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 185

Didn't This Mean There Was a Chance?

2,169 words 7 min read 1 view 1mo ago Source ​
Skip to the end of the chapter
Prev
Next

Monday, January 11th.

At nine o'clock in the morning, Han Luyi walked into the office of Source Code Technology on the twelfth floor of the Qiantan Center.

The weekend had just ended, and the office had already returned to its weekday rhythm. A few engineers who had arrived early were scattered around the open workstation area, discussing things in hushed tones while looking at their screens. Someone was carrying coffee back from the break room.

They all stopped to greet Han Luyi when they saw him.

No one was surprised that Han Luyi had been away from the company for several days and then suddenly returned; everyone gradually got used to the big boss's unpredictable appearances and disappearances.

Han Luyi didn't go to his own office, but went to Su Niannian's office first.

Su Niannian's computer screen was displaying the review process data sent by Chen Jianye. She saw Han Luyi approach, nodded, and then turned her attention back to the screen.

"You're back."

Han Luyi: "I'm exhausted. I slept for two days over the weekend to recharge. Did things go smoothly while I was away?"

Hearing him say that, Su Niannian looked up and carefully observed his expression before feeling relieved and said:

"You should take a look at the Agent programming project that Xiao Jiang did, the one I told you about last time."

"It works very well?" Han Luyi eyebrow.

Su Niannian smiled and said, "I'll save this for you. Ask Xiao Jiang when she arrives; she prepared a report in advance."

Han Luyi went outside and saw that Jiang Yixin was already at her workstation.

Han Luyi walked over, stood next to Jiang Yixin's workstation, and gently tapped the partition.

"Boss Han, you're back!" Jiang Yixin turned around and was startled.

"President Su said your Agent is very effective?"

"Qian Xiaole and I did it together," Jiang Yixin emphasized as she pulled up the data. "Boss Han, look, this is last week's data."

"The company has a total of 30 frontline developers, and last week merging 127 code branches."

Jiang Yixin pointed to the bar chart on the right.

"The previous data was an average of forty-seven per week."

She then pointed to the bar chart on the left.

"But this number hasn't peaked yet, because everyone is adjusting."

Jiang Yixin added another filtering condition to the data.

"Look, Qian Xiaole, she was the first to use it. Last Tuesday alone, she merged ten code branches, and these are all production projects."

Upon Han Luyi leaned forward and placed his right hand on the table in front of Jiang Yixin.

Having worked as a programmer for over five years, he knows what ten merges a day means.

A code merge involves a series of steps, from understanding the task requirements, communicating with other departments, synchronizing information, making changes, and writing unit tests. The changes vary in size, but on average, one merge per person per day is a reasonable workload.

Qian Xiaole is now doing the work of ten people all by herself.

For many years, Silicon Valley has been promoting the concept of "ten-times engineers," which refers to the idea that a single, highly talented programmer can make the contributions of ten ordinary programmers.

At least for this day, Qian Xiaole is already close to becoming a "ten-times engineer".

More importantly, this ability is not a talent.

It's the tools that bring it!

As long as the process continues to be optimized, the output of the thirty developers at Source Code Technology will be raised to a new level.

"What about the code quality? Show me the merged changes," Han Luyi asked.

Jiang Yixin was prepared. "The pass rate for the first submission by the Agent is currently 67%. About one-third of the tasks cannot be completed independently by the Agent yet; they will get stuck, and manual intervention is required at this time."

Then she opened the code repository and found the changes Qian Xiaole had recently merged.

Han Luyi opened his Sight and scanned the surroundings, almost everything was green.

This is an incredibly significant increase in efficiency.

Thirty people did the work of a hundred.

Last month, Han Luyi was still planning that if he wanted to expand the development team from thirty to sixty people, it would take at least three months just for recruitment, interviews, and onboarding training. And in the market of Haicheng, the monthly salary of qualified engineers he could recruit would be no less than 30,000 yuan.

Thirty people, that's over ten million in human resource costs per year.

But if we don't hire more people, Source Code business is growing too fast, we can't keep up with the demands.

The thing that Jiang Yixin and Qian Xiaole created means that he doesn't have to spend a single extra penny to have seventy more people out of thin air.

Wrong.

It's better than having seventy more people.

Adding seventy people means seventy more management costs, communication costs, and an adjustment period.

Agent do not need to hold meetings, team building activities, or one-on-one communication, which greatly reduces management costs.

Han Luyi looked at Jiang Yixin and said seriously, "Xiao Jiang, prioritize this tool the highest. If you need any resources, just ask President Su directly."

This was the first time Jiang Yixin had ever heard Boss Han speak in this tone.

"Yes, Boss Han!" Jiang Yixin, then added:

"By the way, we also have to thank Mr. Chen for his newly designed review process. We've recently been developing a review tool based on Large Model, which, in addition to using BugKiller for bug detection, also includes suggestions on code style and maintainability."

Upon hearing the words "General Manager Chen," Han Luyi was momentarily confused, thinking it was Chen Bowen.

Then he realized that Jiang Yixin was referring to Chen Jianye.

Oh, by the way, he got promoted to technical lead.

Han Luyi praised Jiang Yixin a couple more times, and only left when Xiao Jiang was embarrassed by the praise.

He also needed to go to the model group on the thirteenth floor to find Zhao Wenyuan.

"Wenyuan, what exactly do you mean by 'not going smoothly'?" Han Luyi asked.

"Boss Han, I'm afraid domestic compatibility is not feasible," Zhao Wenyuan said with a bitter face.

"Is it technically difficult?" Han Luyi eyebrow, already preparing to open his Sight himself.

"No," Zhao Wenyuan sighed, "The workload is too large."

"CUDA has been building its ecosystem for over a decade. You're asking me to lead this small model team? Let alone adapting to the ecosystem, we can't even handle the migration of an Operators."

Han Luyi wasn't surprised.

If building an ecosystem were easy, domestic Hardware manufacturers would have done it long ago, and these problems wouldn't exist.

"Let's look at an example," Han Luyi said.

Zhao Wenyuan felt that Han Luyi were somewhat unnecessary.

Even if you're a great writer, I can't expect you to write them all down one by one.

Besides, haven't you already added annotations?

Zhao Wenyuan didn't say these words aloud; instead, he obediently found an Operators's code in the codebase as an example.

scaled_dot_product_attention

This is a crucial Operators in the Transformer architecture; it's fair to say that without it, Large Model would be impossible to create.

"Nvidia has dedicated functions that are deeply optimized for performance and accuracy; I don't even have an equivalent implementation on hand."

Han Luyi pulled up a chair and sat in front of the computer. He took the mouse, opened the browser, and opened the relevant CUDA source code, the IR document of the domestic graphics card, and the HCCL SDK.

Zhao Wenyuan, watching from the side, was stunned:

"Boss Han, what are you going to do? -- You're not going to write it yourself, are you?"

Han Luyi didn't even look up. "Let's give it a try."

"Give it a try?" Zhao Wenyuan thought to himself. "Boss Han, this is several months' worth of work for an entire team."

Han Luyi has broadened his Sight, identified several key aspects of the CUDA implementation, and summarized the key points for adapting to domestic graphics cards.

Then Han Luyi opened Jiang Yixin's AI Agent programming tool on Zhao Wenyuan's computer.

But he didn't input all the key information he saw from the Sight.

He wanted to see first what the model could achieve on its own.

He entered the first suggestion word:

"Translate this CUDA Operators into a domestically developed IR implementation. The required accuracy error is less than 1e-5, and the performance is no less than 70% of the NVIDIA implementation. The following three documents serve as context."

Then I typed in all the link addresses from the browser.

Soon the AI Agent began to break down tasks, solve them, and then summarize the results.

Three minutes later, the first set of results came out. The performance was close to 68%, but the accuracy deviation was too large.

A large red FAIL appeared on the screen.

Zhao Wenyuan breathed a sigh of relief.

That's normal.

He said to Han Luyi, "Boss Han, look, this is the difficulty I was talking about; we can't handle it—"

Han Luyi did not respond to Zhao Wenyuan.

He reopened the CUDA source code and opened the Sight.

When ordinary people look at code, they see characters. When Zhao Wenyuan looks at code, he sees logic.

But Han Luyi saw something else Sight: not just what the code was doing, but why the code did it that way.

The trade-offs behind every design choice appear like annotations next to the code.

Why doesn't softmax use the most intuitive implementation, but instead breaks it down into three stages? Because the intuitive implementation will result in numerical overflow on long sequences.

Why are the matrix multiplication blocks of this size, neither too big nor too small? Because if they were larger, they wouldn't fit in shared memory, and if they were smaller, they would cause memory conflicts.

These things aren't written in any documentation. They are the accumulated experience of Nvidia engineers after countless experiments, hidden in the structure of the code, and can only be deciphered by those who truly understand the Hardware.

Zhao Wenyuan wasn't incapable of understanding code; he simply couldn't extract years of engineering experience from others within a few days.

But Sight is possible, and so is Han Luyi.

Han Luyi turned off the first version of the prompts and re-entered the information.

This time, he did not allow the Agent to act freely.

Instead, it directly inputs what it sees from its Sight.

"Softmax must use a three-stage online algorithm; do not use naive softmax. The current accuracy problem lies in the second stage, reduce. Local maxima, exponents, and update order must be consistent."

"Matrix multiplication uses 64x64 tiles. If the tiles are too large, there won't be enough shared memory; if they are too small, it will increase bank conflicts."

"During the reduce phase, expand using a 4-stride approach to avoid bank conflicts."

The key-value matrix is cached in shared memory by row-major to avoid consecutive collisions across banks.

"Prioritize accuracy, then optimize performance."

Press Enter.

The Agent has started working diligently again.

Five minutes later, a large green PASS appeared on the screen.

Zhao Wenyuan's eyes nearly popped out of their sockets.

"No... what's going on?"

Ignoring Han Luyi still sitting in front of the computer, he leaned close to the screen and read through the test report from beginning to end.

Accuracy error: 2.3e-6, far below the requirement of 1e-5.

Performance: 83% of what Nvidia cards achieve.

It's not 70%, it's 83%.

Zhao Wenyuan pulled out the generated code again and looked at it line by line.

He grew increasingly silent as he looked at it.

This code is not a crude implementation that's just "good enough to run": the softmax uses a three-stage online algorithm, the reduce strategy is clean and efficient, and the use of shared memory is almost entirely efficient.

This is something only someone with a deep understanding of the underlying Hardware could write.

No, to be precise, only someone with a deep understanding of the underlying Hardware can guide the AI in writing things.

Zhao Wenyuan turned his head and looked at Han Luyi.

"Boss Han, how did you know those suggestion words you entered the second time—softmax three-stage, tile 64x64, 4-stride expansion—?"

Han Luyi leaned back in his chair. "I've looked at the document."

"Wow, so you read the documentation, huh? Why didn't you say so earlier?" Zhao Wenyuan joked first, then his voice suddenly rose, "I've been reading the documentation for two days and running over a dozen tests, and I still couldn't find this tile size. You figured it out in just a few minutes?"

Han Luyi didn't answer, he just smiled.

Zhao Wenyuan stared at him for a long while, then leaned back in his chair as if deflated.

"Alright," he said, "I won't ask anymore."

He had considered using AI for these tasks before, but AI simply couldn't do it. Every time it ran, the results either froze or were too far off.

How come Han Luyi works so well right away?

Zhao Wenyuan just wanted to quickly delete the title of the shared document he had sent to Han Luyi.

Han Luyi did something right before his eyes that he thought was impossible.

What if this wasn't a coincidence?

If scaled_dot_product_attention can do this, what about other Operators?

What do you mean it's not feasible at the moment?

What does "Forget it, it's hopeless" mean?

Doesn't this mean there's hope?

He really wants that former colleague to come and see for himself.

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