近日，在2023亚布力中国企业家论坛年会【闭幕论坛】上，微软（中国）有限公司首席技术官韦青，北京中科深智科技有限公司CEO成维忠，达观数据创始人兼董事长、CEO陈运文，彩云科技创始人&CEO袁行远，一起深度探讨ChatGPT未来的挑战与机遇。中国科学技术大学科技商学院执行院长叶强主持了本场论坛。
以下为对话内容：
叶强：各位嘉宾能否在通俗的语境下，说说你们对ChatGPT的认识。
韦青：大家千万别因为新技术一出现，就觉得无所适从。我们要积极探索，共同摸出这头“大象”到底长什么样子。建议大家不要专注在ChatGPT的名词上，这只是一个做出来的Demo，它真正的底层逻辑是被称之为Large language model的大语言模型。你可以将它理解成针对人类历史上所有发生的行为、现象、知识的提炼机或蒸馏机。
举例来说，如果你家的花园里种的是狗尾巴草，它蒸出来的就是狗尾巴草油；如果是茉莉花，蒸出来就是茉莉花油；如果全是玫瑰，蒸出来就是玫瑰花油。我们可以把它抽象成“人对世界”的关系，而这种关系中间靠“知识”来衔接。只有了解这些，我们才能真正理解，为什么这可能是新一轮的文艺复兴。
成维忠：我认为，GPT-4已经是一个思维模型，而并非一个大语言模型。这是通用人工智能的一个大突破，无论是ChatGPT还是GPT-4，我们都要深度学习和积极拥抱。如果单纯从大语言模型来看它，就会产生很多认知偏差和误判。
现在新的工业革命已经开始了，ChatGPT尤其是刚出来的GPT-4相当于历史上的蒸汽机。这是一个新时代，而这个时代可能会影响未来20年甚至更长时间。
陈运文：ChatGPT火了以后，每个人都能用它进行对话和聊天。企业中各种财务、人事、合同、订单、报销、报告、投标等日常白领的工作，未来都可以由专属企业自己的大语言模型系统去完成。这个大语言系统可以学习企业每个员工日常工作的规则和行业知识，最终能自动化地完成上述工作。
达观数据正在致力于开发专业化的、垂直化的、国产的GPT模型。在接下来的2-10年，希望通过我们大语言模型的助力，再造中国商业生态，让企业效率和成本大幅降低，让中国的企业能在全球的竞争中，获得更强劲的竞争优势。
袁行远：前阵子，我用GPT-4去做北京2022年的高考语文，它的有些能力非常惊人，它能理解题目中复杂的指代关系，最终得到正确答案。经过我们的测算，它的语文试卷能得128分左右，在高考总分750分的情况下，如果每个科目它都能达到这个水平，那它的高考大约能考个630分。
这相当于“一个具有大学生智力水平的人” ，这个“人”还是可编程的，你可以跟它进行互动，可以把你的问题抛给它，它进行解决后再返回来。但它和普通大学生有两个不同：一是它知道世界上几乎所有的语言；二是它几乎知道所有的领域。可以理解为，你有一个几乎知晓所有领域的知识且可编程的“个人智能助理”。
叶强：最近有几位做NLP（自然语言处理）领域的学者跟我说，ChatGPT的表现实在太惊艳，这让他们感到不太自信。你们业界对此是什么感觉，还自信吗？ChatGPT的出现，对你们来说，是威胁更多还是机遇更多？
袁行远：突然发现行业里有个人专业做得特别好，一开始当然压力倍增。这种感觉就像自己喜欢的女生最后跟别人结婚了。自己喜欢的这个女生，就是通用人工智能，她跟别人结婚，而你没有追到她。
但从另一方面看，它也告诉我们，新大陆是存在的。这年头最重要、最困难的问题，其实不是怎么到达新大陆，而是有没有新大陆。现在有人告诉你，新大陆是有的，那你应该非常兴奋才对。而且这扇门刚刚被打开一个缝隙，你应当非常有干劲，并且朝着未来继续努力。
不能因为别人引爆了原子弹，我们就坐以待毙。我们应该奋起直追，至于能不能做到，可能没有那么重要，但首先你得相信，得开始干，这才重要。
陈运文：大语言模型GPT的参数规模大到一定程度后，都是上千亿神经元的参数规模，而我们人脑总的神经元数量也才140亿个。今天我们训练的大语言模型都是1700亿甚至两三千亿的参数规模，它的容量大约是人脑的10倍以上。
以前任何人工智能系统还从来没有训练过这么大规模的参数量，当它的参数规模突破一个临界点时，会出现一种智慧的“涌现”现象，这个人工智能系统的能力就会有一个质的飞跃。
过去训练一个人工智能系统，要花费很大的精力去给计算机标大量的数据，做大量模型的调参等，才能让计算机勉强达到一个人类白领的工作能力，但现在拥有具备“涌现”能力的人工智能系统后，它立刻就会举一反三、触类旁通，它学习新的工作岗位技能的能力前所未有，未来这个技术可以用在各行各业各岗位。这是一件对未来的10-20年都非常有影响力的事情。
虽然我们国内的国产大语言模型整体还是落后国外1-2年时间，但我们还在紧密追赶。我们跟别人差距没那么大，中国的国产大模型一定会做出来，而且一定会能发挥很大的价值。
成维忠：人机交互的历史主要有三个阶段：第一个阶段是文字用户界面（CUI）；第二阶段是图形用户界面（GUI）；ChatGPT和GPT-4发展以后，就到了第三个阶段“自然用户界面（NUI）”。但GPT-4这种大语言模型，和我们真实的人和人之间的交流，还有较大差异。人和人之间交流主要依靠视觉，而视觉占人类交流的80%多。ChatGPT和GPT-4现在还做不到视觉交流，需要它们跟生成式AI虚拟人相结合，实现人和“人”的交流。
与国外相比，国内的大语言模型还有差距。国内刚出来的新模型我们都测过，国内在文本生成方面，赶超国外的步伐很快，但在文生图等多模态的理解上，差距还非常大。这个差距能不能在短期里弥补上，现在还不知道。也许会像前面几位老总讲的，一两年的差距，甚至更长时间，但我们行业迭代速度很快，1—2年的时间其实差距已经很巨大。
此外，现在大家还把GPT-4和ChatGPT看成是大语言模型，这也是有问题的，它正在往多模态的思维模型的方向去发展，不仅仅是个大语言模型。我们要站在更加开阔的全球视野上，打开心胸积极学习和拥抱世界的最先进的技术和方法。
韦青：判断新事物发生的脉络，首先要察其言、观其行、审其因、辩其果。 “草蛇灰线，伏脉千里”是有脉络的。我们要警惕的是，是否还在用前一时代的思维方式，去想下一时代的问题。
世界变了，整个知识体系、思维体系都变了。我们很多人还不理解信息时代是怎么回事，语言如何表征我们的思想，行动如何表征这个世界？这方面我们的研究还是很弱的。
我自己是随着中国改革开放起来那一代，那时的一个原生逻辑是“做”，但“做”又不能不去“想”，而“想”和“学”，在这个时代是经常缺失的。我们以为在“想”，却是基于旧范式在“想”，那怎么基于新范式？
无论一个模型有多牛，回到第一性原理，它都有脉络和体系。确定了脉络之后，我们就会更加清晰。很多时候大家不太明白什么是新范式，很多文章把“机器”神话或妖魔化，但很难把“机器”回归本原。机器是机器，人是人，如果把人活成了机器或者把机器弄成了人，那一定会出问题。
很多语言对我们潜意识造成了误导。Robot在英文中就是代表一种“自动机”的含义，但在我们的语言中被翻译成了“机器人”。为什么要把一个自动机翻译成一个机器人？这种过早丧失人的主观能动性的物化，到底是把机器抬太高了，还是把人压成了机器？
总之，一定要反思，我们的每一个念头是不是有偏差。信心肯定要有，但成功不是来源于信心，而是来源于我们自己的改变和努力。
叶强：ChatGPT在哪些方面体现了它非常牛？它还有哪些地方还比较弱？大家能否举两个例子具体说说。
成维忠：企业过去做的关于“流程”方面的工作，而大多数流程的设计是为了让机器更好去理解人的意图和需求，现在要重新去认知它的合理性和价值。
这次微软的Copilot出来以后，通过GPT可以很简单地让机器知道我们要它去干什么。举个例子，原来我要三步才能完成购物，现在我旁边有一个小秘书，我直接让他帮我买一瓶水，他就帮我干完了，GPT就是这个小秘书。过去我们不知道怎么让机器理解人类，现在机器理解人的能力变高了，所以它能够干的事儿就非常多。
这对每一个企业都会产生巨大的冲击，要积极拥抱它。原来我们公司的流程也很复杂，通过不同业务流程去管理各项工作。同时，我们还有跟流程配套的资源库、知识库、图片库，等等，这些都是我们的财富和优势。但有了这种文字自动生成图片的技术以后，我们的机器能自动生成各种需要的内容，库的价值就大幅下降了。同时，机器更理解人了，“流程”也可以大幅简化。原来很有价值的业务流程和各类资源库，现在很有可能是没价值的。
陈运文：关于“流程”，我也补充一点。达观数据专门给企业提供定制化的大语言模型，用机器人帮公司完成自动化的审批工作，在这之前，我们有大量的流程可能要层层审批，但有了大语言模型以后，一个机器人可以快速且客观公正地帮你完成很多事项的自动化审批工作，能够大幅提高工作效率。
但它当前的缺点是对一些特别具体数字类的文字推演、分析方面的能力，还是有短板。比如，让它做一个“鸡兔同笼”的问题，它还是做不好的，尽管做不好，但它会给你讲得头头是道，最后给你一个错误的答案。所以在涉及特别精密数字的运算推理领域，我们还是需要叠加一些有监督的专用技术，才能弥补当前的短板。我们在用它做精确数字混合运算时，还是需要保持警惕。
袁行远：我举两个例子。有个高考语文题：
“核酸检测排队时需要两米安全距离，一些社区为两米间隔线设置了安全贴心、形式多样的标志，有的是撑起的晴雨伞，有的是贴在地上的古诗词图片。请选择一个检测点，依据其环境特点，设计两米间隔线标志，并写出设计理由。”
ChatGPT针对这个问题，给出了一个答案：
“可以选择一个公园设计成核酸检测点，在地上贴上公园里的植物介绍，这样大家可以一边等待检测，一边看植物介绍。”
还有一个用户案例：
最近某天深夜，我老婆突然肚子不舒服，我把这个情况告诉了GPT-4，它很快给出了几种可能，我按照它分析的可能性去做，后面果然也没什么问题。我对孕妇相关的事一窍不通，它却能给到我一些看起来还可以的建议。这给了我很大的震撼。不过语言模型的弱点就是，它与现实世界、物理世界交互之后，还是会经常出各种错。
韦青：人是不是用“概率”思维，我们还要去研究，但机器是用“概率”思维，它是以概率的方式理解我们的，懂得的是概率推理，而不懂得概念推理。机器擅长记忆，它的这种记忆表现方式目前来看是静态的、平的，所以它在连续对话时有可能就会犯错，因为它没有过去时间轴、单向轴的这种方向的记忆。
总的来说，一是概率，二是记忆，三是时间和空间。这几个关键词理解完之后，人才能认识这个工具。比如，你发明了一把枪，没见过枪的土著会怎么用这把枪？你给了他一把枪，结果没有打到野兽。到底是因为这把枪不好，还是由于他不懂什么是枪而把它当成打狗棍来用了？这是一个问题。最重要的还是得去试用。
叶强：由于时间关系，最后再问两个问题：一是从管理角度看，我们的技术创新会发生哪些变化，企业要做哪些管理创新？第二，对人才的培养，各位有什么见解？你们可以挑一个问题来回答。
陈运文：从管理的角度来说，这次大语言模型有一个特别大的特征就是典型的厚积薄发。它把人类历史上所有的教科书、论文、报告、资料都沉淀下来，积累了巨大规模的参数量，在六七年的时间里投了无数的资金，而且一分钱不挣。OpenAI走的这条技术道路原本学术界并不太看好，因为这条路大家都觉得特别费时费力，吃力不讨好，还都感觉好像走不通，但OpenAI能够坚持下来，积累这么多，最终登顶，这非常了不起。
很多时候，我们的科技创新要有耐心，也需要有厚积薄发的精神。有时大家太过于急功近利，一个技术还没有成熟就立刻想着怎么变现挣钱，或者遇到困难就半途而废，这很难真正做出巨大的创新成果。希望未来大家能更加脚踏实地地积累更多的数据，打磨好更多的模型，扎扎实实地把技术做出来之后，再在各行各业应用。
成维忠：这次OpenAI的成功，与企业家精神有非常大的关系。我干的这个行业做的是多模态大模型，我的合伙人是我的中科大同学，但我俩一不是美国名校博士，二不是教授，经常会有人问，你们为什么会干这个事儿？你们能干得出来吗？你们是从美国回国的教授吗？这些质疑其实是有问题的，OpenAI的CEO连大学都没有读完，他们能成功，企业家的使命感和执着探索的精神在其中发挥了主要作用。
所以我们不在乎被质疑，一直在干。支撑我们坚守的核心，就是企业家精神，就是干自己相信的、喜欢的事。
韦青：经常听到有这种说法：大学是培养“会思考的人”的大学，不应该只是培养“学位”的大学。这话说得很容易，但在中国推行起来很难。我就想给各位企业家们提个建议，企业在招人时，如果不是以最后“会用和会思考”的标准来招聘，那么这种社会风气还是改不掉的，建议应聘的方式尽量选择一种实操型的方式，而不是只看学历和证书。
袁行远：如果有机会，希望管理学能去研究一下ChatGPT这种类似的技术，以帮助我们进行管理。是不是将来包括OKR、年度报表等在相关的绩效考核，可以丢给机器去做？希望新技术能帮助企业找出更多问题，发现更适合的人才，增强管理的团队建设，以及带来新的管理体验。
翻译：
Recently, on the closing forum of the 2023 Yabli China Entrepreneur Forum Annual meeting, Wei Qing, Chief Technology Officer of Microsoft (China) Co., LTD., Cheng Weizhong, CEO of Beijing Zhongke Shenzhi Technology Co., LTD., Chen Yunwen, founder, Chairman and CEO of Daguan Data, Yuan Xingyuan, founder &CEO of Caiyun Technology, Discuss ChatGPT’s future challenges and opportunities in depth. Ye Qiang, Executive Dean of the Business School of Science and Technology, University of Science and Technology of China, chaired the forum.
The following is the content of the conversation:
Ye Qiang: Would you please tell me your understanding of ChatGPT in a popular context?
Wei Qing: Don’t feel confused just because of the emergence of new technology. We need to actively explore and find out what this “elephant” looks like together. I don’t want you to focus on the nouns in ChatGPT. This is just a Demo. The actual underlying logic is a Large language model called the Large Language Model. You can think of it as a distiller or distiller for all the actions, phenomena, and knowledge that have occurred throughout human history.
For example, if you grow setaria grass in your garden, it will steam setaria oil. If it is jasmine, it comes out with jasmine oil. If it’s full of roses, it comes out with rose oil. We can abstract it as the relationship between “man and the world”, and this relationship is connected by “knowledge”. And only by understanding that can we really understand why this could be a new Renaissance.
Mr. Cheng: I think GPT-4 is already a thinking model, not a big language model. This is a big breakthrough for general artificial intelligence. Whether it’s ChatGPT or GPT-4, we need to learn deeply and embrace actively. If you look at it purely in terms of a large language model, it leads to a lot of cognitive biases and misjudgments.
Now that the new industrial revolution has begun, ChatGPT, especially the new GPT-4, is the historical equivalent of the steam engine. This is a new era, and one that could shape the next 20 years or more.
Chen Yunwen: When ChatGPT became popular, everyone could use it to have conversations and chat.
In the future, all kinds of daily white-collar work such as finance, personnel, contract, order, reimbursement, report and bidding can be completed by the exclusive enterprise’s own large language model system. This big language system can learn the rules and industry knowledge of the daily work of every employee in the enterprise, and finally can automate the above work.
Dar Guan Data is committed to the development of professional, vertical, domestic GPT model. In the next 2-10 years, we hope to re-create China’s business ecology with the help of our big language model, greatly reduce the efficiency and cost of enterprises, and enable Chinese enterprises to gain stronger competitive advantages in the global competition.
Yuan Xingyuan: Some time ago, I used GPT-4 as the language of Beijing 2022 college entrance examination. It has some amazing abilities. It can understand the complex reference relationships in the questions and finally get the correct answer. We calculated that it would get about 128 points for the Chinese paper and 630 points for the gaokao if it achieved this level in every subject out of a total of 750.
It’s the equivalent of “a person with the intelligence of a college student,” a “person” that is programmable, that you can interact with, that you can throw your problems at, that it can solve and come back. But there are two differences between it and the average college student. One is that it knows almost all the languages in the world. The other is that it knows almost everything. In other words, you have a programmable “personal intelligent assistant” that knows almost everything.
Mr. Ye: Recently, a couple of NLP (natural language processing) scholars told me that ChatGPT’s performance was so impressive that they didn’t feel very confident. How does your industry feel about that? Are you confident? Is ChatGPT more of a threat or an opportunity for you?
Yuan Xingyuan: Suddenly found that there is a professional in the industry to do particularly well, of course, at the beginning of the pressure. It’s like when the girl you like ends up marrying someone else. I like this girl, is general artificial intelligence, she married someone else, and you did not catch her.
But on the other hand, it also tells us that the New World exists. The most important and difficult question these days is not how to reach the New World, but whether there is one. Now that someone tells you that the New World exists, you should be very excited. And since the door has just been opened a crack, you should be very motivated and keep working towards the future.
Just because someone else detonated an atomic bomb, we don’t stand idly by. We should catch up. Whether we can do it or not may not be that important, but you have to believe in it first and start doing it. That’s what matters.
Chen Yunwen: The parameter scale of the large language model GPT reaches hundreds of billions of neurons, while the total number of neurons in our human brain is only 14 billion.
The large language models we train today have a parametric scale of 170 billion or even 200 or 300 billion, which is about 10 times the capacity of the human brain.
Before any artificial intelligence system has never trained such a large number of parameters. When its parameter scale breaks through a critical point, there will be a wisdom “emergence” phenomenon, the ability of the artificial intelligence system will have a qualitative leap.
In the past, it took a lot of effort to train an artificial intelligence system to label the computer with a large amount of data and make a large number of model adjustment and so on, so that the computer could barely reach the working capacity of a human white-collar. But now, after having an artificial intelligence system with the ability of “emergence”, it will immediately draw inferences from one example and draw inferences from another. Its ability to learn new job skills is unprecedented, and the technology could be used in all industries and positions in the future. This is a very powerful thing for the next 10-20 years.
Although our domestic large language model is still 1-2 years behind foreign countries, we are still closely catching up. Our gap with others is not so big, China’s domestic model will be made, and will be able to play a great value.
Cheng Weizhong: There are three main stages in the history of human-computer interaction:
The first stage is text user interface (CUI); The second stage is the graphical user interface (GUI); After ChatGPT and GPT-4 evolved, the third stage was called the Natural User Interface (NUI). But GPT-4, a big language model, is still very different from how we interact with real people. Vision accounts for more than 80 percent of human communication. ChatGPT and GPT-4 are not yet capable of visual communication. They need to be combined with generative AI virtual humans to communicate with “people”.
Compared with foreign countries, there is still a gap between domestic large language models. We have tested all the new models that have just come out in China. In terms of text generation, China is catching up with foreign countries very quickly, but there is still a big gap in the understanding of multi-mode such as Vincennes diagram. Whether this gap can be closed in the short term remains to be seen. Maybe it will be like the first few bosses said, a gap of one or two years, or even longer, but our industry iteration speed is very fast, 1-2 years in fact, the gap has been very huge.
In addition, GPT-4 and ChatGPT are now considered large language models, which is also problematic. It is moving towards a multimodal model of thinking, not just a large language model. We should take a broader global perspective and open our minds to learning and embracing the world’s most advanced technologies and methods.
Wei Qing: To judge the context of the occurrence of a new thing, we must first observe its words, observe its actions, examine its causes and debate its effects.
The “grass snake grey line, Voluminescent Li” is a vein. What we should be vigilant about is whether we are still thinking about the problems of the next era in the way of the previous era.
The world has changed, the whole system of knowledge, the whole system of thinking has changed. Many of us still don’t understand what’s going on in the information age, how do words represent our thoughts, how do our actions represent the world? Our research on this is very weak.
As a member of the generation that followed China’s reform and opening up, one of the primary logic at that time was “to do”, but “to do” could not be done without “thinking”, and “thinking” and “learning” are often missing in this era. When we think we are “thinking”, we are “thinking” based on the old paradigm. How can we “think” based on the new paradigm?
No matter how impressive a model is, going back to first principles, it has a vein and a system.
Once we’ve identified the vein, we’ll have more clarity. In many cases, people do not quite understand what the new paradigm is. Many articles mythologize or demonize “machine”, but it is difficult to return “machine” to its original origin. Machines are machines and people are people. If people are made into machines or machines into people, there will be problems.
Many languages mislead our subconscious mind. Robot means “automaton” in English, but it translates to “robot” in our language. Why translate an automaton into a robot? This kind of premature loss of human subjective initiative of the materialization, after all, is lifting the machine too high, or crush people into machines?
In short, we must reflect on whether each of our thoughts is wrong. Confidence must have, but success does not come from confidence, but from our own change and efforts.
Ye Qiang: In what ways does ChatGPT stand out? Where else is it weak? Can you give me two examples?
Cheng Weizhong: Most of the “process” work that enterprises used to do was designed for machines to better understand people’s intentions and needs. Now they need to re-recognize its rationality and value.
When Microsoft’s Copilot comes out, GPT makes it easy to let the machine know what we want it to do. For example, it used to take me three steps to finish shopping, now there is a little secretary next to me, I directly ask him to help me buy a bottle of water, he did it for me, GPT is this little secretary. In the past, we didn’t know how to make machines understand people, but now machines are better at understanding people, so they can do a lot of things.
This will have a huge impact on every business, so embrace it. The process of our company is also very complex, through different business processes to manage the work. At the same time, we also have the resource library, knowledge base, photo library and so on, which are our wealth and advantages. But with text generating images, our machines can automatically generate all the content we need, and the value of the library is greatly reduced. At the same time, machines understand people better and “processes” can be greatly simplified. Business processes and repositories that used to be valuable may now be worthless.
Chen Yunwen: I would like to add something about the “process”.
Dantan Data specializes in providing customized large language models for enterprises, using robots to help companies complete automatic approval work. Before this, we have a large number of processes may require layers of approval, but after the large language model, a robot can quickly and objectively help you complete a lot of matters of automatic approval work, which can greatly improve work efficiency.
However, its current shortcoming is the ability to deduce and analyze the text of some special specific numbers. For example, if you ask it to do a “chicken and rabbit in the same cage” problem, it still can’t do well, although it can’t do well, but it will explain to you, and finally give you a wrong answer. So in the field of computational reasoning involving particularly sophisticated numbers, we still need to add some supervised specialized technology to make up for the current shortcomings. We still need to be careful when we use it to do precise numerical mixing.
Mr. Yuan: Let me give you two examples. There is a Chinese question in the college entrance examination:
“The line for nucleic acid testing requires a safe distance of two meters. Some communities have set up a safe and considerate sign for the two-meter line, from umbrellas held up to pictures of ancient poems posted on the ground. Please select a detection point, according to its environmental characteristics, design a two meter interval line sign, and write the design reason.”
ChatGPT has an answer to this question:
“You can choose a park to be designed as a nucleic acid test site, and paste descriptions of plants in the park on the ground, so that people can wait for the test while reading plant descriptions.”
Here’s another user case:
Late one night recently, my wife suddenly had an upset stomach. I told GPT-4 about the situation. Gpt-4 quickly offered several possibilities. I didn’t know anything about pregnant women, but it gave me some advice that seemed OK. It gave me a big shock. But the weakness of language models is that they often get things wrong when they interact with the real world and the physical world.
Wei Qing: We still have to study whether people use “probabilistic” thinking, but machines use “probabilistic” thinking.
They understand us in a probabilistic way. They understand probabilistic reasoning, but not conceptual reasoning. The machine is good at memory, and its memory representation is currently static and flat, so it may make mistakes in continuous conversation, because it has no memory of the past time axis, one-way axis.
In general, one is probability, two is memory, and three is time and space. After understanding these key words, people can know the tool. For example, if you invent a gun, what would a native who had never seen a gun do with it? You gave him a gun and it didn’t hit the beast. Was it because it was a bad gun, or because he didn’t know what a gun was and was using it like a dog club? This is a problem. The most important thing is to try it out.
Ye Qiang: Due to time constraints, I would like to ask two questions at last. First, from the perspective of management, what changes will happen to our technological innovation and what management innovation should the enterprise do? Second, what are your views on talent cultivation? You may choose a question to answer.
Chen Yunwen: From the perspective of management, one of the great features of this big language model is the typical accumulation. It deposited all the textbooks, papers, reports and data in the history of mankind, accumulated a huge amount of participation, invested countless funds in six or seven years, and did not earn a single penny. OpenAI took a path that was not viewed very well by the academic community, because it was a path that everyone thought was very time consuming, thankless, and seemingly impossible, but OpenAI has persevered, accumulated so much, and finally reached the top, which is remarkable.
In many cases, our scientific and technological innovation requires patience and a spirit of accumulation. Sometimes people are too eager to achieve quick results. They immediately think about how to make money before a technology is mature, or give up halfway when faced with difficulties, which is difficult to really make great innovation results. I hope that in the future, we can accumulate more data more down-to-earth, polish more models, and then apply the technology in all walks of life.
Cheng Weizhong: The success of OpenAI has a lot to do with entrepreneurship.
What I do in this industry is multi-modal large model. My partner is my classmate from USTCM, but we are not doctors of famous American universities or professors. People often ask me, why do you do this? Can you do it? Are you professors returning from America? These doubts are actually problematic. The CEO of OpenAI didn’t even finish college, but the entrepreneurial sense of mission and the spirit of persistent exploration played a major role in their success.
So we don’t care about being questioned, we just keep doing it. The core that sustains us is entrepreneurship, which means doing what we believe in and love.
Wei Qing: You often hear the saying that a university is a university for cultivating “thinking people”, not just “degrees”. That’s easy to say, but hard to implement in China. I would like to give a piece of advice to all entrepreneurs. If enterprises do not recruit people based on the standard of “being able to use and think”, then the social atmosphere will not change. I suggest that the way of employment should be practical, rather than just based on academic qualifications and certificates.
Yuan Xingyuan: If there is an opportunity, I hope management can study similar technology like ChatGPT to help us manage. Will the future include OKR, annual report and other related performance appraisal, can be thrown to the machine to do? The hope is that new technologies will help companies identify more problems, find more suitable talents, enhance team building for management, and bring about new management experiences.
本文由数字化转型网（www.szhzxw.cn）转载而成，来源：网络；编辑/翻译：数字化转型网宁檬树。
免责声明: 本网站(http://www.szhzxw.cn/)内容主要来自原创、合作媒体供稿和第三方投稿，凡在本网站出现的信息，均仅供参考。本网站将尽力确保所提供信息的准确性及可靠性，但不保证有关资料的准确性及可靠性，读者在使用前请进一步核实，并对任何自主决定的行为负责。本网站对有关资料所引致的错误、不确或遗漏，概不负任何法律责任。
本网站刊载的所有内容(包括但不仅限文字、图片、LOGO、音频、视频、软件、程序等) 版权归原作者所有。任何单位或个人认为本网站中的内容可能涉嫌侵犯其知识产权或存在不实内容时，请及时通知本站，予以删除。
