All the challenges posed by technology are challenges to human inertia.
- Wu Xiaobo
All products are worth redoing with AI.
——Ye Jun, President of DingTalk
Finishing / Ba Jiuling (WeChat public account: Wu Xiaobo channel)
Dialogue / Wu Xiaobo, Ye Jun
From small models to large models
Xiaobo Wu: Welcome Dr. Ye Jun, today I invite him to talk about ChatGPT, Big Language Model (LLM) and AIGC (Artificial Intelligence Generated Content).
I first learned about AIGC before last year's year-end show, when ChatGPT 3.5 was released, and the CTO of a stock software company told me that there was a ChatGPT, you entered a requirement, it would help you program, you entered a paragraph, it would help you write a little essay. I was surprised.
Many students have not used or heard of the "big model". It's like Alvin Toffler added a big in front of data and invented the concept of "big data." What is the difference between the current "big model" and the model?
Ye Jun: What we are most familiar with is the system of recognizing license plates when entering the community, the car drives over, the pole is raised, and that can be understood as a small model. Large models identify larger amounts of data, and the trained algorithm is different from the previous algorithm.
Because the license plate only has a few letters after all, it will be easier to train the model. You can think of a model as a washing machine, a bread machine, and so on. Like a model of license plate recognition, it can only do one thing.
The large model has stronger capabilities, enters a large amount of data, and the parameter level is hundreds of billions. You can lose public data and information to it, and it constantly models neural networks. After the calculation, save the analysis results, and the next time there is input, it will take this model to process and give you an output. If the result is not right, you can also correct it and keep improving.
Wu Xiaobo: I'm a little confused, Chinese content only accounts for 1% to 2% of Internet information, and the remaining 90% of the content is English, French and other languages.
Then our Chinese companies, when doing large model data input, will there be no way to get it for some ideology and some reason? Or can we tap into the entire stock of human knowledge?
Ye Jun: This is a very critical issue, and input is very important for large models. What kind of information is entered determines whether the results of the training are good or not.
It is difficult for anyone to get all the knowledge of mankind, and everyone is trying to find a better corpus. Some specially trained graphics generation, such as Midjourney (MJ), which has done relatively well abroad, and the open source Stable Diffusion, are based on hundreds of millions of pictures to train, and the training corpus is very important.
Wu Xiaobo: At present, many large and small companies at home and abroad are making models, how supportive do you think the original capabilities of enterprises, such as computing power, server capabilities, and technical capabilities, are at this time?
Why am I asking this question? Because MJ has only 11 people and makes $100 million a year. Do small companies have the potential to disrupt the world, or do larger companies have greater opportunities in the future?
Ye Jun: I think in the future world, there may be tens of millions of models, and maybe each enterprise will generate its own models according to its own business needs, so there are very many models.
Every business has a chance. The big model is just a bottom model, combined with each enterprise, such as Mr. Wu's company, you can make your own exclusive model based on some expansion capabilities. So there will be a lot of companies like MJ in the future.
From iPhone moment to next iPhone moment
Wu Xiaobo: A while ago, DingTalk also held a press conference to demonstrate your AI products live. How were you feeling?
Ye Jun: At that time, I was still a little apprehensive, and we were worried that something would go wrong. Because there were problems during the rehearsal before. In fact, the product still needs to be polished, we hope to do more stable, of course, now it is better, we have given more than a hundred customers in the co-creation.
The big model we are based on is Tongyi Qianqian. DingTalk's AI product is "/" (slash), we call it magic wand, when chatting, writing DingTalk documents, and in meetings, as long as you hit "/", you can interact with the machine in natural language.
Wu Xiaobo: In addition to painting, what other functions does the magic wand have now?
Live demonstration of AI painting function
Ye Jun: Write a plan, write an email, write a plan, for example, you can ask it to write a plan, I am going to do a marketing event today, the theme is about, or write an email on your behalf, what needs to be purchased.
Wu Xiaobo: You said that there will be countless models in the future, for example, our company is a DingTalk user, purchased your large model service, and fed my data into it?
Ye Jun: Deploy an exclusive model for you, only you can use it. Therefore, every company will have a model in the future, just as every company has a printer now, and maybe this will be a standard in the future. So, data will become more and more important.
Wu Xiaobo: With AI, what kind of DingTalk is it?
Ye Jun: We are also working hard to redo each product of DingTalk with AI. DingTalk has a feature, especially many functions, we have developed a large number of functions after eight years of development. With AI, the first change is that DingTalk will become simpler.
We think back to the development of computers, which used to use keyboards and mice, and the day Jobs released the iPhone, we thought it was an iPhone moment and began to accept finger input. He slid his finger and turned on the iPhone to everyone remember.
Today, with AI, the interactive interface is neither a mouse and keyboard, nor a finger, and you have a new channel, which is natural language.
The transformation of finger interaction into natural language interaction is actually a very big change. In the past, a DSLR camera was designed, there were so many buttons on it, and in the future, maybe just one button, just press it. Product design concepts will change a lot.
Of course, the dependence on data behind this will become heavier and heavier. This is the second change, DingTalk is the same, we will pay more and more attention to data assets, every enterprise should pay attention to, data assets will really become a business element of a company.
So now could be another iPhone moment, the same thing as the fingers back then.
Wu Xiaobo: Nvidia Huang Jensen said that now is the next iPhone moment.
Ye Jun: It can be understood that AI can indeed make a lot of changes.
Wu Xiaobo: What will happen to DingTalk in that year's time?
Ye Jun: One year later, I think all DingTalk core products will be intelligent, I just reviewed the R&D progress of our products yesterday, search, customer service, documents, videos, every corner, will have intelligent capabilities, will be much more convenient than before.
Will AI eliminate me and my company?
Wu Xiaobo: The computer was invented in 1945, and by 1950, Turing had put forward a concept - machine thinking. He thinks it's possible for computers to have the kind of thinking that humans do. Since then, everyone has begun to push Turing's so-called machine thinking forward through various algorithm models.
In 2015, Yuval Harari published A Brief History of the Future, which already said that the so-called uselessness of man and the substitution of machines for man began to be reflected in the level of thinking. Today, after ChatGPT, it seems that something intimidating is starting to become a reality.
Is it possible for robots to defeat humans intellectually, or will humans always defeat robots?
Live demonstration of AI writing function
Ye Jun: I think it depends on the scene. In many places, robots certainly replace human jobs, but not every corner of the job can be replaced.
Wu Xiaobo: For example, I am a clothing company, 200 people, annual sales of 100 million yuan, what can the big model help me?
Ye Jun: Every industry is thinking about this. There are already companies in the clothing industry that are doing it, such as taking outdoor photos of models, changing dozens of sets of clothes a day. Now with AIGC, all I need is a dummy model, put my clothes on and take pictures, and then tell the AI what kind of action I can change into and where. It can generate 1,000 photos, and the effect is not obvious whether it is a real person or a virtual person.
Wu Xiaobo: Do you have a sense of fear that many professions will disappear? If I am a clothing company, now I want to reduce costs and increase efficiency, with a thousand big models, which positions will be killed first?
Ye Jun: When it comes to being killed at the post, everyone is panicking. I think it has an impact on the industry as a whole, not on an individual. For companies like this, designers should first try to learn to use AI tools. I won't use AI in the future, as if I can't use my mobile phone or WeChat now, I guess it's quite hard to go out and find a job.
Wu Xiaobo: Kevin Kelly said that in the era of artificial intelligence, what is our first ability as a person in the real world? It is the ability to ask questions, which is particularly important.
Ye Jun: So now there is a new position, prompt engineer, prompt is equivalent to asking questions, you can ask questions well, the level is also quite high.
We used to spend a lot of time thinking about how to execute better, and now everyone is upgrading. The boss is good at asking questions, the employees are good at execution, and when everyone has good questioning skills, you will improve to some extent.
Wu Xiaobo: DingTalk has an app store on it, and there are many application platforms, such as Cool Academy. So with AIGC, with Qianqian, will they have more opportunities, or will they be eliminated? Will their barriers be broken down?
Ye Jun: I think there should be more opportunities, but they must embrace AI. We are just a base model. They have this data and knowledge stored, and they need to keep a firm grip on their assets to generate their own professional models. The relationship between many models in the future should look like this.
Wu Xiaobo: They have to be strong enough. But strong enough means that there may have been 20 companies offering HR courses in DingTalk, but in the era of AIGC, this opportunity only belongs to 3.
Ye Jun: There are definitely several of them if there is this possibility. As for who can keep up, see whose AI capabilities enter in advance.
Wu Xiaobo: If I am doing decoration, I want to find a special human resources training in the decoration industry, and I will find more professional segmentation. Inclusive public courses may be eliminated.
Ye Jun: If the ability has 100 points, it should be said that the 80-point large model will eat it, when you enter a new field, such as the field of traditional Chinese medicine, through the ability of AI, you may quickly have the level of 80 points of old Chinese medicine, but between 80-100 points still need old Chinese medicine.
I feel that most industries will undergo such changes, especially in those industries that originally use models to build knowledge systems, doctors, lawyers, designers, in the past, you have to do a lot of training, copying, in fact, you are practicing this model, then there may be problems in the future.
Wu Xiaobo: People find that no, what is easy to replace are a few things: the first is the ability that can be obtained only through calculation, and the second is the ability to be more common-sense, the ability that you are doing and what I am doing.
It is not only individuals who are eliminated, but possibly some companies. There are several categories left: the first is the first person to practice these tools, who will be half a step ahead and have to act; The second is the future of human-computer interaction, the ability to ask questions is very important, you must be able to ask the right questions; The third is the ability of know-how in specific industries, the more professional and focused, the easier it may be to be problems and crawled by users in the future. Not only do you have it, but I am also good, and the overlapping part in the middle will be eliminated.
When optimists face social inflection points
Wu Xiaobo: In such a new era of AI, what advice and suggestions do you have for each of us as individuals and friends who run enterprises? Let's start with the individual.
Ye Jun: The core of an individual is to continue to learn and then practice it. You can't just listen to us, you will feel it when you run out. It's like everyone says that the car, you go to drive it, the impression it brings you is very deep, you can feel the designer's thoughts.
Wu Xiaobo: In a big era, we have a social inflection point - this inflection point is not an industry inflection point, but a social inflection point, and there are not many times when people encounter social inflection points in their lives. Electricity has been applied, computers have been invented, mobile phones have been invented, and this is called a social inflection point. Now, AIGC is out, which is called the social inflection point.
When social inflection points appear, many people will have an instinctive panic, will I be replaced? What to do? There is only one way to do it, and that is to rise to the challenge, face it, use it, and keep yourself alive.
What about advice for businesses?
Ye Jun: Enterprises should use digital tools to embrace digitalization very decisively and resolutely.
Wu Xiaobo: If you are in the 50s, 60s, and 70s, if you are running and managing a business, if your company wants to enter digitalization, remember that this matter cannot be entrusted to the office director. If you don't enter, your company will never enter.
Industrial intelligent transformation, office collaboration, digital capabilities, the first person who needs to practice, is the number one of the enterprise, you must go on your own, rush to the front line to use.
Ye Jun: It must be the entrepreneur's own first strategy.
Wu Xiaobo: If you use this tool and weapon first, you will be half a step or even one step ahead of your peers.
Kevin Kelly has a quote that I find particularly good, he says that I am an incorrigible optimist, and optimism is almost my faith and belief. I think that since human beings entered the agricultural civilization to the industrial revolution and the information revolution, every iteration and every social inflection point have eliminated a number of people, and even eliminated some industries. Who stayed? Those who remain are the optimists, those who can face difficulties, face challenges, and constantly change themselves.
I think all the challenges that technology brings are challenges to human inertia.
Organize | Ba Jiu Ling | Current value edit | Yang Shuai
Editor-in-Chief | He Mengfei | Image source | VCG