In early January 2014, when it was the time to say goodbye to the old and welcome the new, Meng Fanqi had many things at hand, both fast and slow, and it happened to be harvest time at this point in time.

After returning from rebirth, Meng Fanqi has been unsure how much he, a butterfly with flapping wings, can change the world.

The reason why I was able to obtain these funds and start these two companies was completely dependent on another time and space that people in this world could not observe, which is quite similar to the physical meaning of "horizon".

In addition, "horizon" is also homophonic with "world", implying the determination to go to the world in the future.

Therefore, he registered the two companies he founded for the first time as Vision Vision Technology Co., Ltd. and Vision Intelligent Medical Co., Ltd. respectively according to the specific technical direction.

Among them, intelligent medical care is still just an empty shell, and nothing is being promoted for the time being.

On the Depth Vision side, although the camera app, related filters, function interfaces, and functions such as payment are already in full swing.

But in reality, the real core part of the model is still being trained.

After handling some of the initial affairs on his own company's side, Meng Fanqi returned to Google's Shanghai branch again.

On the one hand, it is to finish the follow-up two axes of the recommendation advertising algorithm, and the second is to accept its own FaceGAN face generation model.

"Am I a public weapon for private use?" Meng Fanqi really couldn't say clearly.

FaceGAN is a subdivision direction variant based on Meng Fanqi's previous generative network, which is mainly trained on a huge dataset of face images, and its purpose is to perform realistic generative fitting for a single element of face.

The technology doesn't have a very strong monetization effect at the moment, partly because the number of pixels supported is far from enough, and secondly, because the computing power required to achieve good enough results is quite large.

In the future, it can achieve video face swapping, AI painting, style switching of various pictures, or high-definition restoration of pictures, which will have actual commercial value.

Therefore, this technology will be the first serious visual technology achievement that Meng Fanqi has contributed after signing a contract with Google Brain.

Although this version does not have high commercial value, the functions it realizes are very symbolic, and it can be said that it has sounded the alarm bell for human society.

On the day Ian came up with the idea one step behind Meng, Bengio had already predicted a similar dilemma, and that in the near future people would begin to face ethical issues related to artificial intelligence.

How is this AI-generated content regulated, defined, restricted, and handled?

Humanity is still blank in this regard.

In Mountain View, California, after receiving the FaceGAN paper from Meng Fanqi, Hinton suddenly felt relieved.

Some time ago, he also heard that Meng Fanqi specialized in the recommendation advertising algorithm in Google Shanghai and made money, which made him very sad.

In his opinion, anyone can do these things to make money, but Meng Fanqi's talents are not used to promote artificial intelligence, especially visual intelligence, but to follow the advertising team to make money.

This is simply a tyrannical thing, nonsense! Not at all.

But Google is a commercial company after all, and advertising is a fist revenue product.

As an employee, Hinton had no choice but to give up because it was inconvenient to express any opinions.

At this moment, when I saw the work of FaceGAN, I learned that Meng Fanqi was still very attentive to visual intelligence, and I was immediately relieved.

"Meng! Long time no see! I'm glad you didn't forget your visual algorithm while doing the recommendation advertising algorithm! After the video call was made, Hinton's face was palpable.

"I just finished reading your paper, and your ideas and operations are very ingenious. Change the random noise input from the generator to a low-resolution version of the face image, and let the generator do the upsampling and fine-tuning repair by itself. ”

Meng Fanqi's generative technology has been unveiled for more than two months, and so far there has been no particularly decent follow-up.

Hinton saw the most eye-catching, and it had to be the GAN author himself, Meng Fanqi's series of operations on the face.

"The potential to generate network variants is still very large, and in addition to making some fine-tuning of the structure and components, I have also made some slight modifications to the function of measuring loss, adding adversarial losses to encourage the generator to make high-resolution faces to deceive the discriminator."

Meng Fanqi and Hinton had an in-depth exchange and discussion, but they were all "generated" in this regard.

Another powerful feature of the GAN series network, "editing", Meng Fanqi deliberately did not mention it.

This part will be the primary buying point of the camera app in the past few months, and Meng Fanqi does not want to make a difference.

Due to time constraints, he didn't have that much energy to do a very detailed comparative test and make a website launch.

The meticulous part of the experiment, as well as the fake face generation site [none of these people existed], was entrusted to the rest of the Google Brain team.

The hidden layer editing function that has not been released is handed over to my own vision company along with the model, so that there will be an initial test version tomorrow.

"You've updated two more versions of your recommendation algorithm?" Zhou Huan came back from the holiday to see Meng Fanqi, and before he finished greeting, he heard a message that made his jaw drop on the ground.

"I haven't been able to digest that exaggerated result last time." Chen Shen said a little sourly beside him, it is really better than people to die.

The two of them haven't been able to recover from the previous shock even after Christmas and New Year's Day, and this kid is good enough to bring two more bombshells.

"It's not going to be a mess again this time, is it?" Zhou Huan still remembers that after the last update was all launched, although the revenue increased, due to the wide range of traffic punishments, some related customer service auxiliary employees were all tired.

There are many unethical members of the garbage dump website, pretending to be innocent and not reluctant, which is very tormenting.

"Fortunately this time, the optimization of the mobile terminal is basically moisturizing and silent. In terms of language understanding, it is only to give users a better search experience, and the impact will not be as great as last time. ”

The current search patterns of Google and Whiteness are just a temporary product, a transitional product, of the inability of machines to understand the meaning of language.

It's a bit like the telegraph before the invention of the telephone, and before the invention of the telephone, there were many people who thought that "the telegraph was good enough that no one would use something like a telephone".

But it's the phone that gets closer to what people really need.

Search engines should also become more intelligent, rather than mechanically retrieving every word of user input.

For example, the same keyword has completely different meanings in different sentences and contexts, and this phenomenon is especially serious in Chinese.

In the early days, when Internet users searched for things, they often used a few keywords with spaces, just because search engines couldn't understand long sentences.

If your sentences are too long and there are too many contents, the search will be cluttered or unsearchable.

But now everything is different.

With this rudimentary T-method model, search engines can not only handle long sentences.

And it can also deal with those very well, the difference between one word, the meaning of the very different.

Of course, the kind of word that doesn't need to be changed, but has multiple meanings, is still more difficult to do.

For example, "liking someone" has at least four distinct meanings.