Walking out of the private club with Li Yanhong, Meng Fanqi was in a pretty good mood.

Robin Li is quite optimistic and persistent about the cutting-edge technology of AI, and he is very good at many conditions in this regard.

Meng Fanqi came into contact with him, and the feeling was quite good.

If this trial cooperation is done well in general, Meng Fanqi is too lazy to continue to look for other partners in China.

After all, at this point in time, the popularity of AI technology is far from high enough.

"Then let's meet again in a week or two." Li Yanhong shook hands with Meng Fanqi, "When the time comes, you prepare the algorithm and model, and we will prepare the contract and other relevant regulations." After the matter is finalized, I hope to come up with a real-time smart detection app to see how the market response to this technology is. ”

For the detection of content on images, there is a great qualitative key point is real-time or immediate.

Not all images that need to be inspected are prepared and pre-processed in advance.

The input of many scenes will be real-time shooting by the camera, and in this case, the speed of the algorithm is a very big bottleneck.

Gamers should be very familiar with the concept of FPS, which is, to put it bluntly, the number of frames per second, how many pictures are shown per second.

The higher the number, the smoother and silkier it will be. But at the same time, the number of images that the algorithm needs to process will also grow exponentially.

At present, the traditional algorithm with high accuracy, one or two seconds to calculate a picture, has been very fast.

Generally speaking, the FPS must reach at least ten or twenty before it can really be considered smooth real-time detection.

If you can reach 60-70 FPS, you can obviously have that extremely smooth, completely real-time processing feel.

"In this period, for the problem of slow detection algorithms, sampling is usually the method." Meng Fanqi recalled.

The principle of sampling is very simple, I can only process one image per second, then I don't care about other frames, just extract one frame per second for detection.

However, this practice will cause the frame to chase people, and when the detection frame is predicted, the objects in the frame have already run out of the frame, giving people a very poor user experience.

If most of the objects in the video content are moving fast, it is empty frames flying all over the sky, pure visual pollution.

"There is no problem, I will keep my promise and will not submit the paper or publish the technical details until at least six months later." This is also an important part of the conditions just discussed, and whiteness does not exclusively possess the technology.

But Meng Fanqi had to wait until June '14 before he was allowed to submit or disclose the technology. This time Meng Fanqi is completely acceptable.

After Meng Fanqi shook hands with Mr. Li and said goodbye, he refused the kindness of Li Yanhong to send a driver to him, and left Baidu's park alone.

Li Yanhong stayed where he was, looking at the thin pieces of paper in his hand, thoughtful.

After returning to the headquarters, Robin Li lay on his boss's chair and continued to think for a long time.

Immediately, he made two important decisions.

First, Baidu will open an AI research laboratory in Silicon Valley overseas.

Meng Fanqi's reason for rejecting himself made him wake up to a certain extent, and the overall disciplinary atmosphere and environment are also very important.

Last year, whiteness lost a sky-high auction with Google, and Canada's proximity to the United States was a big factor.

If you have an excellent Silicon Valley AI research laboratory, then the team that won for Hinton last year and just won Meng Fanqi to join will have a lot of excellent competitiveness.

At this point, my own reaction was slow. Li Yanhong reflected.

The second point is that whiteness needs to hire industry leaders who really have an international vision, and it is best to have experience in working with giant companies.

It takes a person like this to check whiteness.

Robin Lee's favorite role at the moment is Andrew Ng, a Chinese-American who is the director of the Stanford AI Lab and one of the founders of Google Brain.

If you don't move faster, then the whiteness will only lag behind more and more.

The third point was recommended to him by Meng Fanqi in the previous conversation.

The AI deep learning platform should be easy to use and open source.

In fact, this platform is already under secret development, and Meng Fanqi's suggestion surprised him. This shows that the technical talent in front of him also has a leader or entrepreneurial vision like himself.

By building such an open source community, we can attract a steady stream of AI algorithm talents to conduct research based on these tools.

This direction is still a desert, waiting to be explored.

Meng Fanqi's suggestion made Li Yanhong feel like seeing the sun.

Even if you can't recruit enough talent, as long as you maintain the platform, Baidu will never lack relevant conveniences.

It is necessary to lower the threshold for ordinary people to access AI algorithms.

It is up to everyone to use these models and techniques with as little manipulation, exposure, training, modification, as little as possible.

It is necessary to make the whiteness AI platform from a product that only a few scientists and large enterprises can afford, to the infrastructure of the whole society.

This is indeed something that Li Yanhong has not begun to consider at this stage, and Li Yanhong, who has just woken up from a dream, has improved his evaluation of Meng Fanqi by a few points.

While Li Yanhong contacted Yu Kai and other capable subordinates to discuss the details of various aspects, Meng Fanqi returned to the school library as usual to make final adjustments and tests on the code that was about to be handed over to Baidu.

He still prefers the feeling of a library, which is too comfortable at home or in his rented place, and there are countless temptations that can interrupt his work and thoughts.

Once the rhythm of learning and work is interrupted, the efficiency is likely to be very different.

"I'll just swipe my phone." Thinking so in my heart, pop, the afternoon passed.

Meng Fanqi, who was immersed in revising the code, didn't expect Li Yanhong to be so resolute at all, and today's meeting directly advanced many of Baidu's decisions by 5-7 months.

In fact, in the final analysis, it is the wage earners who underestimate the determination of the big technology companies to invest in this area.

Baidu's net profit exceeded 10 billion yuan for the first time last year, but Robin Li himself has a strong sense of crisis.

On the one hand, the core business of Baidu is still contributing to the steady growth of profits; But on the other hand, the mobile Internet is rapidly penetrating, and great changes in the business world are about to emerge, which makes Baidu fall into a certain confusion.

Meng Fanqi's AI algorithm performance in the first two and a half years is like a shot in the arm, which makes Li Yanhong, who has already locked the AI track, like a chicken blood.

After a year, Robin Li re-opened the internal letter he sent to all Baidu employees at this time last year, "Change, Start with You and Me".

"When the business is still growing rapidly, you should not quickly pursue net profits, but invest more money in more new business and innovation."