Skip to main content
  • Home
  • Library
  • Read
    • Popular
    • Charts
    • Browse
    • Categories
    • Wikis
    • Glossary
  • Request
  • Activity
  • Leaderboard
  • Help
Our Apps
  • DaoSearch
  • Narratorr
Social
  • Discord
  • Patreon
The AI Wave Tycoon's Rise
Chapter 285

Giving the Boss Work

2,270 words 8 min read 2 views 1mo ago Source ​
Skip to the end of the chapter
Prev
Next

On Tuesday morning, when Han Luyi arrived at the office, he was surprised to find Chen Jianye waiting for him at the door.

Old Chen looked like he hadn't slept well yesterday, with dark circles under his eyes, but he seemed to be in a good mood and quite relaxed.

"Boss Han, do you have a moment? I'd like to take two minutes of your time." Chen Jianye's tone was quite sincere.

"Okay, come in and talk."

Han Luyi opened the door, led Chen Jianye to sit on the sofa, and poured him a glass of water.

"Boss Han, I've come to thank you." Chen Jianye said, "President Shen told me about my child's schooling; it was you who specifically instructed me to do so."

Han Luyi smiled and said, "It's all thanks to the good policies the district has implemented. These are for key talents, and since you are a key talent, you deserve them."

In the past, Chen Jianye would have stopped there and said nothing more. A grown man keeps his gratitude in his heart and shows it through his actions. But his wife had already spoken to him yesterday, and today he was here to show his loyalty, so he forced himself to continue, even with his bald head.

"Boss Han, from now on, I, Chen Jianye, will be completely devoted to you and to Source Code Technology! I will never abandon you!" Chen Jianye stated firmly.

Han Luyi laughed when he heard this; it sounded like wedding vows.

"Okay, I believe you. Go back to work. Communicate promptly if there are any problems."

When Han Luyi saw Chen Jianye off, he brought up his panel again and took a look.

[Chen Jianye]

[Technology S | Management B | Collaboration B | Stability S]

[Special Effects: Large-Scale System Architecture | AI Agent Design]

[Entry: Cornerstone Pioneers]

Wow, once the issue of the child's schooling was resolved, the stability was immediately raised to S level.

Han Luyi that having government support really makes a difference; it significantly increases employee loyalty, especially since some resources are simply unavailable through purchase.

Unfortunately, I don't have kids yet, so I can't take advantage of these deals.

Han Luyi shook his head regretfully.

Looking closely at Chen Jianye's attributes, they have improved significantly since he joined the company last year, giving Han Luyi a sense of accomplishment, like playing a simulation game.

-Although the idea of raising Chen Jianye sounds a bit strange.

After all, appearance is a very important aspect of second-generation games. It's hard to imagine how much second-generation game players would criticize a game character dressed like Chen Jianye.

However, Chen Jianye's attribute changes prove that although the Sight panel can reflect a person's immediate attributes, it cannot predict a person's growth.

Chen Jianye's previous work habits of working alone and not listening to others' opinions were largely formed in the Dingsheng work environment, but Source Code Technology has changed a lot after a year.

When Riding the Wind was first created, he was one of the opponents, but after Su Niannian talked to him, his mindset changed quickly, and he made many forward-looking explorations in the company's AI transformation process.

Therefore, for a company, in addition to attracting talent, establishing a stable talent development system is also very important.

A company's long-term growth depends on three key elements: the ability to identify, retain, and cultivate talent.

After all, when expansion really accelerates, it's impossible to rely on external recruitment for all key personnel and management staff; otherwise, the corporate culture would quickly deteriorate.

Speaking of which, Han Luyi was actually quite interested in the project that Jiang Yixin Friday, and wondered how it was progressing.

Just then, Liu Huo knocked on the door from the outside.

"Boss Han, you have a technical meeting with President Zhao and Engineer Jiang in two minutes. Shall we go now?"

"Coming right away."

In the conference room, Han Luyi was holding a tablet and looking at the L100 Operator Adaptation evaluation report.

This was something Jiang Songran and Zhao Wenyuan worked overtime to create recently.

Since the acquisition of Ximing has been finalized, the next development path has also been determined. The two of them have put all their energy into this project, which is why they produced this report.

After reading Han Luyi didn't say anything. He put the tablet back on the table and turned his gaze to the two people sitting opposite him.

Jiang Songran with his computer on, frowning. The screen displayed a dense list of Operators and priority labels—the raw data from the report.

Zhao Wenyuan sat next to him, holding a Coconut Latte, looking even more relaxed than Han Luyi.

Today's meeting is about discussing how to Model Training in the L100 Cluster.

To be honest, Jiang Songran amazed that the reasoning could run in such a short time.

This was originally a project that would take years to complete, but Source Intelligence finished it in just one month.

Of course, Jiang Songran also put in a lot of effort in this.

He bypassed all the problems that could be avoided during the reasoning stage, which reduced the engineering difficulty and allowed Glutinous Rice Ball to be launched on time.

But now we need to do training, which means not only doing the work that wasn't done, but also making up for all the lessons we skipped during the reasoning phase.

"Let me start by outlining the general direction." Jiang Songran projected the content from his computer onto the large screen in the conference room. "We can divide this project into two phases. One is the Operators phase. After the adaptation is completed, we'll first validate it on a small cluster, allowing only small-scale training, not exceeding 7 bytes."

"The second phase involves using this system to build large-scale clusters with thousands or even tens of thousands of cards. This involves many challenges in inter-card communication, which we will discuss later."

Jiang Songran pointed to the dense Operators of arithmetic on the screen. The top dozen or so rows were marked in green, the middle dozens of rows were in yellow, and the rest were in red.

Green indicates that both forward and reverse operations were performed; yellow indicates that only the forward operation was performed, with approximate values used for the reverse operation; and red indicates that no operation was performed at all, with approximate implementations used instead.

Jiang Songran said, "The forward Operators is only the first layer. The real trouble is that training requires the backward Operators, and we also need to ensure numerical stability. Many Operators can be approximated during inference, but not during training. To achieve high precision, we have to work on it bit by bit. It's a lot of hard work."

7

He finished speaking and gave a wry smile.

If you use CUDA, Nvidia has already done the hard work for you; you only need to consider how to design the training pipeline and training strategy.

Now that we're using Ximing's L100, we can't count on their software team anymore. All the work has to be done by the Source Intelligence team. Although the people we've recently recruited have extensive experience in optimization and are ready to tackle tough challenges, these things can't be rushed.

Having a lot of work isn't scary, but you can't set unreasonable expectations.

Jiang Songran, because Han Luyi had already hinted at something at the beginning of the meeting: prioritize tackling L100.

The training ability was improved to complete the training of Glutinous Rice Ball 1.0.

Although he didn't say how long he expected to take to complete it, just listen to the name: the current release is version 0.9, and the priority is to complete training version 1.0. Can this task really wait another year?

He could only offer the most optimistic estimate. "I estimate that the first phase can be completed in three months, the second phase of communication will take six months, and another six months will be needed for Operators optimization. This is the most conservative estimate. After fifteen months, we can try to do training."

Anyone who has worked in engineering knows that such an estimate would take more than double the time to actually implement. But he simply couldn't bring himself to say that Glutinous Rice Ball 1.0 would start training three years later.

Version 0.9 has been released for three years, and they've only just started training version 1.0? Is this even human language?

Jiang Songran understood Han Luyi's decision to insist on not using any cloud service providers, as it was a commitment to data security. However, they currently only had domestically produced cards. Alternatively, if they couldn't wait any longer, they could purchase some Nvidia cards and use them temporarily. This wouldn't hinder their efforts to develop domestically produced cards, and it wouldn't be shameful.

He didn't say that, waiting for Han Luyi's response.

Unexpectedly, Han Luyi remained silent, but Zhao Wenyuan, who had been preparing the report with Jiang Songran for the past two days, spoke first.

"I think three months is too long. One month should be enough to complete the basic training process for the first phase."

Jiang Songran in shock at Zhao Wenyuan beside him.

Old Zhao, do you know what you're saying?

This is training, not reasoning!

That's not how you brag. You have to expose the truth of your boasting in front of Boss Han.

He couldn't hold back any longer and wanted to cover Old Zhao's mouth. "No, no, Boss Han, a month is definitely not enough. Look at how many Operators are in this table!"

He knew that Han Luyi was technically strong; before he joined Source Intelligence, Han Luyi had also contributed to the early Operator Adaptation work.

But he was terrified that Han Luyi might have some misunderstanding about the scope of the project.

The Source Intelligence team is currently using a working method called "Operator Targeting Method".

This working method was actually proposed by Han Luyi back then.

The so-called "Operator Targeting Method" means that for each Operators to be adapted, it is Operators against Nvidia's operators, Riding the Wind is used to convert and generate adaptation code. As long as the efficiency reaches 60% of that of Nvidia, it is considered to have met the standard.

This 60% figure is a reduction from the initial 80%, in order to allow reasoning to be deployed quickly, at the expense of sacrificing some Hardware efficiency.

However, to train Large Model now, 60% is not enough, not even 80% is enough; it needs to reach at least 90%.

That being said, is improving efficiency really an easy task?

While Riding the Wind generates code quickly, the necessary verification work is still indispensable. Every parameter of every Operators must be tried one by one to find the optimal solution for the L100 Hardware design.

If there's a 60% chance of finding it after ten tries, there's an 80% chance it will take a hundred tries, and a 90% chance it will take a thousand tries.

This is just one parameter of Operators.

That's why Jiang Songran said that doing adaptation work is hard work.

Before Han Luyi could speak, Zhao Wenyuan asked again, "Boss Han, do you think a month is feasible?"

Jiang Songran was speechless. What do you mean by "can it or not"? You're asking the boss? Of course it's possible! What if we can't produce anything in a month?

That's definitely out of the question; making even ten Operators would be beyond our expectations.

Han Luyi did not answer directly, but thought about it for a moment.

Jiang Songran hoped that Han Luyi his words to heart and realize how incredible a month was.

Reject Old Zhao, Boss Han, reject Old Zhao!

"Okay, I'll give it a try," Han Luyi said as Jiang Songran silently prayed.

Jiang Songran dejected.

Old Zhao, look, the boss agreed, right? What are we going to do?

Boss Han said he would give it a try.

Wait a minute, what did Boss Han say?

Should he give it a try?

What is he trying?

Zhao Wenyuan showed no surprise whatsoever and nodded in satisfaction.

The old man was indeed reliable.

Boss Han says it's okay, then it's okay.

Before Jiang Songran could even process what was happening, Han Luyi spoke up again, "Engineer Jiang, Operator Adaptation, you and your team only need to ensure the logic is correct. Leave the specific parameters to me; I'll find them."

Jiang Songran then realized that Old Zhao was assigning tasks to the boss!

"You--you--" Jiang Songran "you" twice, but couldn't get the words out.

What he really wanted to ask was, "What makes you think you're so good at parameter tuning?"

But it's not a very pleasant question to ask.

Zhao Wenyuan, who was standing nearby, had already explained for him, "Boss Han is particularly good at reading academic papers."

After hearing Han Luyi nodded, suppressing a laugh.

Okay, Wenyuan has found a reasonable explanation.

Jiang Songran was left standing there, bewildered in the wind. I had only heard that Boss Han had written the implementations of a few early Operators; I had no idea he was going to write the implementations of all the Operators.

Furthermore, what does it mean to entrust all parameter tuning to him?

Is this a job one person can do?

Zhao Wenyuan, seeing Jiang Songran frozen in place, gently patted him on the shoulder.

It takes time to accept some facts.

Han Luyi glanced at his phone; the meeting was almost over. He clapped his hands and said, "That's enough for today. We'll arrange the next steps after we finish the first phase of work."

Then he said to Zhao Wenyuan, "Wenyuan, I have a question for you privately. Come to my office in a bit."

End of Chapter
Prev
NovelSummaries
Next

Explore the wiki

Han Luyi
Su Niannian
Jiang Yixin
Zhao Wenyuan

Comments

No comments yet.
AboutStatsAPIChangelogStatusContactPrivacyTermsDMCA
© 2026 RandomTranslator·v3.26·RSS