Nearly thirty years old, established in science and technology

Chapter 35 Wife, do you know what a quasi-thinking model is?

"Husband, you are awesome!" As soon as Ji Hong entered the door after returning from the scene, Cheng Hui rushed over and praised him excitedly.

[Recognized, ultra-large-scale GPU scheduling professional ability +0.1]

"What's going on?" Ji Hong said while changing his shoes, "Did you watch our live broadcast?"

The signing ceremony was broadcast live by Moore Technology, and the relevant video can now be seen on the official website of Hezhou University. Seeing the starry expression in Cheng Hui's eyes, Ji Hong guessed that she had been paying attention to him. side situation.

"Yes, my husband is so handsome!" At this time, Cheng Hui was so excited.

When the cyber girlfriend became popular and made a lot of money, she knew that her husband would have a good time, and she was determined to support her husband in starting a business, but sometimes she would have some involuntary worries in her heart.

After all, starting a business is so difficult that even the second generation of rich people can’t handle it. Aren’t they saying that they are not afraid of the second generation of rich people losing their fortunes, but are they just worried that they want to prove themselves?

And my husband has never had any relevant experience, so it is completely normal for him to be a little worried.

However, it has only been a month, and the situation has developed to such an alarming point - Juaner Intelligent Technology has begun to dominate the construction of intelligent computing centers of this level and scale.

Cheng Huina wiped away her previous worries, and deeply felt her husband's strength, and recognized it from the bottom of his heart. This gave Ji Hong the ability he urgently needed now.

"But is it something I urgently need? Or is it because Cheng Hui's approval is related to this technology? Or is it because I have been studying this recently?"

Although it is a bit outrageous to get +1 thinking or +0.1 ability after being recognized, and it is impossible to get to the bottom of it for the time being, Ji Hong is also summarizing the rules from time to time and trying his best to find out some things.

In the past few days, after confirming with Zhang Jian to join the Hezhou University Thousand Card Intelligent Computing Center project, he has indeed been studying the issue of comprehensive scheduling of large-scale graphics cards.

Even without today's ability +0.1, he actually already has some relevant thoughts and ideas.

After all, AI smart scheduling of GPU stream processing core has been done before, and the computing power balancing scheduling algorithm of large-scale graphics cards has something in common with it.

Ji Hong has even made several sets of algorithm plans, waiting for implementation verification, and he was full of confidence in this.

And today I got related ability +0.1, and my thinking is even broader and clearer.

"Classic thinking model, this is the root." While Cheng Hui was busy in the kitchen, Ji Hong sat on the sofa to sort out his thoughts and discovered the root of the problem:
"Actually, if I had been thinking more about this in the past few days, I could have come up with the same plan without even using this ability +0.1!" Ji Hong thought angrily. He was already capable.

……

"Honey, do you know what a quasi-thinking model is?" At the dinner table, Ji Hong was going to tell Cheng Hui about the quasi-thinking model: "Do you remember the last time I told you about reinforcement learning in AI training? "

"Of course I remember." If it comes to skills, Cheng Hui may not really remember it, but when Ji Hong said about spanking children, she remembered part of it:

"Just like teaching children, you can set a reward rule. Give a reward if you do well, and give a slap if you don't do well. In terms of the model, it is through this kind of constant feedback and adjustment to find the maximum reward. The best choice. That’s about it, right?”

Ji Hong nodded. Although Cheng Hui was not a related professional, his understanding of the concept was still very good. The so-called reward rule is a preset reward function in reinforcement learning.

"This is the current ordinary reinforcement learning. Its learning ability is very poor, just like a poor student in school who is not enlightened. No matter how you teach him, he just can't learn.

"So it is necessary to use a large amount of data over and over again to train in a continuous cycle to achieve the desired effect. Therefore, training is very slow, time-consuming, and costly." Ji Hong thought without thinking. , try to use plain language: "Now, we have a new model here. It is different from the traditional model. How is it different?

“The neural network nodes of the traditional model are simple nodes, just like people, who have a single mind and can’t turn, they are fools.

"As for us, we replaced this node with a cluster. This cluster is an artificial intelligence micro-model. We use the micro-model to replace the original neural network node. Do you understand me?"

Cheng Hui was listening attentively with her big eyes wide open. When Ji Hong asked, she also expressed her opinion: "Artificial intelligence has certain intelligence capabilities, and this small model must also have it, right?"

"Yes, if the node of the traditional artificial intelligence neural network is a fool, our current model, just as a neural network node, already has certain decision-making and judgment capabilities. It is this ability that I Call it quasi-thinking.

"Based on this, its processing logic of things has actually undergone fundamental changes. It used to be tried over and over again, and the optimal path was obtained through statistical analysis and comparison, but now it is different. Before trying, the node itself There can be some judgment.

"This is actually a pre-qualified method, similar to qualitative thinking. This ability and method, used in artificial intelligence training, can save at least 99.99% of the amount of data in the training process. Therefore, our model training The cost is extremely low, and the effect is even better.

"Of course, it is still very different from human thinking. Its judgment is also based on the model itself, and the current training is all proprietary. It is almost impossible to make a universal model, unless a new model can be designed. Architecture hardware.

"I have an idea, which is to add a..." to the existing hardware logic.

Ji Hong became more and more excited as he spoke, forgetting to explain it in plain language. Cheng Hui had just understood it quite clearly, but now she was a little confused, but she could still vaguely sense it. The logic and profound meaning of this also make me sigh: "Husband, you are so awesome!"

[Recognized, +1 for the whole line of thinking related to model theory. 】

"Honey, are you confused?" It wasn't until he heard a voice in his ear that Ji Hong looked at Cheng Hui's confused expression and realized that he was becoming more professional as he spoke, and she was already confused.

"Well, I have to digest it. Your theory sounds quite interesting. Tell me about it next time."

Cheng Hui likes this atmosphere of mutual communication very much. Even if she doesn't fully understand what she hears, she still has some unfinished ideas. Putting aside her professional knowledge, it is impossible for her to change her career now, but it is also very useful to open up her mind. What's more, the situation in front of her is It's her husband!
"Okay! I'm going to pack some things. After you finish packing, go to bed."

……

The wooden wall between the study room and the bedroom itself in the large bedroom has been replaced by transparent glass. While sorting out the information, Ji Hong raised his head and glanced at Cheng Hui over there from time to time. This was a romance that only belonged to two people.

"There are more than one hundred ideas in the entire model theory related to quasi-thinking, among which there are twelve points that cannot be bypassed at all."

Ji Hong thought while sorting it out: "If I use these twelve points and use an AI personal assistant to make up a batch of seemingly correct but actually wrong routes, and directly apply for a large number of related patents, will the problem be solved?"

Ji Hong considered this. Previously, due to confidentiality considerations, the patent information it had compiled had not been submitted.

But now, NVIDIA has set its sights on him, and this matter is about to be put on the agenda. There is no way to guard against thieves even if you are a thief thousands of miles away.

What's more, it's not like no one has mentioned his theory before. It's just that the voice is too low and there are no examples to prove it, so it doesn't attract attention. In fact, there is no problem with the idea, but it is poor in the design of details.

"Without announcing the details, it is difficult to judge whether the route is true or false. If the route is true, but the details are not available, the conclusion will be wrong. In this case, it should be feasible."

Ji Hong nodded secretly and decided to take action immediately. With his personal assistant here, the real and fake techniques and methods could be combined very quickly.

(End of this chapter)