After the layoffs, I became a tech godfather

Chapter 16 Silence is Songshan Lake tonight

He Gang recounted his experiences and observations of the day word for word without making any mistakes, while Yu Dongshir frowned and bit his jaw from time to time, but in the end, he fell into a rare silence.

"This..." At this moment, Yu Dong couldn't even find a suitable word to describe what was in his heart. He just had a thought: "No, I must go back immediately!"

As soon as he got back to the hotel, before he even had time to sit down, he immediately packed his luggage, then rushed straight to the Capital Airport and flew back to Pengcheng on the night flight.

He didn't even go home, but went straight to the company. On the way there, he had already called the team together for a meeting.

"We have no intention of prying into the secrets of our partners, but the facts are before our eyes. They have already made their products and their capabilities are so strong. This is an objective fact."

Yu Dong came up with the conclusion without saying anything: "This can't possibly be trained by accumulating massive amounts of data and using large-scale computing power. There must be a new method."

The reason he said this is because he has absolute confidence. If we really compare computing power, who in China can compare with Huawei?
What's more, Yu Dong already knew clearly that Hao Cheng did not use any particularly high-performance hardware to train these things - there was not even a data center or a computing power center there, and the deployed servers were all rented from Alibaba Cloud - there was no way to use them even if he wanted to!

This time, before He Gang went to Linzhou, he specifically asked Hao Cheng to give the project of "color imaging device based on mobile phone multi-spectral lens" combined with AI to conduct food safety testing to Hao Cheng, which was a test in itself.

After all, there is no possible connection between AI capabilities that surpass all language models in the world and individual developers.

It is normal for him to have doubts.

It just so happens that the AI model they are currently training has an accuracy rate of only about 60. Although it has some effect, it is not much better than blind guessing and does not meet the conditions for commercial use on the market.

If Hao Cheng did it...

But now it seems that even if he can't do it...

Just as He Gang said, what we already have is powerful enough:
"If training is not achieved using massive amounts of data, then it must be based on known laws, theorems, or even common knowledge. There is no doubt about that."

Silence, silence is Songshan Lake tonight.

Everyone held their breath and remained silent. They knew from the beginning that knowledge and rules must be more powerful, more accurate and faster materials for training AI than data.

However, it is absolutely useless to have only materials but no methods.

They have also proposed using knowledge to drive intelligent networks to achieve continuous evolution of intelligence towards a higher level.

However, AI itself does not have the ability to understand, and its essence is imitation. Therefore, direct training is not feasible, and to this day, they still have no idea how to proceed.

Various related work has also stagnated to some extent, and all have turned to supporting AI training with big data and high computing power.

I originally thought that this road was completely impassable.

but……

Today, there are already people at the finish line!

Even if you don’t know how others get there, at least you can be sure that there is a way!
"Make every effort to deploy personnel and funds, spare no expense, and speed up the progress of Plan B!" Yu Dong quickly made such an arrangement.

...At the same time, Hao Cheng was not idle either. After finishing the negotiations with He Gang and returning home, he had been calculating the costs on his side.

According to the test results of the first batch of users trying out the application, the maximum concurrency that a server with more than 3,000 users per month can meet is 1,500.

It was originally thought that the 15,000 would be enough to satisfy 10,000 registered users. After all, it is impossible for all users to use it at the same time.

But the test result surprised Hao Cheng - there were 10,000 user registration slots available, and now the server was obviously overwhelmed.

The response speed has been greatly affected, and the user experience has been obviously affected.

"User stickiness is too high, and the ratio of server computing power to users can only be configured at a peak of 1:2 in the early stage." Hao Cheng said with emotion: "In this case, the server cost alone is one yuan per user per month."

If other Internet companies heard Hao Cheng say this, they would probably curse him and wish him to be struck by lightning - companies have always been trying to retain users, and it is outrageous that they are still complaining that user stickiness is too high.

Of course, Hao Cheng just said that. The purpose of his calculation was to calculate how many users a computer room with a computing power of 10,000 TFLOPS can satisfy.

Ten computing power units mean one thousand concurrent connections, and ten thousand computing power units can be used by one million users at the same time. If calculated at a ratio of 100:1, it can accommodate approximately two million registered users.

This number is still considerable. The key lies in how to consider the operational strategy.

Hao Cheng is a technician and has never considered this kind of problem before, but now it is no longer possible. He has to make all the decisions. Fortunately, he has Xiaosha's assistance.

"Advertising can be ruled out first. Not only is it unstable, but it also involves risks." Hao Cheng dislikes this approach from the bottom of his heart.

Moreover, for a product with such high user stickiness, it would be most appropriate to set up a reasonable membership system, right?
Hao Cheng and Xiao Sha worked together, each with their own ideas, and soon came up with a preliminary plan, which was roughly as follows:

1. Basic service: This is directly free, and the purpose is to increase the number of users. The number of users is almost positively correlated with the energy value. This thing is very important to Hao Cheng now, and it is almost the most important one.

This version includes almost all the main functions, and its main form is conversation-based AI question-answering, and its upper limit is the capacity of the current large ordinary language model.

To be more specific, referring to the capabilities of the 59.9 model that costs 4.0 yuan for Wenxin Yiyan, if it were 20% stronger, it should be about the same.

16. Advanced service: 1999 yuan/month for individual users, yuan/month for enterprise use per terminal.

This version fully unlocks all of Xiaosha's capabilities except for the professional enhancement. Not only is the ability based on AI question and answer enhanced, but it also adds many functions such as "help me do it", "follow me", and "you do it".

Just as Hao Cheng negotiated with He Gang, he can be treated as a human being.

3. Additional enhancement of professional capabilities.

As for professional capabilities, Hao Cheng thought about it again and again and decided not to develop other apps again, but to integrate them all into this one app and integrate them into the Xiaosha app in the form of "professional capability enhancement".

As for the cost of integration, like the advanced services, each item is 16 yuan/month for personal use and 1999 yuan per terminal per month for corporate use.

Of course, now there is only one professional ability available: "Legal Services".

(End of this chapter)