I am the only one who practices magic: I practice magic in the city

Chapter 222 Announced in Advance (First update of 4000 words!)

A 42% winning rate does not mean that there is a % chance of winning, but it means that you are bound to lose if your opponent does not make any mistakes.

Is alphaGo really going to lose? And it's losing with white?
You should know that in DeepMind's internal evaluation, in the modern rules where Black has a handicap of 7.5 points, it is equivalent to Black having lost a piece of territory before the game even starts, and White's winning rate is significantly higher than Black's.

In machine games, alphaGo has a winning rate of over 55% when playing with white pieces against itself.

In fact, the human Go community has had the same view in recent years, especially after Gu Jie's rise to fame.

Last year, Gu Jie lost only one game with white pieces throughout the year, and his win rate with white pieces was 97%. However, his win rate with black pieces suddenly dropped to 60%.

And now, AlphaGo is about to lose when playing with white?

And they lost without any bugs?
Can't win.

When the black chess piece's flying move in the 125th move once again split the white chess piece, even Hassabis, who was only an amateur th dan, could see that the white chess piece's disadvantage in the whole game was already very obvious.

After White's 146th move, AlphaGo's self-assessed winning rate had fallen below 12%.

But AlphaGo is still trying its best to resist, but everyone can see that there is no hope.

Its opponent is too powerful!

How can it be! ?
How could there be such a chess player?

Hassabis couldn't understand it at all.

Huang Shijie, who kept making moves with the white chess pieces, also couldn't understand it.

He is an amateur 6th dan and the actual developer of AlphaGo. No one understands AlphaGo's chess strength better than him. Even when facing a top chess player like Lee Se-sil, he can confidently say that his understanding of AI's thinking is more accurate than Lee Se-sil's.

But now, Huang Shijie completely cannot understand his opponent's chess moves, and he dare not say that he can understand AlphaGo's thinking.

Even after dozens of moves, he still couldn't understand why both sides played like this. He only knew that somehow, black seemed to have the upper hand.

Why are many of the moves I use never seen before?
Is the person opposite really a human being?

No, even AI needs to learn from human chess records. The moves have not really deviated from the scope of human thinking. It just explores human moves to the extreme.

But now, the moves used by this opponent, both in terms of layout and local patterns, are completely different from the existing routines of humans.

Unless... the opponent is really an AI, and has never learned any human tricks, and has mastered the game solely through self-playing, self-learning, and self-reasoning!
Huang Shijie was shocked by his own thoughts.

If his guess is true, it means that the other party is not only ahead of DeepMind in training, but also far behind DeepMind in algorithms!
Previously, they had imagined letting alphaGo play completely against itself to master Go, but the results were very poor.

In the experiment, they found that the policy gradient algorithm used by alphaGo had serious instability in balancing the exploration of new strategies and the use of known strategies, which made it impossible for the AI ​​to consider the advantage of winning or losing from the perspective of the overall advantage.

Moreover, the noise of Go victory and defeat signals is too strong. Without the guidance of existing knowledge at the initial stage, the model will not be able to learn effectively in the initial training stage.

In addition, it is also necessary to solve the problem of "catastrophic forgetting" of neural networks and the problem of a large number of random moves without an initial strategy.

It can be said that any of these problems may require a lot of manpower and material resources and extremely smart brains to spend several months to solve.

In particular, "catastrophic forgetting", the phenomenon of forgetting old knowledge after learning new knowledge, is very common in the field of deep learning. It can be said that the entire deep learning field has not yet found a good way to deal with catastrophic forgetting caused by various reasons.

In the end, the alphaGo team could only use human chess records to train alphaGo. When it grew to a certain level, they let it play against itself and evolve iteratively.

And now this opponent...

I wish he was human.

Although Huang Shijie did not need to consider the complex changes on the chessboard, beads of sweat still appeared on his forehead.
-
"Hey, everyone, why do you look so bad?" Sergey Brin and Larry Page had just rushed back to the monitoring room, and saw that everyone in DeepMind had a grim face, staring silently at the dozen or so monitors placed side by side on the table.

"The white pieces are in danger. It looks like the opponent is about to surrender." Sergey Brin also plays Go, and his level is pretty good, with an amateur level of one or two dan. He can tell at a glance that the white pieces are in danger.

"Hey, be happy, you made history! Guys, defeating Lee Se-sil 3-0 is not a failure of mankind, but another victory of mankind!" Sergey Brin patted David Silva's shoulder, who was closest to him.

After he said this, David Silva's face became even uglier.

The monitoring room was completely quiet.

"Sergey, alphaGo... is the white chess piece." Hassabis squeezed out a sentence with difficulty.

AlphaGo is the white chess piece? Sergey Brin was stunned and took a closer look at the screen. Sure enough, AlphaGo is the white chess piece today.

"Did the opponent find a bug?" Larry Page doesn't understand Go, but he is well aware of how powerful AlphaGo is. In previous simulated games, it was almost impossible for AlphaGo to lose unless the opponent found a bug.

"No." David Silva said stiffly, "The opponent played very strange chess today. Three of the first 30 moves were beyond AlphaGo's prediction range. At the 3rd move, although the opponent still played abnormal chess, from the perspective of the value network, AlphaGo's response was still reasonable."

David Silva frowned and glanced at the two bosses secretly: "After today's game, we will re-analyze this game. Only then will we know what went wrong with AlphaGo."

Sergey Brin scratched his messy hair. "You mean to say that the opponent's chess skills have surpassed AlphaGo? Is there such a chess player?"

After hearing what Hillgebrin said, almost everyone in Deepmind suddenly remembered the name that Fan Hui mentioned in the morning.

tangerinr.

Is this mysterious foreign aid you?
Are you human... or...

At this moment, on the screen, AlphaGo's winning rate finally dropped to 1%.

AlphaGo resigned!

After alphaGo randomly placed a chess piece in the corner, a line of text popped up on the screen.

alphaGo admits defeat!
Huang Shijie took out two white chess pieces from the chess box and placed them on the chessboard with trembling fingers.

The invincible alphaGo actually lost.

Sergey Brin saw the words "AlphaGo surrenders" on the screen and touched his beard.

Should I still go to the post-match press conference that I was originally scheduled to attend?

Sergey Brin suddenly hesitated.

For some reason, Sergey Brin suddenly remembered what the Asian young man he met in the elevator today said.

"I hope you can still be in a good mood tonight."

Sergey Brin, who originally thought that losing a game was no big deal, suddenly became really upset.

Let's wait until we really win this game before attending the press conference.
-
"Seungri Heita (Korean for victory)!!!" The Korean reporters at the scene cheered, and almost all the studios were filled with joy.

"Zalehita (well done)!!!"

"Li Jiudan! How do you feel about today's victory!?"

"I would like to ask, since you won today's game, does that mean that the previous two losses were caused by your unfamiliarity with the AI's moves?"

"Excuse me, Li Jiudan, who is the person wearing a mask sitting next to you today? He doesn't seem to be a professional chess player. Can you introduce him?" "Your chess style seems to have changed a lot compared to the past. Is this the secret weapon you came up with for AI?"

"Mr. Hassabis, what do you think about AlphaGo losing today?"

Li Shishi and Hassabis, who had a gloomy face but tried to maintain his composure, looked at each other, as if they both saw helplessness in each other's eyes.

"Today's game was very exciting. We need to go back and check carefully what problems occurred with AlphaGo today." Hassabis picked up the microphone and said dryly, then stopped talking and looked at Lee Sesil.

He had already prepared his speech after today's victory and the questions that reporters might ask.

But now that he had lost, he didn't know what to say or what not to say.

How to face the media is a professional skill that also requires training. Unfortunately, Hassabis has not yet truly mastered this skill.

Even a genius is not good at everything.

In comparison, Lee Se-sil, who has been facing the media since childhood, is much more at ease.

"I'm very sorry, but I must first state that today's chess talk is not based on my personal strength."

"According to the agreement reached between me and the Goryeo Chess Institute, since I lost the first two games, and the second game was a complete defeat, I do not have the initiative in this game. My role today is the same as Dr. Huang's, and I am no less surprised than everyone else about this game," Lee Se-sil said softly.

As soon as Li Shishi said this, the media at the scene was in an uproar.

No one expected that such a proud Lee Se-sil would actually give up the right to play the main game.

What happened? And which chess player could be so strong that the fearless Lee Se-sil would willingly give up the mastery of the game?

"As for my partner, I haven't met him and he probably wasn't at the scene either."

"The gentleman next to me only informed me of the placement based on the information transmitted by my partner in real time. As for the name of my partner, I believe many chess fans know him and have guessed that he is the one who completed the game for me today. He is the mysterious chess player, Orange, who defeated Gu Jie and me online two months ago."

tangerine!!!

This name, which sounds like it's following Apple's lead, has officially appeared in the mass media.

Before Orange defeated Gu Jie and Lee Se-sil, only a small number of Go enthusiasts knew about it. Although Gu Jie also said something when he won the Menglily Championship, his influence was obviously not comparable to what it is now.

Now we are facing the global media!
It is conceivable that within two hours, this name will spread throughout the world and occupy countless hot searches!
Li Shishi looked indifferent.

"As for Orange's identity, you can ask Mr. Hua next to me later." Li Shishi pointed at Hua Zecheng who was hiding in an inconspicuous corner.

The flash lights were clicking as Hua Zecheng, who was still wearing a mask, kept taking pictures.

"Excuse me, who is Orange? Since he is so powerful, why doesn't he challenge AlphaGo himself?"

"Some people say that Orange is artificial intelligence. Is that true?"

“How do you feel about this victory?”

……

Hua Zecheng was a little confused by the unexpected questions from reporters.

A voice suddenly came from the headphones on his head.

Hua Zecheng listened attentively and couldn't help showing a surprised expression. He glanced at Fang Yu who was still sitting on the chair, and saw Fang Yu nodded to him.

"Yes, Orange is indeed artificial intelligence." Hua Zecheng took off his mask, picked up the microphone, and replied concisely, "Orange was developed by Yuzi Technology. It is the world's first truly large-scale artificial intelligence model."

Hassabis was shocked.

Yuzu Technology?

That Yuzu Technology that Google wanted to acquire? Did they develop Orange?

What is the big model?
Damn!

In an instant, Hassabis figured everything out.

Damn! This was premeditated!
deepmind was sniped!

Hassabis's face turned pale instantly.

Angry.

I should have known not to sign that pre-match agreement.

So careless.
-
"What do you think? I'll tell you!" In the studio, Gu Jie jumped up excitedly and punched the air several times, like a 1.8-meter-tall monkey.

He has been scolded a lot these days, and he just doesn't understand why so many people can't see such an obvious thing?
I told you to scold me! ?

Are you embarrassed?

"This, this..." The other three hosts in the studio were dumbfounded.

"Uh, Professor Liu. In the interview just now, it was said that Orange AI was developed by a company called Yuzu Technology. Do you know anything about this company?" Chen Ying quickly changed the subject and handed Liu Ziqing a handle.

"Ah?" Liu Ziqing was still confused. Aren't you just bringing up something that is irrelevant?

How the hell would I know what Yuzi Technology does?
-
"Yuzu Technology!"

"Bang!"

Wu Ende pounded his fist on the table in annoyance, and without thinking about anything else, he picked up the phone and called Robert.

"Andrew, I know what you are going to say. I will give you an explanation." Before Wu Ende could say anything, he heard Robert's gentle but firm voice on the other end of the phone, "Qianxun, you must go all in on AI now!"
-
"Didn't you say that the results would be announced after the final battle?" Shen Shuyao pulled Fang Yu and asked quietly with her peach blossom eyes flashing.

Fang Yu shook his head: "No, I just took a look at the atmosphere at the scene, too much is as bad as too little."

"Originally, I wanted to intervene after Lee Se-sil was defeated 3-0, but I didn't expect Lee Se-sil to lose the will to fight so quickly and announce the Xiangtanqi in advance."

"If there were only two games, and they were played after the winners and losers were decided, it would be fine to create some hype. But now, Orange has intervened in the third game, giving everyone hope that humans will win. If we keep it hidden and announce it after the last game, people will criticize it as a gimmick to fool the audience, which will be difficult to deal with."

"So, we can only announce it now and then challenge AlphaGo head-on!"

After saying that, Fang Yu used the Wilderness in his sleeve to perform a fourth-ring "sweet talk" on Hua Zecheng.

Note 1: The current large model LLM refers to the large language model, which mainly uses the transformer architecture.

The concept of LLM only officially appeared in 2017.

However, the Orange model in the book is different from the Transformer-based large language model that mainly relies on the attention mechanism, such as GPT. Instead, it is built based on the underlying principles of Pomelo, and is built by Pomelo's newly developed corresponding neural network algorithm. Therefore, the large model here does not refer to a large language model, but a large parameter model.



(End of this chapter)