"What is generative? For example, draw me a picture of a busy traffic. This kind of thing, in the past, people didn't think it was something that artificial intelligence should do, but now it can be done." On May 18, Baidu CEO Robin Li was in Tianjin When I gave a speech at the Seventh World Intelligence Conference held, I said such a sentence.
Friends who don't know this stalk may not get the joke in this sentence and the more meaning behind it.
In March, when Wenxin Yiyan was just released, some netizens discovered that if Wenxin Yiyan was asked to draw a picture of "heavy traffic", it would really draw a picture with cars, water, horses, and dragons picture.
Some netizens joked that this is a word-for-word new art genre, and some people questioned Wenxin Yiyan’s technical strength based on this, and even said that AI that cannot read Chinese idioms may be a foreign model with shells, right?
In the past 3 months, this bug has long since ceased to exist, and the hot discussion at that time seemed to have died down. After more and more generative AIs came online, everyone found that no matter which model is Chinese or foreign, most of them have strengths and weaknesses. However, Wenxin Yiyan, which was first released, has become better the more it is used, and gradually has a unique meaning.
At this time, Li Yanhong once again mentioned "heavy traffic" with a smile, which seemed to have more meaning. First of all, this is not a kind of self-certification and whitewashing, because Wenxin Yiyan has completed self-certification in terms of technology and application. What Li Yanhong is expressing is more of a kind of calmness, calmness, and sincerity.
This is Baidu's technical view on AI, and it is also China's AI development view on the future.
birth of technology
Often greeted with cold eyes and ridicule
In fact, any technology that can attract attention is often accompanied by criticism and ridicule instead of welcome. This may be due to human nature, because destruction is easier than affirmation, and it attracts more attention, or it may be that innovation will inevitably break through the comfort zone of some people and certain things, so it will inevitably be negatively evaluated.
The earlier example is a famous joke of the Industrial Revolution: when the car was invented, it was used to race against a horse-drawn carriage, and the result was lost. At that time, countless coachmen laughed at this cumbersome iron lump, but later, there was no later.
In the middle of the last century, when the transistor was just born, people from all walks of life in the world were generally not optimistic about this fragile semiconductor material with a low yield rate. In the depths of history for people to pay tribute to.
To give a more recent example, when ChatGPT was first born last year, some people laughed at it for describing Li Yu and Nalan Xingde as people from the Ming Dynasty. They didn’t even know that Lin Daiyu was a character in "Dream of Red Mansions". Meanwhile, users who are more discovering than laughing at have found 10,000 ways to use generative AI.
In fact, it is always easy to find mistakes. There is always a shortcoming in any technology, and the same is true for a person. In an HR interview, you need to ask what the person is capable of, not what he does not know. Otherwise, if you look for it with an encyclopedia, a person will always have a knowledge blind spot. Problems like "heavy traffic" almost inevitably existed when Wenxinyiyan's model involving huge data and knowledge was just released. Today and for a long time to come, Wenxinyiyan, like other large models, has uncertainty and irrationality in many vertical functions.
Only by quickly discovering problems and quickly fixing them is the correct way to solve problems, and this happens to be the value of rapid iteration of large language models based on user feedback. Facts have proved that Wen Xinyiyan has indeed achieved this.
Then think about it from another angle, can Wen Xinyiyan wait for everything to be perfect before releasing it? First of all, perfection itself is a relative concept, and absolute perfection cannot exist. Baidu finds problems by itself, which is far less efficient than finding problems with global users.
Secondly, although it is safer to delay the release, it will waste industry opportunities and waste the window of opportunity for all walks of life to understand and participate in generative AI. Chinese AI needs to race against time, but Wen Xinyiyan under pressure and the possibility of being questioned and ridiculed, rushed to complete the release on March 16th.
Therefore, the birth of new technologies is almost inevitable to be questioned and ridiculed. The important thing is not what you can say to prove yourself, but what you can do next to prove the technology.
What can't beat you always makes you stronger
Li Yanhong can now mention "heavy traffic" in a calm manner, because Wen Xin has proved his technical strength and verified the correctness of the strategy of efficient release and rapid iteration.
The so-called "heavy traffic" and even more problems have not become obstacles on the development of Baidu's large-scale model, but have become evidence that Baidu can quickly hear feedback and complete iterations.
The end result can already be seen in the development data of the past few months.
From the perspective of operating performance, on May 16, Baidu released its unaudited financial report for the first quarter of 2023. Both revenue and profit exceeded market expectations, of which revenue was 31.1 billion yuan, a year-on-year increase of 10%, and net profit (non-GAAP) was 5.7 billion yuan, a substantial increase of 48% year-on-year. Analysts believe that Baidu's good momentum in Q1 is closely related to the positive user feedback and market response received by Wenxinyiyan. Baidu's head start at the level of large language models has been widely recognized by the market.
From the perspective of technical data, Wenxinyiyan not only maintains a leading position in many core technologies, but also brings the most efficient upgrade and development methods to the industry. Through the joint optimization of Wenxin Yiyan and Baidu's self-developed industrial-level deep learning framework Flying Paddle, the efficiency of an average of one iteration per week after the release of Wenxin Yiyan has been achieved. After 4 iterations, the reasoning of Yiyan has been realized. The capacity has been increased by nearly 10 times. This means that the model reasoning cost of Wenxinyiyan is greatly reduced, and it can provide services for 10 times the number of users.
It is not difficult to see that problems such as "heavy traffic" have not become Baidu's shortcomings, and have not hindered the pace of AI development. On the contrary, the existence of these "stalks" makes Wenxin Yiyan stronger and makes China's AI stronger.
selfless heart
In the Chinese science and technology circle, there are always some companies that are very taboo about the problems of their own technology, and they are always twitching, as if they have been caught by something. And this kind of attitude will naturally stimulate more heated discussions among netizens, and eventually form a vicious circle of "If you don't let me say it, then I will say it."
On the contrary, if you are really like Robin Li, take the initiative to talk about this little embarrassment, and face all this in a humorous and calm way, but it will open up the situation and broaden your horizons. Of course, what is needed here is the breadth of mind of an entrepreneur, and more importantly, a technical confidence, knowing that some problems do not constitute a problem at all.
Baidu's technical confidence is that Baidu does not only have a few models and a few algorithms. Large-scale models are going to all walks of life, thousands of lights, and models alone are not enough. Where did the model come from? What is the model supported by? It also needs multiple powers such as computing power, framework, and ecology to provide sufficient technical momentum to support the adaptation of the large model to the industry and continuously complete upgrades.
This accumulation is Baidu's real trump card in the field of AI. As mentioned earlier, the joint optimization of Wenxin Yiyan at the model layer and Flying Paddle at the frame layer has achieved a unique effect. Baidu has its own accumulation in the full-stack architecture of models, frameworks, development kits, tools, etc., and has leading core technologies. In the end, they adapt to each other and collaborate efficiently to bring Baidu "surprises" in the field of AI time and time again. ".
These "surprises" include the first release of Wenxin Yiyan, including efficient bug fixes, including the 4 iterations in January, and Li Yanhong's calm and calm.
On the contrary, other manufacturers that do not have self-developed deep learning frameworks and low-level technology accumulation, but rely on the wind to enter the AI field, either need to spend huge resources and experience on research and development from scratch, or redevelop based on deep learning frameworks such as TensorFlow and PyTorch, and often tend to lose sight of others. , In the end, I also lost my confidence.
On the road of innovation in China's science and technology, it seems that we are always talking about shells, and we will always talk about shells. In fact, how to prove the "whether the shell is covered or not"? Only with sufficient infrastructure and core technologies will gradually form more and more obvious differences, end users will prove everything, and ultimately time will prove everything.
Only by talking and laughing about the past, being calm about the future, and being brave about the present can we be selfless and broad-minded. Baidu is like this, and so is China's AI.
