On Monday morning, Jiang Songran handed over the caregiving arrangements for his father to his mother and the caregiver, went home to tidy up a bit, and then went to the office of Source Intelligence Technology.
Although Han Luyi said that he could join the company remotely, Jiang Songran still wanted to go to the office more often when he joined the company, since his father's health had gradually stabilized. This would be beneficial for him to familiarize himself with the situation and carry out his work.
The job offer letter stated the office was on the thirteenth floor, but upon entering the elevator without swiping the access card, the only floor accessible was the twelfth floor, and the floor sign read "Source Code Technology".
Jiang Songran hesitated for a moment, then pressed the button for the twelfth floor.
The elevator doors opened, revealing an elegant and sophisticated reception area with a large, stylish sign on the wall reading "Haicheng Source Code Technology Co., Ltd."—not the company he was supposed to work for, Source Intelligence.
Jiang Songran slightly surprised.
Shen Congyun was already waiting at the door. When he saw Jiang Songran come in, he quickly reached out his hand to greet him.
"Dr. Jiang, hello, I am Shen Congyun."
"Hello, President Shen."
The two shook hands and walked inside, with Shen Congyun explaining the company's structure to Jiang Songran.
Jiang Songran then realized that the bustling scene on the floor beneath him wasn't Source Intelligence, but rather another company owned by Han Luyi, Source Code Technology. Source Intelligence Technology was upstairs, and you needed a card to get there.
"Let's go get the computer first and get an access card. I'll take you up in a bit," Shen Congyun said.
Shen Congyun also gave a brief introduction to Source Code Technology's current products.
Have you heard of BugKiller, a code vulnerability detection tool for programmers?
Jiang Songran shook his head. "Supreme Purity."
"The Exploitation of the Works of Nature Kaiwu? The most popular no-code app generation platform in China?"
Jiang Songran looked embarrassed. "I just returned to China and don't know much about the ecology here yet."
When Jiang Songran was in Silicon Valley, he focused on the world's top companies and unicorns. He had never heard of Source Code Technology, and his attitude towards Source Intelligence was that it "has potential and a possibility of success."
Now that Shen Congyun was bringing up the company's existing products in person, Jiang Songran didn't want to show his disdain, but he also found it difficult to offer any flattery, so he could only gloss over it.
Seeing that the atmosphere was getting a bit awkward, Shen Congyun quickly changed the subject.
After completing the formalities quickly, the two went up to the thirteenth floor.
Jiang Songran could clearly feel that this floor was completely different from the floor below.
The orderly atmosphere of the downstairs Source Code Technology is nowhere to be found here. Most areas are still messy, with unopened workstation boards and cardboard boxes piled on the floor, and several workers dismantling partitions.
Following his gaze, Shen Congyun explained, "Originally, we only rented a small corner, but recently, after the funds from Source Intelligence Technology's investors arrived, Boss Han rented the entire floor. He also plans to set up a new reception area to clearly separate the two companies."
Jiang Songran realized that Source Intelligence Technology had already raised at least an angel round of funding.
When he signed the employment contract, he only focused on the cash salary and the clauses regarding leave and business trips. As for the 2% Stock Options incentive, he didn't really pay attention to how much it was worth; it was probably just paper money.
Shen Congyun led Jiang Songran to Zhao Wenyuan's workstation.
Zhao Wenyuan wasn't at his workstation; he was in front of a whiteboard not far away, holding a marker and scribbling. Several engineers were beside him, discussing and updating the whiteboard with their findings.
Shen Congyun was about to go up and say hello, but Jiang Songran stopped him.
The two stood a few meters apart, watching.
This was a familiar area for Jiang Songran, and he quickly understood that Zhao Wenyuan was assigning tasks. Each engineer was given several Operators to adapt, and then they brought up the parts of their work that they were stuck on, and everyone discussed them together.
This Jiang Songran had ever seen this kind of work method.
It's not that the format of the discussion is novel.
The problem is that these people were assigned the wrong amount of work—each person had three to five Operators.
It's just one of the Operators. To get it running on domestic graphics cards and tune its performance to near Nvidia levels, based on his previous experience doing adaptations in Silicon Valley, it would take at least a team of five people two or three months to get it to a basic form.
He was a Google engineer.
It's no exaggeration to say that Google has one of the highest talent densities among companies worldwide; one Google employee could easily be worth two Google employees in the market.
He's now facing a domestic startup. The number of people, the workload... something seems off, doesn't it?
At this moment, Zhao Wenyuan noticed them, and Shen Congyun waved to him.
Zhao Wenyuan spoke a few words to the people around him, put the marker back on the whiteboard, and walked over.
"President Zhao, this is Dr. Jiang, who joined the company today," Shen Congyun introduced.
Zhao Wenyuan extended his hand, and the two shook hands.
"Dr. Jiang, I've heard of you since I joined Google, but unfortunately we never met in Silicon Valley. It was only in Haicheng that we finally met," Zhao Wenyuan said politely.
"President Zhao, that paper on domestic adaptation is exceptionally well-written; it gives me hope for domestic technology." Jiang Songran wasn't just being polite; his tone was enthusiastic. Although the two had never met, he already recognized Zhao Wenyuan's technical and academic abilities. "That paper was a major factor in my decision to Source Intelligence."
Upon hearing this, Zhao Wenyuan's expression became somewhat unnatural.
Jiang Songran he was just unaccustomed to being praised in person and didn't pay much attention. He was more concerned about something else.
What is the current training stack? What stage are we at?
Zhao Wenyuan was taken aback.
On their first day at work, most people exchange pleasantries, get a computer, familiarize themselves with the environment, and get to know their colleagues. That's how the day goes by.
Jiang Songran just arrived and didn't even know where his seat was, but he was already in work mode.
Zhao Wenyuan didn't mind.
He first completed the handover with Shen Congyun, then led Jiang Songran to his computer and opened a progress chart.
"Ten key Operators have been implemented so far." The screen displays a row of Operators names, highlighted in green, followed by data comparing them to similar implementations on NVIDIA cards; all are above 80%.
A large portion of the remaining Operators names are grayed out, and many also have numbers after them, but these are all below 80%.
Jiang Songran guessed those were ongoing projects, but he didn't know why so many were being pursued simultaneously. Did Source Intelligence Technology have a large number of engineers working elsewhere besides the engineering team present?
But he didn't ask.
Zhao Wenyuan continued, "Our current pre-training is running on the NVIDIA GPU cluster provided by Dingsheng, and it's expected to finish by mid-February, after which we'll begin post-training. Boss Han's goal is that after the first version of the Model Training, we need to build our own domestic GPU cluster by early March, so we can run inference on it."
"Wait, what did you say? Running deduction on a domestically produced card? When?" Jiang Songran couldn't help but ask.
"Early March," Zhao Wenyuan replied.
"This year? In early March this year?" Jiang Songran asked again.
"Yes, early March this year," Zhao Wenyuan replied affirmatively.
Jiang Songran instantly plunged into self-doubt.
He wasn't sure which of them, him or Zhao Wenyuan, was at odds.
His papers and our conversation clearly showed him to be a pragmatic computer scientist, but now he's saying things like a conman making empty promises?
And even if you wanted to lie, you could lie to investors or superiors, but there's no need to lie to me, is there? How could I possibly fool you?
Zhao Wenyuan glanced at Jiang Songran, seemingly noticing his odd expression, and smiled slightly.
"Dr. Jiang, are you curious about our development process?" Zhao Wenyuan asked.
"I'd like to hear the details," Jiang Songran said patiently.
Zhao Wenyuan pushed his chair towards Jiang Songran and said, "Please sit down."
Then he pulled up another chair and sat down.
Zhao Wenyuan closed the document on the desktop, opened the code editor and terminal, and ran the programming Agent in the terminal.
Zhao Wenyuan typed on the keyboard while explaining to Jiang Songran.
"We're using the 'Operator Targeting Method,' the path set by Boss Han," Zhao Wenyuan said. "First, we select an Operators, set a benchmark, typically 80% of the performance of similar implementations on Nvidia cards. Then, we have the Agent generate an initial version, run tests, and if it doesn't meet the benchmark, we iterate again based on the test report."
Jiang Songran shifted his gaze from Zhao Wenyuan's face to the screen.
He didn't pay attention when he heard the words "Agent," assuming it was some kind of test script or an open-source code generation model wrapped in a shell.
However, he saw that Zhao Wenyuan had opened a new, empty Operators folder in the terminal.
The two had already looked at other folders, each containing over a dozen iteration records, with each iteration corresponding to code, test reports, and performance analysis reports.
Zhao Wenyuan typed a prompt message, which essentially stated the domestic adaptation for generating this Operators, as well as some other technical constraints, and then sent the prompt message out.
A green thought process message appeared on the terminal, describing the Agent's operational status line by line:
The Agent first breaks down the task and plans the steps.
Then came the code generation: a few hundred Operators core code, CUDA-style but adapted for domestically produced cards, which was generated very quickly.
Next is the automatic writing of test cases to cover various boundary conditions.
Finally, we ran the tests.
The test cases were run one by one, and the first thirty or so were all green "PASS".
When running a test labeled "performancebenchmark", the output turned yellow: [Functional: PASS]
[Performance vs Nvidia: 67.3%]
[Threshold: 80%]
[Status: FAIL(below threshold)]
Jiang Songran saw it clearly: the function was implemented, but the performance did not meet the standards.
The test failed.
But that doesn't mean the process isn't amazing.
Jiang Songran's current thoughts are hard to describe.
First, he breathed a sigh of relief—if he saw a project that used to take several person-months to complete in five minutes, he might have been terrified.
However, it's already quite remarkable that the first iteration could achieve 67.3% of the performance of Nvidia cards.
And this took less than five minutes.
Zhao Wenyuan hovered the mouse over the output on the screen Jiang Songran, "It didn't pass the benchmark, but we also have a complete analysis report. It's very clear which direction to adjust next. The engineers just need to figure out the direction for the next round, turn it into new prompts, and then run it again."
Jiang Songran silent.
He had been holding his breath for a while, and it took him a long time to exhale. He still couldn't catch his breath.
Jiang Songran wasn't shocked by any specific data that the Operator Adaptation to in this round.
The data adaptation was certainly impressive, but what shocked him even more was the process of obtaining the results.
It's something that can break down tasks, write Operator Adaptation code, generate tests, and run performance analyses on its own.
That was a new paradigm, a new tool.
It's not code completion, nor is it a question-and-answer chat interface.
First, there are intervention-based programming tools that can autonomously arrange multi-step, complex tasks.
"Agentic programming," Jiang Songran murmured.
He doesn't work in this field, but after returning to China, he heard about it from friends in Silicon Valley. It's currently a hot topic of discussion among AI companies and Large Model teams in Silicon Valley.
Experts in every related field are talking about it.
Everyone says this thing has the potential to change the way humans interact with computers.
But in his memory, no one had actually made it yet.
But now, in Haicheng, 10,000 kilometers away from Silicon Valley, in a startup company, Jiang Songran has personally witnessed the product that people were discussing, the one that "hasn't been made yet."
Moreover, this is a highly refined new tool.
This new tool is doing something that even a company of Google's caliber would spend several engineers and months refining.
Jiang Songran swallowed hard.
He finally raised his gaze and looked at Zhao Wenyuan.
"President Zhao," he heard his voice was a little dry, "Which company made this 'Agent'?"