Nearly thirty years old, established in science and technology

Chapter 54 I think it’s nonsense!
"What do you think?"

As soon as Huang Renxun finished speaking, a voice came out: "I think it's nonsense!"

The person who spoke was Nvidia’s senior technical officer, Chris Malachowski, who went on to say: “It is theoretically impossible to use a small-parameter AI model as the core of decision-making to replace the original free neural network nodes. "

“If the original model was a ball of thread, we were running smoothly, but now it turns into a complete mess!

“Not only has the training efficiency not been improved, it has also caused great trouble. As far as my current tests are concerned, the loss during training of a million-parameter model has been as high as 70%.

"Moreover, the larger the model, the greater the loss. At the level of GPT4, the loss can be less than one thousandth.

"In other words, if you want to train GPT4 in this way, you will consume a thousand times the original computing resources! What kind of bullshit thinking model? Put together small parameter models and it will be analogous to humans? What nonsense! "

Chris sounded quite resentful. Ever since the disclosure of more than 120 patent applications for Curlear Smart Technology, Chris had not had a good night's sleep in the past few days. Not to mention quasi-thinking, he felt like he was fast. No more thinking.

However, the conclusion he reached coincides with Huang Renxun's ideas.

"But Whisper..." But as soon as Whisper was mentioned, they all fell silent again.

Because the Whispering fact exists and they have all seen the effects, there must be a plan that can do it. It is not a good choice to choose to believe Ji Hong's nonsense or continue to bump into each other like a headless chicken.

"We still have to do it! We have no choice, don't we?" Huang Renxun said.

"We are just a hardware supplier, can't these things make AI headaches go away?" Chris said:

"Aren't you in China now? Have you considered meeting the founder of Juaner Intelligent Technology? When it comes to GPUs, no company in the world can match us. Isn't this a natural collaborator?"

Chris is Portuguese and he doesn't understand the thinking of the Chinese people, but Huang Renxun understands a little bit, and he keeps shaking his head in his heart. If it were possible, he would have gone there long ago.

But when Juan Er Intelligent Technology first came to Moore Technology, he knew that Nvidia was definitely out of business. It's better to hope that OpenAI can learn from its shame and be brave enough to catch up quickly.

"But let's go and meet. Even if it doesn't work, there's nothing to lose, right?" Huang Renxun thought, he's already here, and it's not too far from Hezhou.

……

Not to mention companies in the industry like NVIDIA, even netizens who can't stand it choose to stay awake this night.

It was already late at night, and the technology circle was still very lively, almost like celebrating the New Year.

There is only one protagonist in "this year's" Spring Festival Gala, and that is the quasi-thinking model.

“I have sorted out more than a hundred patents, and I have also thought over and over what Jida said at the launch ceremony of the Intelligent Computing Center project at Hezhou University.

"Based on this, I designed an implementation plan [training a large model of quasi-thinking from scratch]. The source program was uploaded to Gitee code cloud. The address is: https:‖gitee.com｜cla***** king. If you are interested, you can download it and try it out together and discuss it together."

"I tried it, but it doesn't work at all. This efficiency is worse than building GPT directly from scratch!"

"You're talking nonsense. It's definitely impossible for a patent document to disclose details to you. This requires you to try and adjust it bit by bit to achieve the best result."

"Can it be achieved? Why do I think this is a pit? I used a mathematical model to make an evaluation. All conditions were taken into consideration and how many parameters of the small model replaced the neural network nodes. Do you know what the result is? The answer is 1 .”

The professional IT community is made up of people in the industry, especially today, the CSDNBlog artificial intelligence partition is a gathering place for a group of big guys in the AI ​​field. The answer is what 1 means, they can't understand it - that is, the efficiency of the existing neural network model is the highest, and any changes will only affect the efficiency, but not improve it.

"Post your calculation process and let's take a look together..."

"Okay!" This person's name is Sun Bo. He graduated from the University of Information Science and Technology with a bachelor's degree, and received a master's and doctorate from the National University of Science and Technology. He is now a researcher at the Artificial Intelligence Basic Frontier Theory Research Laboratory of the Institute of Automation, Academy of Sciences.

……

Ji Hong is also peeking at the screen. He is an old user of CSDN and often posts something there. After returning from Hezhou University today, he also came in to take a look as usual.

I happened to see this group of people discussing the quasi-thinking model, so I kept staring.

When Sun Bo posted all the calculation processes, he downloaded them immediately.

The calculation process is extremely complex and involves the intersection of a large number of disciplines such as differential calculus, integral calculus, linear algebra, discrete mathematics, optimization theory, and information theory.

If his ability hadn't been improving at +0.1+0.1 recently, he might not be able to understand such complicated things.

"This person is not simple!" After a long time, Ji Hong sighed. I wonder who the big boss is. Is there any chance to recruit him?

With this in mind, Ji Hong selected some purely theoretical and non-confidential things and sent him a private message: "I think the blogger is very high-level. Are you interested in this direction? We can chat if we have the opportunity."

After sending it, Ji Hong ignored him and started looking at the screen again.

Not to mention, there are a few big names who don't often bubble up in professional forums. Today, they all came out on this "Reunion Night", and Ji Hong saw several of them.

"It seems to be true. This theory doesn't work!" A blogger named "Artificial Intelligence Miscellaneous Talk" said:
"This puts me at ease. I thought Juaner Smart Technology was stupid for letting such confidential things out. It turns out it's a fake, so that's right!"

"It's not entirely false. A small model will definitely not work, but if we can solve an optimization theory problem and replace this micro model with a new thing, then..."

Sun Bo has done in-depth research on cutting-edge theory, and he always feels that the quasi-thinking model is a bit interesting, and it is related to an optimization theoretical problem that he has been thinking about. Maybe this is his success in solving that problem.

However, as he was typing, he suddenly realized that he was not discussing issues with his colleagues, but playing in a technical forum.

This is high for fun! This thing cannot be sent out, so I hurriedly deleted this text and checked my computer.

Finally he replied: "Well, it is indeed fake, but even if it is fake, wouldn't you feel itchy if you don't give it a try?"

"It's true! Regardless of whether it's true or not, the whispering system already exists. How did it come about? I won't give up until I give it a try. I'll give it a try first. Maybe although the efficiency is low, the real effect is very good. ?The results calculated by mathematics may not necessarily be absolutely consistent with the facts!"

"hey-hey……"

Ji Hong smiled, this is a conspiracy.

Who can not give it a try when they have no other ideas? Even if you know this is a trap!

I believe that Nvidia, OpenAI, especially OpenAI will definitely give it a try - it consumes a lot of their computing power resources in vain, no matter where they come from, it is a good thing!
(End of this chapter)