Reinventing the Millennium

Yi Ke's "Taibai" was born and broke through the human Go camp at a lightning speed, but even in 1V5, even among the top human players, many people are still dissatisfied with Yi Ke's claim that AI has surpassed human intelligence in the field of Go.

There are many reasonable reasons for this, such as AI appeared too suddenly, and no one had time to carefully study its chess moves; for example, the time was too close, and the players had not fully adjusted to the best state...

However, it is too late to regret for the human players now, and the "Alpha" from Google has arrived on the battlefield.

After AI's unparalleled victory over humans, Yi Ke quickly determined the time for the technical exchange with Google, which was directly set on July 3, and the whole process was still live broadcast.

Fang Zhuo does not understand Go, but he is very interested in the verification and exchange of this technology.

This time is not just a game between two AIs, Yi Ke also held an exchange event with Google's Deep Learning in Shanghai, and NVIDIA and Silicon Valley's recent R\u0026D personnel and startups in this field also participated.

Yike is really willing to learn from each other, and Google and Silicon Valley don't have much intention to compete with each other, because there is really no hope of breakthrough in this track, and it is still in the accumulation stage.

It is precisely based on this situation that many uninvited developers and companies have seen the interaction between the two companies and actively signed up for the event. After discussing with Google, Yike simply expanded the scale and postponed the time to July 10.

This is not an AI interaction of Go, but a seminar on deep learning DL with it as an appetizer.

On July 10, Yike, Google, NVIDIA and other companies held a meeting at the Yike Center in Shanghai. At the same time, Yike also invited Ke Jie, Lee Sedol, Shin Jin-seo and others who were defeated before as guests to witness the duel between "Taibai" and "Alpha".

These two AIs have a record of defeating top human chess players. Taibai's multi-faceted play and public opinion hype last month have successfully made the public interested in AI. In addition, there is a live broadcast, so there are many viewers.

At nine o'clock in the morning, two robots sat in the center of the stage. A large screen was set up on the scene. Fang Zhuo, Larry and others sat in the first row. Chess players Ke Jie, Shin Jin-seo and other guests commented in the live broadcast room.

Unlike the last live broadcast of 1V5, "Taibai" and "Alpha" made moves very quickly. There were occasional hesitations, but these moments were extremely short compared to humans.

The game was a traditional rule, with three hours for each player. However, in just 32 minutes, this highly anticipated game ended with "Taibai" winning 1 piece.

These 32 minutes were 32 minutes that most people could not understand. However, they could see the performance of the chess players and the process of disagreements in the live broadcast room at the beginning, disputes in the middle, and finally silence.

When Ke Jie was invited to the stage to comment on the game as a professional, he was very confused in front of the camera. After a while, he said: "AI may have completely exceeded human imagination in Go. Last month, I seemed to see the shadows of Gu Li, Lee Chang-ho, and Wu Qingyuan in AI's chess, and I seemed to see my own shadow."

Ke Jie's expression was full of struggle and pain, and he covered his face with his hands: "Today, I can't see it, I can't see it at all. I don't seem to understand, I don't understand what Go should be like..."

The host saw that Ke Jie had lost his composure, so he quickly asked this top human chess player to step down and smooth things over, but this scene undoubtedly left a deep impression on those watching the live broadcast.

After the appetizer of the AI ​​game ended, Fang Zhuo took the microphone and made a brief speech.

"Go is a masterpiece of human wisdom, but so is AI."

"My expectation for AI is that it can greatly liberate human hands. This day may be far away, but just like today's 'Taibai' has made progress compared to last month."

"AI will evolve at an amazing iterative speed, and we are gathered here today to find the right direction of development."

"AI has won over humans in the field of Go, but this is not the end of human wisdom, but an extension of human wisdom, another advancement of science and technology, and another exploration of the future."

Fang Zhuo's expression of this view still won a lot of applause and praise from the live broadcast room.

For many people, this excitement is only here, but for practitioners and R\u0026D personnel, the real part has just begun. Both Yike and Google have deep research in the field of deep learning DL. This kind of Go game is just the appearance, and the internal operation and thinking are more important.

As one of the leaders of Yike's "Venus" project, Andrew Ng communicated with Silva of Google's company on the model logic of DL.

Whether it is Taibai or Alpha, they are all based on the development of convolutional neural networks. The foundation is similar, and its breakthrough originated from the paper on AlexNet deep convolutional neural network published by Alex, Ilya and Hinton in 2012. It was after that that related research experienced explosive growth.

What Ng Enda and Silva talked about was the architectural innovation after AlexNet, which is the effective integration of traditional search algorithms and deep learning models, and what the entire team has done in terms of local receptive fields, parameter sharing and sparse connections, and translation invariance. effort.

This kind of communication between Yike, Google and the guests was extremely pleasant, and it also made Fang Zhuo quite satisfied. Although he didn't understand, looking at such a scene, he felt that knowledge was stuffed into his brain.

However, the next day, when Andrew Ng brought up the team's confusion in research and development, a heated debate ensued.

Yike has voice assistants like "Siri" as a practice for artificial intelligence, and Andrew Ng's team is not only doing research on convolutional neural networks (CNN), but also on recurrent neural networks (RNN). They believe that the latter is more suitable for use with voice. The combination of assistants, but the effect is not very good and cannot achieve the desired results at all.

where is the problem?

Ng Enda expressed his confusion and also talked about Yike’s internal solution direction.

Some people attending the meeting agreed with Yike's problem-solving ideas, but there were different voices from Google.

"Why do we have to use recurrent neural networks?" Google's Uskert was on vacation, but he signed up because he was interested in DL exchanges. "Why not try Self-attention? I think it is very useful for NLP." There will be better changes in the field.”

“Self-attention can perform better parallel computing capabilities instead of sequential processing like RNN. It can also directly compare the vector representations of any two positions in the sequence, so that it can more effectively capture and utilize long-distance dependencies. , but not RNN!”

“Although RNN can theoretically capture long-distance dependencies, in practice it is often difficult to achieve due to vanishing or exploding gradient problems!”

Uskert is studying Google's machine translation improvement methods. His father is a professor of computational linguistics. Although he disliked the work of language translation when he first joined Google, he eventually focused on research in this field. , and what he is currently thinking about is the improvement of "Self-attention" in related fields.

Andrew Ng quickly understood what the Google researcher meant, and after some thought, he rebutted: "Self-attention does not explicitly encode positional information, which means that if the model with it as its core cannot distinguish the same words in different sequences in different The meaning of position is different, and in natural language processing, the semantics of words are closely related to position.”

“Moreover, the self-attention model must have a huge number of parameters due to the attention weight calculated for each pair of elements in the sequence, which is likely to lead to overfitting.”

Just as he spoke, Silva from Google's own DL also refuted the new route proposed by Uskert. One of the important reasons is that the cyclic structure of RNN is too consistent with everyone's understanding of sequence data processing, that is, the current state depends on Past information, and the global dependence of self-attention is not as intuitive as RNN at first glance.

The two leaders of Yike and Google both criticized self-attention, but Uskert was not convinced and went directly to the stage to explain more of his ideas.

Moreover, some solutions to the shortcomings criticized by Andrew Ng and Silva were also given, such as introducing positional coding, for example, conducting research on multi-head attention.

Some people find it bright, some find it whimsical, and some people perform quick analysis and calculations on the spot.

Fang Zhuo, who was in the first row, was extremely confused. He turned around and asked Jensen Huang, the head of Nvidia who was thinking deeply next to him: "What are they discussing?"

"Uskert said that GPU is the most suitable hardware for deep learning technology." Huang Renxun gave a summary.

Fang Zhuo: "???"

He wondered: "How come I haven't heard similar expressions at all?"

"Because Self-Attention puts more emphasis on parallel processing, this is what GPUs are better at." Huang Renxun said with a smile, "As for the other things, it's not important. We just need to extract what is beneficial to us."

Fang Zhuo observed the atmosphere at the scene. This was no longer an exchange of knowledge, but more like a battle of knowledge.

He silently put away his little presence on such occasions.

However, in the evening, Fang Zhuo still asked Andrew Ng, who participated in the debate, face to face, wanting to know what this red-eared discussion was about.

It was really difficult for Ng Enda to explain to Mr. Fang what happened.

"Mr. Fang, let's think about it and write a report later." He said sincerely after thinking for a long time, "Perhaps we should have more exchanges like this. His self-attention, I think about it now, is not a shortcoming. Can’t be solved.”

Fang Zhuo asked patiently: "Then what should we do?"

Ng recalled today's entire debate, thinking about the different ideas put forward by different people, and murmured: "Maybe we need to try a new type of software first so that it can talk to the computer."

Fang Zhuo said "hmm" and expressed his consistent support: "Okay, let's get started."

Ng Enda didn't know whether to laugh or cry, and said sincerely: "Mr. Fang, the support you give us is really generous."

"I can only give what I can." Fang Zhuo said with a smile, "Let's try it. If there are not enough people, then we will recruit more people from home. I have a simple understanding. Since it can cause controversy, There is merit in new things.”

Ng gave a thumbs up and liked the support from his boss.

However, Yike's current situation is not so wonderful.

He raised a small concern in this regard.

Fang Zhuo said nonchalantly: "How much money do you guys invest? It just so happens that a batch of NVIDIA GPUs are about to arrive. Let's try them out first. I see Huang Renxun is smiling very happily today."

If Ng wasn't good at business matters, he would have gone shirtless to help Yike sell goods.

However, although he cannot sell goods, Yike's exchanges with Google and Silicon Valley companies have directly further stabilized the company's stock price.

Yike seems to have solid research and development. Although it is unknown when it will have commercial value, AI defeating humans in the field of Go is undoubtedly a real breakthrough.

Yike's stock price has fallen badly enough, isn't it enough now?

In response to this voice in the media, Russell of RC Hedge Fund was quite angry. He and a group of short sellers were also paying attention to Yike's recent exchanges with Google and other companies. However, they did not recognize the impact of this field on Yike's stock price. promote.

Not only that, Russell also invited Conborn, a senior executive of the American manufacturer DuPont, to let professionals judge the current situation of Yike and Bingxin.

"The photoresist in the ice core must have been used up! This is determined by the material itself, and the ice core cannot violate the laws of reality!"

"China's photoresist is very backward, and it will be impossible to provide usable raw materials for ice cores for at least five years!"

Senior executives from DuPont swore in front of the camera to give their conclusions.

This professional judgment in the monopoly field has indeed caused some people to cast doubt on the future of Yike and Bingxin. As time enters July, the market performance of Yike mobile phones seems to have verified the possibility of the discontinuation of the cutting-edge technology of Bingxin.

However, Russell's private communication with Conborn was slightly shortened in terms of time limit.

"It can take as short as two years or as long as three years, and Ice Core will have usable photoresist." Kang Boen still swears by his judgment, "You can't overestimate the research and development of Yike and Ice Core, but you can't overestimate it either. Too underestimated.”

He doubled his expectations on TV. That was a genuine exaggeration, but now he is also sincerely helping Russell analyze the situation.

"Two or three years is enough, completely enough." Russell was very satisfied with the stimulation such action brought to Yike's stock price. "This is enough to cause Yike to suffer an irreversible decline."

He commented on Yike: "Kang Boen, Yike has a shortcoming, do you know what it is?"

Cornborn listened to the short seller's advice: "What is it?"

Russell smiled and said: "Yike has only one shortcoming, it is not strong enough."

Kang Boen has reservations. There will be no comprehensive company in the world. Yike is already strong, but he also respects the short seller who is attacking Yike's stock.

Moreover, he also shared his exchange with this person on Twitter.

——Yike has only one shortcoming, it is not strong enough.

On July 19, Yike, which is not strong enough, released its Q2 financial report, stating to the market the current financial situation of the company and the changes during this period.

Throughout the second quarter, Yike's revenue dropped again from Q1, from US$29.344 billion to US$18.14 billion, a rate of 38.18%, while key data such as operating profit and net profit also experienced a more consistent decline.

Generally speaking, the second quarter is the off-season, and Yike's second quarter figures in previous years have been lower than those in the first quarter. However, leaving aside the normal impact, this 38.18% decline is still very ugly.

Although there are different voices on Wall Street regarding Yike's Q2 financial report, they believe that "advertising revenue has increased significantly", "technology licensing fees have increased significantly" and "cloud business has grown unexpectedly", which together support Yike's transformation, but , the market value of Yike, which has been stable at US$280 billion recently, has plummeted.

As of July 26, Yike’s market value fell to US$251.3 billion within a week.

Yike is known as the two poles of YA in China, and with its decline, it is very close to Alibaba's current market value of US$228.9 billion, only one and a half degrees difference between the two.

RC Hedge Fund's Russell immediately laughed at Yike's decline.

"Hahaha, Mr. Fang has been working for these years, and the market value of the company is only that of Alibaba!"

"The concept is only temporary. AI has not made any substantial breakthrough. Its popularity will soon pass. Yike is returning to where it should be!"

"I have already said that Yike has no magic power, and Mr. Fang has no magic power anymore!"

"He's messed up everything! Yike has been messed up by him!" (End of Chapter)