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 200

Looking forward to working with you

2,828 words 9 min read 1 view 1mo ago Source ​
Skip to the end of the chapter
Prev
Next

Zhao Wenyuan glanced back at Jiang Songran's expression, secretly pleased.

I see other people showing off all the time, so today I'll show off for a day.

He casually remarked, "It's developed by Source Code Technology themselves, currently in internal testing, and should be released to the market soon."

Jiang Songran silent for a long time before finally asking, "You... you made this yourselves?"

Seeing that Jiang Songran's mental fortitude was acceptable, Zhao Wenyuan continued, "Yes, the earliest inspiration came from a product called Source Code, which was originally Open Things for users without a technical background. Later, it was developed into Large Model that users didn't need to understand code or intervene in its execution; they could iterate using only Chinese characters."

Zhao Wenyuan observed Jiang Songran's reaction and concluded that he likely already knew about Open Things.

"President Jiang works on the Source Code, applied this idea to the development process and created this Agent tool." After saying this, Zhao Wenyuan suddenly asked, "Did you see President Jiang downstairs just now?"

Jiang Songran shook his head. He had come up directly with Shen Congyun and hadn't had a chance to contact the employees from Source Code Technology.

However, he knew in his heart that this President Jiang, who could lead a project of this level, must be a key figure in Source Code Technology, and he secretly looked forward to seeing his demeanor.

"I'll introduce you to her sometime; you'll like her," Zhao Wenyuan said, then steered the conversation back to work.

"In short, this is our workflow: we distribute the Operator Adaptation and iteration tasks to each team member, and if someone gets stuck, we discuss solutions together."

Jiang Songran nodded; no wonder the work pattern he had just observed was like that.

It's not surprising that such tools exist.

From this perspective, this tool significantly reduces the time spent on implementing the code and conducting actual testing, but it's neither some kind of cutting-edge technology nor black magic.

Problems that can be solved will still be solved, just faster.

Problems that cannot be solved with this tool still cannot be solved.

For those who truly understand the principles, the time saved is the greatest, and the help is the most significant.

On the contrary, for those who already have only a superficial understanding, even with such tools, they are not very useful, and the problems they cannot solve will still remain unsolved.

The fact that this technical team can adapt ten Operators shows they are quite capable.

He believes that with his technical skills and experience, combined with this set of tools and such a team, it is not impossible to achieve full-chain adaptation for domestically produced graphics cards.

Jiang Songran nodded to himself.

Then he pointed to the results on the screen and said, "This Operators doesn't need to be 80% complete for now."

Zhao Wenyuan was taken aback and asked, "Why?"

"Because what we lack right now isn't the single-point performance of this Operators," Jiang Songran pointed to the screen. "It runs, functions correctly, and has 67% performance. If the goal is to publish a paper, of course we should adjust it to over 80%, but if the goal is to get the first batch of domestically produced card clusters running by early March, what we should be doing now isn't continuing to obsess over this one point."

Zhao Wenyuan's expression turned serious.

"Then what should we do?"

"Run end-to-end."

Jiang Songran opened several folders on the screen.

"Pull out the Operators that have already met the standards and assemble a minimum link. The model can be small, the fragments can be small, and the data can be fake, but the link must be complete to complete end-to-end debugging."

77

As he spoke, he began to write on the whiteboard next to him.

Model loading, forward pass, backward pass, gradient synchronization, optimizer update, checkpoint backup, and recovery.

Jiang Songran wrote the seven steps on the whiteboard without any pauses.

Everything he's done in the past few years has been on this chain, and he's thoroughly familiar with this knowledge and the problems that might be encountered in practice.

When I was creating this flowchart, the most time I spent was actually mentally finding the Chinese translations for all the English words.

After finishing the flowchart, Jiang Songran said, "Although we'll only be running reasoning in early March, training is the main focus. We need to understand the end-to-end process now, otherwise it will be too late to catch up later. The 'Operator Targeting Method' we're using now is correct, but given the tight schedule and heavy workload, the most important thing at this stage is to have a holistic view and adjust priorities from a macro perspective."

As he spoke, he made dots on the whiteboard with a marker. "Otherwise, everyone will be optimizing their own Operators, and by the time they're all put together, the real bottleneck will have appeared, and by then it will be too late."

Zhao Wenyuan stared at the contents of the whiteboard, lost in thought.

Jiang Songran is right.

So far, Zhao Wenyuan and his team have been working on a point-based approach, advancing Operators, one Operators at a time.

This is certainly necessary; without all the necessary points, it's impossible to connect them into a line.

However, system training requires more than just points; to ensure error-free connection, we need to start connecting now.

Zhao Wenyuan had truly witnessed Jiang Songran's prowess. On his very first morning at the company, before even having reviewed the entire project schedule, Jiang Songran had already pointed out the biggest hidden danger in the current project path.

Zhao Wenyuan turned his chair to face Jiang Songran and said, "Dr. Jiang, what do you think we should do now?"

Jiang Songran organized his thoughts and replied, "If the goal of the first stage is just to run inference, there's no need to pursue 80% efficiency on the Nvidia card. We can't fully utilize it. We should prioritize completing the necessary Operators and getting them running. I just checked, and besides the ten that are already marked in green, there are twenty-three others that have already passed the 60% threshold."

Jiang Songran glanced down at the screen and corrected himself, saying, "Twenty-four, that makes thirty-four."

Jiang Songran in his chair, moved it closer to the screen, and began Operators in the table.

"Reverse engineering, optimizer updates, and gradient synchronization should be prioritized. These won't be needed for inference in early March; we'll come back to them in April when we're fine-tuning the training."

Zhao Wenyuan nodded. These categories made up a lot of the table, and they were indeed only used in training, so it was perfectly reasonable to put them aside for now.

"Let's look at the reasoning path again." Jiang Songran moved the cursor to the first half of the table, "Matrix multiplication, LayerNorm..."

Softma and activation are already highlighted in green. Where are the attention-related Operators?

Zhao Wenyuan scrolled down and pointed to one item. "Standard Attention, 61%. The Flash-Attention version hasn't been ranked yet."

"We won't use Flash; we'll use the standard one for now, as long as it runs," Jiang Songran said. "What about KV-cache read and write?"

"Here, 64%."

"That's enough." Jiang Songran leaned back. "Add One sample and one RMSnorm. These two are simple; merging them or not doesn't matter. Let's get them first, and I'll count them."

He tapped the table, stopping at the very end.

"Two short."

Zhao Wenyuan was taken aback.

His original plan was to complete 80% of all the Operators, and then call on his personal "grandpa "Han Luyi to solve any problems he encountered. At the current pace, it might take until the end of February, at which point he would work on the end-to-end solution.

But now that Jiang Songran dismantled it...

Only two left.

"You—" Zhao Wenyuan was momentarily at a loss for words, "I'll have the team start preparing for these two today."

"No rush." Jiang Songran waved his hand. "First, connect the thirty-four that have already met the standards using the existing scaffolding, run them through under no-load conditions, and establish the connection. We can add the other two after that. Then it will be clear what's wrong."

Zhao Wenyuan nodded, stood up, and turned to call for help.

He walked along, feeling a little annoyed.

I failed at showing off again today.

Where did this deity come from?

Zhao Wenyuan called all his team members over, assigned them tasks, and said that they would start setting up scaffolding today and try to pull out a link using the smallest model, and run it in an 8-card environment.

The surrounding engineers were all taken aback. "Today?"

Jiang Songran added from the side, "When running it, turn on the Profiler. Don't just look at the final success or failure; the time taken at each step, the waiting time, the memory copy, and the communication ratio—all this data is valuable."

Since the crowd didn't recognize Jiang Songran, they all looked at Zhao Wenyuan. Zhao Wenyuan nodded, and then everyone dispersed.

Zhao Wenyuan turned to Jiang Songran and said, "Dr. Jiang, thank you. This was indeed something I hadn't thought of. It's good that you pointed it out now."

Jiang Songran showed no arrogance whatsoever. He said, "This isn't your fault; we've moved too fast."

Zhao Wenyuan asked in confusion, "Too fast?"

Jiang Songran said, "Do you know how long it took us to realize this problem when we were making TPU?"

Zhao Wenyuan then asked, "How long?"

"Three years," Jiang Songran said. "How long have you been doing this? At most, half a year?"

Zhao Wenyuan had lost all desire to correct this; it was better not to bother with such a petty boast.

Jiang Songran continued, "Normally, when a team is still stuck on the first SDPA, they don't need to think that far ahead."

Our current problem is that we've already broken through the most difficult part to start, so systemic issues have been exposed prematurely.

"That's a good thing."

After saying that, Jiang Songran opened the folders on Zhao Wenyuan's computer containing Operators with over 80% efficiency and said, "You oversaw these implementations, right? This requires a deep understanding of the underlying mechanisms. Regardless of whether there are tools like Agent, this understanding cannot be faked."

Jiang Songran was a few years older than Zhao Wenyuan and had managed dozens of people before. He said these words because he was afraid that his previous behavior had damaged Zhao Wenyuan's authority in front of his subordinates. So he wanted to encourage him, repair their relationship, and prevent any future problems.

After all, the two will most likely be working together for a long time in the future.

Unexpectedly, Zhao Wenyuan's face showed that uneasy expression again after hearing this. After hesitating for a while, he finally spoke, "Actually, I didn't write these."

Jiang Songran asked curiously, "Who wrote that?"

"It was written by Boss Han." Zhao Wenyuan said, glancing at the elevator. Recently, Han Luyi has been coming directly to the thirteenth floor after starting work. Today, he heard that someone from Beijing came—the due diligence team from State Grid Investment, led by Huang Yu —and Han Luyi needs to be in contact with them, so he hasn't arrived yet.

"Boss Han?" Jiang Songran been surprised so many times today that he was almost speechless with surprise. "Which ones did he write?"

Zhao Wenyuan is already numb to it; why do you have to bring up something he doesn't want to talk about?

"All of them."

"All of them?" Jiang Songran took a deep breath, then exclaimed in surprise. He recalled Han Luyi, the young man he had met at the hospital that day. They had only met once, and he had initially thought Han Luyi was an HR person. Jiang Songran had no intention of joining another startup, but Han Luyi's words seemed to have a magical power, directly addressing his deepest needs. After only a few tens of minutes of conversation, he was convinced.

Since then, he thought Han Luyi a talented manager and that he could make good use of his skills here. But he never expected that he would find out today that Han Luyi himself was also so skilled.

"Boss Han— " Jiang Songran asked gossipily, "Which university did you graduate from for your doctorate?"

"An undergraduate degree from Jiaotong University." Zhao Wenyuan's feelings were complicated. He hadn't managed to show off his own abilities, but he had done so for Boss Han, and he felt inexplicably pleased.

After saying this, they both fell silent, and then tacitly ended the conversation.

Jiang Songran returned to his new workstation, set up his computer, and looked around; it was still a mess.

He suddenly remembered what Shen Congyun had said: Source Intelligence Technology had already received funding. Curious, he turned on his computer and did a quick search on Lark, and indeed there were relevant announcements.

Source Intelligence Technology completes RMB 450 million Series A financing to accelerate the R&D of domestic Computing Infrastructure.

450 million RMB in funding, valued at 3 billion RMB.

Jiang Songran looked puzzled. He had originally thought that he had turned down offers from several other major companies and joined Source Intelligence Technology at a lower salary for the sake of his family. He hadn't expected that in addition to the 2 million yuan in cash, the 2% Stock Options were now worth 60 million yuan, which was much more than what Dingsheng and Kunlun had offered.

Jiang Songran became somewhat uneasy after that.

He wasn't short of money, but so much money and such trust gave him a feeling of "a scholar would die for one who understands him."

Finally, Han Luyi sent him a private message on Lark: "Dr. Jiang, how is your first day? I'm sorry I had other things to attend to and couldn't welcome you right away. If you have any questions, you can consult @Zhao Wenyuan; he's the most knowledgeable about this area."

Jiang Songran a moment, then replied, "Boss Han, everything is going well for me, thank you for your concern. The Stock Options reward is truly generous, thank you again."

"The Stock Options are already specified in the employment contract, right? Does Dr. Jiang have any additional requirements regarding the travel clauses?"

Jiang Songran thought to himself, "I did write it, but I didn't expect it wasn't paper money."

Before he could reply, Han Luyi sent another message.

"I believe you are worth this price. This isn't just a salary; it's proof of partnership. I look forward to working with you."

Jiang Songran silent for a moment before typing his reply.

"Boss Han, I also look forward to cooperating with you."

With that topic settled, Han Luyi started a new one.

"I just heard from Zhao Wenyuan that you've roughly figured out the training process. This workload is definitely not enough for the existing team, so I'll continue to recruit, but this field is too narrow, and regular recruitment is too inefficient."

Jiang Songran understood what he meant. "You want me to recommend some people?"

"If you have a suitable candidate for any of the positions under you, you can recommend them for an interview."

"Okay," Jiang Songran replied.

Then he closed Lark and opened a WeChat group that hadn't been used in a long time.

[AIInfra Chat Group]

It's a large group of 500 people, from all over the world, including the United States and China; some are from Meta, etc.

There are people from large companies like Amazon, people from startups, and some technical leaders from large domestic companies.

The only thing they have in common is that they both work in this industry.

Jiang Songran thought for a moment and decided not to send the message in the name of Source Intelligence Technology. Source Intelligence Technology was not well-known at the moment, and sending it out would only make potential job applicants hesitant.

He sent a message only in his own name.

"I've returned to China and am now leading an Infra team at a startup, focusing on large- Model Training stacks for domestically produced AI Accelerator Card. The team is hiring, with roles including training systems, distributed scheduling, communication, fault tolerance, and AI Infra engineering. Feel free to message me if you're interested."

Not long after the message was sent, the group chat, which had been silent for a long time, came alive again.

"Jiang Shen, you've returned to China too?"

"Jiang Shen, are you optimistic about the direction of domestic Computing Power substitution?"

I read a paper a couple of days ago about non-CUDA attention; it seems to have been written by a Chinese team? Has anyone been paying attention to it?

"Domestic cards? Training? Does anyone still dream about this?" -- Upon seeing this, Jiang Songran's lips curled up slightly.

A week ago, he would have had the same reaction, but now he just thinks the other person is a frog in a well.

"I have a junior colleague who just came back from Singapore. He's doing scheduling now. Do you want me to recommend him to you?"

Jiang Songran knew his standing in the AIInfra field; sending out such a message would surely attract a lot of resumes.

I need to spend some time carefully choosing something in the next few days.

End of Chapter
Prev
NovelSummaries
Next

Explore the wiki

Han Luyi
Zhao Wenyuan

Comments

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