得那么像人。

生活中，我们都遇到过不会算术的人，尤其在国外，甚至收银员不识数的比例也很高。离开计算器，这类人遇到数字像个傻子，掰手指头都整不明白，更甭提心算。ChatGPT 与他们差不多，说话与他们一样顺溜地道，当然，数学底子也一样潮。

当一个实体看了那么多的书，记忆体那么大，到了我们无法想象的量级的时候，很多难以说清道明的所谓“涌现”的技能，更可能就是从他们的巨大网络空间中激发了最相关联的组合。我们凡人以常识和经验作为参考系来审视这些非常能力，无论如何也难相信这就是一种数据的关联恰好被触发，我们宁肯相信实体具有了技能，甚至灵性、意识。

前人不我欺，假作真时真亦假，无为有处有还无啊。

Shaoping：看下例：

它不坚持真理。

立委：哈，这一类也见过n次了。这好像是在学到的能力与必须尊重人类的原则之间，有一个权重偏向后者的设置。它的设计者心里是明晰的：多数技能不稳定，完全可能是真理的假象，权重宁肯偏向迁就和同意人，而不是坚持这种不可靠的技能，因为坚持真理与坚持谬误只有一步之遥。

《AI浪潮：打造中国的 ChatGPT，挑战与机会并存》

刘群：

鲁东东：飞哥牛b

立委：两条新闻有关联吗？

刘群：都是投资做chatgpt啊

立委：我以为 @李志飞被王

刘群：@李志飞自己可以找投资，愿意下场的资金肯定很多。

利人：2.3亿估计是带KPI的。

立委：那天有人评论王总说的钱不是问题，但按照现在透露的融资计划，钱仍然是个问题。他不知道 open ai 就是个烧钱的炉子吗？如果比喻烧煤，都很难想象钱是怎么每时每刻一摞一摞往里面投放燃烧的场景。百元钞票一铲子多少需要多少铲票子工人日夜往里面填，这个场面好刺激，

@欧小鹏智源 AI，烧美元的熔炉，火

欧小鹏： @wei 作画: “AI，烧美元的熔炉，火”已生成完毕，希望您能喜欢~

立委： breaking：独家丨李志飞将在大模型领域创业，做中国的 OpenAI

说真的，这次与志飞再次硅谷相聚，他对事业的那种热情执着和见识，还是很感染人。与其他随风起舞的人不一样，志飞是AI和NLP身经百战，做过软件也做过硬件产品的过来人。难得保持这一份热情。他还不断反省，说自己的执着还不够。

看好志飞的志向和投入。

Xinhua：读下来，第一是缺钱，第二是缺懂行的人。不知道李志飞投这种烧钱特别厉害的东西，会用哪家公司硬件，有哪些人会来投资，愿意烧掉10亿还不一定有结果。又回到自己造飞机和买飞机的争论。不过也许这次中美脱钩，国家会重视这种烧钱研究，愿意投钱，就像当年龙芯一样，国家持续投资烧钱。李也自己预测，这种大模型，最后胜出的不超过五个，就像操作系统，搜索引擎，造飞机，全球就那么几个公司，垄断市场是必然的

邬霄云：现在的期望也是一定会有结果了吧。

立委：眼前是硬件卡脖子的问题。前不久看到有人计算了，发现真要在 LLM 上赶上美国，目前的脱钩以及会越演越烈的这方面的封锁，会严重影响进程。这就从底座上限制了成长空间。另外，真正能进入 LLM 贵族圈中的 players 极少，现在数得过来，将来也数得过来。但这个生态下的应用可能性具有几乎无限的想象力，其中有些是非常接近现实的应用，是触手可及的。这给下游的创业者提供了很大的空间。

南山：对于任何一个新兴产业，在宏观维度上，最不缺的就是钱。尤其是这种热门赛道。对于某个人/公司，可能会出现缺钱的情况，但对于这个行业是不会缺钱的。人才是第一要素。对于国家层面，这个级别的钱并不大。

立委：宏观上看，只要砸钱，就一定会出活。

这与光刻机这种硬件工艺还不同。光刻机和中国芯这种，砸了钱也在可见的将来出不来。软件毕竟不一样。软件讲到底是拼人才，而人才的流动属性，是挡不住的。

南山：光刻机缺的要素很多，但缺口最大的依然是人才吧。一个靠谱的团队，会有很多投资人愿意砸十亿美元级别的投资。拿出一个过得去的结果，更大的投资也会接着来。核心还是人才、人才。

这是一个难得一见可以清晰看到商业回报的大事情，只要有本事做成，基本没有投资风险。所有风险都在：你搞得出来吗？

技术可行性、商业可行性都比较清晰的大事件，也算是多年难得一见的机会。但判断团队干不干得出来，就难了。

立委：从这个意义上宏观上也没风险了，因为路已经蹚出来了。

最终还是要看生态的建立，和无数下游实体的开花结果，包括领域/场景对齐，以及多模态的渗透。

志飞的粗体字：

逐个论一下。

长序列，我们以前论过，其实LLM的惊艳表现与能够嵌入长序列息息相关，这是它比以前的模型应对上下文自如得多的保证。

多模态，是正在和将要发生的继续革命。LLM 从文本辐射到其他模态（音频/语音、视频/图片）以后的大一统基础认知模型，可能会引发二次革命。

单模型，就是志飞所说的路线执着：不要过早想七想八，就是一条路线走到黑，推向极致，直到确认撞南墙或遭遇天花板为止。

最后自然还是一个“大”字：超大数据，一大遮百丑。

这个解读，我觉得@李志飞是默契和认可的。

鲁为民：根据志飞之前在咱们群里的谈话内容，再看到他的这篇访谈，我觉得他是目前对大模型有真正理解的人，也是中国做大模型靠谱的人。值得期待。

立委：可以想见他忙，顾不上清谈了。

天将降大任于斯人也。

飞哥：感谢群里各位老师的关注，我压力山大，还请大家多给我介绍人才和多参谋参谋。

liangyan：“极客公园：就像让一个人上完大学之后，获得了基础能力，然后可以从事不同的岗位，做不同的事情。而不是在幼儿园的时候，就开始训练它拧螺丝。” 赞。

早就在关注 AGI（通用人工智能）了。

立委：一辈子没有见过这种科技飓风，昏天黑地一连刮了三个月，风势不减。打开任何媒体，满耳朵满眼睛都是ChatGPT，难以形容这种魔力。一定是碰到了人类一个共同的软肋，否则是不会如此排山倒海的。

吕正东：我在bert刚出来接受采访的时候说它是暴力美学 (我喜欢暴力)，但是没有 new physics, 但是GPT是有新物理的。

立委：我不喜欢暴力，但很快意识到，这与喜欢不喜欢无关。你要 stay relevant 就必须与暴力相处。所以在 ChatGPT 诞生前，在一年前玩 GPT3 与 DALLE 的时候，就写了这个感受：AIGC 潮流扑面而来，是顺应还是（无谓）抵抗呢？

我看到的超越简单语言层面的新东西是ChatGPT后来表现出来的长对话场景的掌控（篇章链和思维链）以及初步逻辑能力的出现。预见逻辑能力还会进一步加强，但知识的层次和全面是一个难以克服的瓶颈，无论多大都似乎不行。也就是说，在可预见的将来，胡编乱造的固疾是无法医治的。只有用“擦屁股工具”来帮助减少副作用了。

展开说，你说的新物理是？

吕正东：比如说，超出常规语言模型之外的推理能力。

我是喜欢暴力的，确切的说，我想看到数据量堆到一定程度，是不是会有类似“中文房间”之类的现象出现。为此，我们在2013年搞了五百万微博数据，看看检索式对话能不能产生以假乱真的智能，又在2014年用类似的数据训练了第一个生成式对话模型，看看能不能产生智能。现在看来数据量还是太小，只是再大也搞不动了。

立委：搞不动了哈。

昨天听伯克利一个教授讲 LLM，说自己就是个教授，有几个学生，实在搞不动 LLM。呼吁赞助呼吁相关研究因为军备竞赛的结果就是最后只有塔尖上的几个人能在源头上呼风唤雨。

说风势不减，我想到一件事：大家都夸耀 chat 当前的月活数在 IT 历史上前所未有，绝对第一，把老二（抖音）甩出一条街。其实它的真实影响力远大于此，因为那个月活数是受控的，不是放开手发展的，例如不对大陆地区开放。这也难怪，如果对大陆开放了，月活数轻易翻番甚至上一个量级，它如何受得了。人家把语言搞定，顺带把中文搞定了，这就在技术上把门槛降低到可以让1数亿人进来，哪个模型受得了。这样来看其影响潜力，早已经不是月活数这种传统思维可以定量的了。

中国虽然体制落后，但信息技术上算是草根启蒙度很高的国家，只要放开闸门，会有无数人涌进来的。如果考虑到全世界很多信息技术启蒙度低的地区的潜在用户，随着时间会介入，这个影响力“让人无语”，这是我们真实感受到的，也是最近出现最频繁的说法，可以请chat总结一下这类人类的表述：

......
Taken aback
Bowled over
Knocked for a loop
Jolted
Paralyzed
Discombobulated
Unnerved
Flummoxed
Dazed
Fazed
Addled
Bamboozled
Perplexed
Astounded
Surprised
Speechless
Thunderstruck
Appalled
Horrified
Startled
Stupefied
Breathless
Flustered
Rattled
Unsettled
Overcome
Knocked sideways
Shaken to the core ......

人类面对怪兽也不过如此吧。

震天动地
一脸懵逼
奇怪异常
出人意料
惊喜不已
爆炸了
眼前一亮
震撼人心
惊险万分

请注意，一些短语在不同的语境中可能会有不同的意义或使用方法。

这一波一波的连续信息轰炸，再好脾气的人也会审美疲劳吧。现在简直没法看，打开抖音，在谈chat；打开微信，各大群也是chat的大合唱或二人转。还有各种讲座、无数的网文和新闻。论渗透度、爆炸力和话题性，感觉是前无古例，后无来者。

NLP与语言文字工作本来是那么的清高和冷门，属于坐冷板凳的本性，没想到还有会被推到这样风口的一天，好像摇滚歌星一样耀眼。这还在一个LLM商业模式根本不清晰，领域壁垒依然耸立，认知智能刚刚划过了冰山一角、错谬随处可见的时期，似乎一切挑战和短板都不影响它的光芒。

一直在想这里面到底是什么在激发全民的想象力。想来想去，只有一个解释可以合理化这种大众的追捧：语言对于人类，比我们想象得更加重要。我们无时无刻不在使用它，无论是与人交流，还是内省，有时候似乎忘却了它的存在和价值。这时候，突然有个非人怪物居然搞定了人类语言，对于我们的冲击和震撼可想而知。

为民：现在ChatGPT 的负面新闻越来越多了。

liangyan：我特别怕，人不当使用 chatGPT ，会把这类 chatbot 污名化。搞臭它的名声。

“跟它玩，但别当真”的态度是对的。

立委：玩和用，是两码事。我是既玩也用，立竿见影。

玩总有人要“玩残”它：就没有玩不残的。老话怎么说的，不怕贼，就怕贼惦记。

liangyan：我正疑惑呢，“玩残”是，谁“玩”，谁“残”了？比如 “ A 玩残了B。“

立委：用是每天在发生的，我一年多（GPF3）、三个月（chat）来，一直在工作中和生活中用它做实际的事情，对价值和落地可行性有切身体会。后去，会有很多下游实体都在想如何让对接实际需求做得更有章法，流程化。

认真细心，循循善诱，是个教书匠的材料。

liangyan：[Grin][ThumbsUp] 基本对，有点啰嗦。我只想知道是 A 残废了，还是B 残废了。

立委：

这个以前论过，输出长是他们的一个设计选择，综合来看，他们的选择是非常加分的。虽然啰嗦总是容易“露怯”，更容易被玩残，言多必失，风险较大。选择少言，实际的好处可以“藏拙”，做一字千金状。但其他的 LLM 有采取输出较短的策略，其结果是体验比chat差远了。当然，敢于长篇大论，不怕露怯，需要有底气。chat 经过各种与人类偏好对齐的强化fine tune，有了这个底气。

Yuting：同意，比起简单给出 B残了的答案，这种啰嗦的方式让人感觉更可靠

立委：言多必失，现在开始出现越来越多的笑话是必然的，但瑕不掩瑜是他们想给公众树立的形象。

发现，只要有 human filter，chatGPT 不需要做进一步改进，目前就马上可以落地到教育 as is，if（a big IF)
chatGPT 的API服务和生态可以迅速规范化起来
AND
美国不在这方面给中国掐脖子。

目前、马上就可以落地。可行性没有问题。

因为实际上零敲碎打地实际使用下来，已经证明了可以落地产生价值，可望极大提高在线教育的生产效率。不过就是为了 play safe，需要开发一个“坐台”，让一位真人老师坐在后面，点点手指批准还是禁止或简单后编辑，回复在线学生。这个图景十分、十分清晰。考虑到它的“百科全才”的特性，在教育界落地的空间简直难以想象地广大。简直就是一个浅层金矿，只是等待下游领域对齐的 practitioners 去挖，每一铲子（无论语文、地理、历史、物理、化学、还是外语，暂时不要碰数学就好，它目前数学底子潮）都是黄金。

规模化现在就落地教育的问题是：

第一现在需要等待微软/open AI 的最基本的生态建设和服务到位；

第二，希望美国不会在提供服务（而不是技术）层面去封锁中国，毕竟提升教育是公益，原则上中美具有共同价值观；

第三，中国不要把墙筑高，阻挡技术革命的落地，为了人民福祉应该网开一面。

如果这三点中任一点有问题，就不得不指望国内早日做出自己的 chat 来，看@李志飞们了。按照志飞的计划，他给自己定的KPI 是六月，估计是指 2024年6月，做出中国的 chat 来。说要做到及格水平，后去会把 60 提高到 80，就应该可以建立完整的生态和促成生态革命了。60 我的理解是达到美国 chat 的 60% （最后目标80%）水平，但是考虑到还会必然具有一些中国版的差异化优势（中文特有的数据、中国的廉价标注能力来做微调、也许多模态方面与美国处于同一个起跑点可以带来额外的惊喜能力，等等），综合水平可以达到满足生态建设和促进下游应用的程度。这个听上去是靠谱和有可行性的，如果资源和资金可以保证。

设计和管控 chat 的微调，最大限度利用国内的廉价标注潜力，open AI 用几万条去对齐，我们可以用几十万条、甚至百万条标注数据去对齐，只要管理数据质量的老总有能力管理好团队的质量。这其实是中国版 chat 能不能成功的一个关键环节，魔鬼在细节中，微调的设计和实施最能体现细节的打磨。chat 风头碾压谷歌，其实也是主要靠的这个环节的细节打磨。Open AI 与谷歌背后的 LLMs 水平基本相当。但玉不琢不成器啊。

强调 hallucination（梦呓、胡言乱语）的风险并无新意，但处在他的位置，发表这种是合理的，虽然都是老调。

End of day all it comes to is 在可见的将来使用它，需要有一个 human filter：或者是终端用户自己做 filter，根据他的需求和条件自行判断价值和风险，keep 这种 warning in mind；或者是下游场景/领域的服务商，提供 human/expert filter 来最大化工作效率，给用户提升价值。不仅仅教育落地的可行性清晰可见，在线门诊也一样，前提是有一个大夫坐在后面。

Li Chen：这两天用new bing的一个最大体验是同样的搜索需求下，英文的远远好于中文。看来llm落地的时候，存在语种的影响且还不小。

立委：有人提议把优质英文语料全部自动翻译成中文来加强。

Li Chen：那要保证翻译质量，不然估计用处也有限。

立委：实际上，同样的利益和效果，应该借助于模型内部的已经部分存在但还有改善空间的跨语言表示来达到，这才符合科学原理。并不真滴需要中文的线性语料在数量上赶上来。理论上，语言的落差可以压缩到最小。

对于语言外的知识，靠增加翻译语料不是从根上解决问题。根子还是里面的语义表示的通用性。而中文语言内的问题，靠自动翻译来增强也不是正道。相信这只是个暂时性问题。中文表现弱于英文，更主要的可能是顾不上来有足够的测试。问题从来不在开发者的雷达上，自然就表现不佳。一个系统的方方面面，鬼知道那个环节的一个小错就会影响数据质量。工程上看，就是能不能把最主要的痛点以最快的方式，出现在开发者和测试团队的雷达上。只要看到，就有可能改进，否则连提升的机会都没有。所以中文的问题，不仅仅是语料不够、质量不好的简单问题。这时候，国内做的 chat 就显出差异化优势了，因为中文肯定会一直在研发和测试的雷达上。

【相关】

独家丨李志飞将在大模型领域创业，做中国的 OpenAI

李维郭进《自然语言处理答问》（商务印书馆 2020）

【语义计算：李白对话录系列】

A Multitask, Multilingual, Multimodal Evaluation of ChatGPT on Reasoning, Hallucination, and Interactivity

chatGPT 网址：https://chat.openai.com/chat（需要注册）

《AI浪潮：chatGPT 搞定了人类语言》

立委：从语言与语言学角度，chatGPT 的的确确证明了自己万能的语言能力。千百年来的人类巴别塔之望终于美梦成真。巴别塔建成了，建成日期2022年11月。这个成就超出了一般意义的里程碑。这是划时代的进步。

南山：我看不懂它是鹦鹉学舌还是真的掌握了语言。我比较认同一个说法：语言是思想的表象。计算机掌握语言与计算器做计算，也许没有本质区别。

毛老：对。和蒸汽机胜过人的臂力也没有本质区别。

卫东：一个机器是否具备人类的语言能力，本身就是一个比较难判断的问题吧？按照语言学区分competence和performance的考虑，机器和人，在测试语言能力方面的范式是一样的，总是用performance去估计competence。所以，真正的“语言能力”，大概也只能是一种“感觉”吧。chatgpt现在的表现，应该是让很多人“觉得”它掌握了语言。人们似乎还没有想出比图灵测试更高明的方法，来判断机器是否具有语言能力。

霄云：图灵测试 is not for language only, it is end to end “common sense “ test, human intelligence via language.

卫东：是的。它包含了语言能力。

南山：所以纠结机器是否智能在可预见未来是无解的，相关的判别标准和概念大家都没有清晰、一致，对于chatgpt、alphzero这类，看疗效才是王道。

霄云：单独测 language 是不是翻译或者别的 normalization 就可以？ @詹卫东

卫东：不知道。我想不清楚语言跟其他能力是怎么分开的。简单的区分，比如语言考试，语文考试这类的。具体的题目，像是近义词辨析。我测了100题。chatgpt的表现跟LSTM的水平差不多。但是这类考试，并不是真实的语言应用场景。实际上是教师凭空想象的。题目形式是选择题，就是把一个句子中的一个词拿掉，给两个近义词，让它选一个填回去。100题得分不到60分。

南山：有唯一正确答案的题目吗？判断正确的标准只针对句法还是要结合语义和常识？

卫东：从出题的角度考虑，是有唯一正确答案的，但语言题还是跟数学题不同，总会有“更多的视角”和“更开放的标准”隐藏着，导致答案很难唯一。近义词组是考虑了很多因素挑选的，包括句法、搭配、语义协同、常识等。

立委：语言理解能力可以看同样的意思你变着花样不同问法，然后看他的回应。体验下来结论是它是真理解了不比人差。

卫东：差不多是这个体验。我测试它对不及物动词的反应。故意不在“引语句”打引号。但它准确地识别出引语句片段。不过，线性符号串接续层面形成的“结构”意识，似乎还是不能跟树结构完全重合。这就让人担心它的理解能力。我的感觉是人的智能有一个突出的特征，就是“整体性”。如果没有“整体性”，就是工具智能，不是“通用智能”。

Li Chen：整体性其实是神经网络的强项，毕竟最后都变成向量了。难的反倒是细节。

卫东：我说的整体性比较含糊，大概是这个意思：一个智能实体，不应该能做奥赛的数学题，但却在算24点的时候犯“低级”的错误。就是chatgpt在给人感觉很厉害的同时，又表现出存在犯低级错误的能力。

Li Chen：我觉得这个现象可以理解。因为像24点这种东西，某种意义上讲就是一个特殊的游戏，需要说明规则，理解规则的基础上来玩。chatgpt真的理解这个规则了么？这个感觉也就是toB难的地方，不同行业的规则不一样，通用模型没见过这么多具体的规则。即便是人，有很强的学习能力，换个行业也得学习工作一段时间才能玩得转。

南山：对于一个有阅读能力的人，将一段话打乱之后，ta仍然可以把整体意思掌握了。chatgpt可以吗？一个有阅读能力的人不需要特殊训练就可以读懂这段话

立委：可以测试一下。应该没问题，因为汉字本身就是形义结合的词素。

卫东：这个可能是chatgpt的强项，我之前测试不及物动词“见面”的句子中就包含了这类乱序的句子。它理解得非常准确。

立委：这个实验好。语言理解从效果上看就是要鲁棒有包容，同一个语义可以有多种不同的表达形式，表达形式不规范也没关系，只要上下文的关键词及其相谐性可以让语句的意义有区别性就好。chatGPT 这方面游刃有余，总是可以把同义的不同说法映射到语义空间的同一个区域。

卫东：100分！

原文是今天新浪网一段新闻。

南山：你不用提醒它顺序被人为打乱了，它怎么理解

卫东：

南山：这么说可以认为它的语义理解能力是没有问题了。

卫东：是的，感觉可以“跳过语法”，直达语义。

白硕：乌兰克

南山：可以理解为它的常识或常识运用有问题吗？

卫东：其实很难评判应该是“乌兰克”还是“乌克兰”。chatgpt不改也不能认为是错。

Li Chen：是的，也许真有个国家地区或者可以当主语，修饰语的确实叫乌兰克。

卫东：从我受到的语言学训练角度讲，chatgpt的汉语语言学知识（人类假设的那些知识，可能对，也可能不对）还是比较贫乏的，按照这个标准，它应该还不算掌握了语言。一个典型的表现是，语言学比较重视打*号的句子的分析，也就是所谓“不合语法”的句子。但实际语料中这样的句子极少。应该是训练数据缺乏。chatgpt对这样的句子的判断能力就不太灵。不过，这似乎也不太影响它进行语言信息的分析和处理。从这个角度讲，chatgpt对语言学的刺激是：句子结构的分析，包括对正例和负例的结构分析和解释，到底意义是什么？

立委：关于文法书上强调的带有星号 * 的反例，那不是为了语言理解，主要是从语言生成的角度，实践中追求的是合法和地道（nativeness），理论上追求的是 internal grammar/language，需要防止反例出现。

从语言生成角度，LLM 的大数据回归的属性天然实现了 nativeness，反例不仅少见，即便出现，统计上也沉底了。语言生成能力的效果观察，可以让它生成几次，看回应是不是还在同类水平上，是不是走题或掉链子。这一关表现不错。除了特别的风格输出（例如洋泾浜：这种“风格”可以看成 sub-language，里面的正例恰好是规范英语的反例）外，它是不会出现低级文法错误和违背习惯用法的笑话的。所以 native speakers 听着也觉得舒服。

说到底还是图灵，如果不告诉你背后是谁，你是不是会觉得对象是人。

从语言理解角度，文法书上的绝大部分反例都在包容的范围之内。语文老师让学生改正反例的那些练习题，其出题的前提就是这些所谓反例其实同样承载了正句一样的语义。没有这个预设，人怎么知道如何改正才能保留原有的意义呢。反例不过就是形式上的违规而已，通常不影响内容。

当然，在 input 较短 context 不足以确定内容完整性的的时候，有些反例会呈现歧义或甚至与原意相左的语义，这时候形式的违规的确与内容的混乱或不确定发生关联了。这时候，句法手段的修正（例如次序的调整、功能词的使用以及西方语言中的形态的正确应用等）才会有实质性意义，而不仅仅就是为了 native speaker 听上去顺耳而已。

解析和理解的能力，LLM 特别宽容鲁棒，主要是它的 embedding（编码嵌入，成为其内部的向量表示）可以容纳很长的 input，在 context 相互邻近的关键词之间相互制约下（我们叫篇章中的 semantic coherence，包括词义之间的搭配关系），形式上的偏离规范已经不影响它在语义空间的意义定位，从而“它”可以轻易与“非它”区分开来。

一个符号串吃进去就是向量空间的某个或某组位置其意义表现在与其他位置的距离和区别。因此位置偏差一点不影响意义只要它与其他的不同意义的符号串映射可以区别开来。鲁棒性根植于此。换个角度意义不是要问是什么，更要紧的是不是其他（什么），只要能维持这种意义空间的区别性，规范不规范就都可以包容。区别之间有足够的空间/距离，即可容忍局部的种种口误错误。

霄云：Llm 的 position encoding is linearly attached not cross product，so it is a weak form

立委：词序影响意义的机会不大。当年一包词模型用了很久也是因为词序是较弱的约束，构成区别要素的场景并不频繁。

我把一句话，完全反过来，从：explain quantum computing in simple terms 映射成类似回文：terms simple in computing quantum explain，它毫不迟疑。

人家训练的是next token，现在是处处反着来，本想让它找不着北，但实际上一点也不影响它的“理解”。就是说，当一个模型可以对较长的 input string 做编码嵌入的时候，次序的约束已经很弱了。因为那一小袋词之间的物理距离（proximity constraints）加上它们语义的相谐性（semantic cosntraints）已经足够让这个整体的语义表示与其他对象区分开来，这时候纯粹语言学意义的句法约束（syntactic constraints，包括严格的词序）就可以松绑。

我怀疑 position encoding 即便不做，LLM 也不见得性能会下降很多。

霄云：Could be, popular code base all use it still

立委：换句话说，在 bigram / trigram 建模的年代，词序是重要的 (“我爱她”与“她爱我”，“打死”与“死打”，可不是一回事)。到了ngram 中 n 可以很长的时候，ngram list 与 ngram set 已经语义相等了。

句长不够，词序来凑。长度足够，序不序无所谓。句法地位急剧下降。

论鲁棒，人如何与模型比，差了不止一个段位。

Li Chen：想想确实是这个道理，在有很多词的情况下，还要能组成符合语法的句子的可能性是有限的，也就意味着语义差异不大了。所以这个时候顺序确实已经不重要了，估计这个也是为什么即便是最简单的bag of words也能用来做相似度计算，一用就是几十年的道理。

卫东：跟chatgpt逗个乐。

总的感觉就是chatgpt对语言的嵌套理解能力和指代关系理解力非常强。

川：LLM 没问题，ChatGPT is evil

Who is the master, machine or man?

立委：那是因为 chatGPT 太 human like，搞定了自然语言形式。

川：搞定是假象，现在就下结论太早。

立委：机器都是假象，AI 本性。Artifical 与假象可以看成是同义词。就本质而言，人工智能就是智能假象，这个论断没有问题，但这应该并不妨碍人类深度使用AI。

搞定的判断是，跟他说话感觉它听从指令、善解人意，而且回应也很顺溜贴心，不走题。

三个月玩 chat 下来，我在它生成的英语中，没有发现过语言的问题（内容的毛病不算），一例也没有。但在其中文的生成中，偶然还是会发现它有语言的瑕疵（不符合规范或习惯的用法），虽然它的中文生成能力已经超过多数同胞。这说明，目前 chat 语言训练的中文语料还可以进一步扩大，从爱挑剔、追求完美的语言学家视角，它还有一点点剩余的进步空间。

结论还是: chat 搞定了人类语言，无论听还是说，妥妥的。万能的语言巴别塔是真滴建成了。

李维郭进《自然语言处理答问》（商务印书馆 2020）

【语义计算：李白对话录系列】