At 7:30, there were still very few people on the thirteenth floor.
Zhao Wenyuan pushed open the glass door to his office on the thirteenth floor. The corridor lights were still in motion-sensor mode, lighting up in segments. He put his backpack on his workstation, opened his laptop, and went to the break room to get a glass of ice water.
This is his habit. He arrives an hour and a half early to use the quiet time to review the data.
The model's training panel is on the first tab of the browser; the results of the sixth round of fine-tuning last night are already in.
He clicked on the loss curve.
It's basically flat.
Just like the day before yesterday, just like the day before that, just like last week, the post-training has reached this point, like wringing water out of a towel, with only the last few drops left. Zhao Wenyuan turned off the control panel and took a sip of ice water.
My phone lit up; it was a notification from LinkedIn.
He usually doesn't look at LinkedIn; nine out of ten messages there are templates sent by headhunters—"high-salary referrals" and "looking forward to exchanging ideas"—they're so lazy they don't even bother to change the names after copying and pasting. But this one was different; the sender's subject included "Neus AI."
Zhao Wenyuan clicked on it.
The message was from the Talent Recruiting Director at Neus AI. It wasn't a mass-sent template. The first sentence mentioned his research direction at Google, and the second sentence mentioned the achievements of the Heavenly Craft model in the field of code generation—which were just announced on the official website yesterday.
The gist of the message is: Neus AI is planning to establish a research team in China, possibly in Haicheng or Beijing. They have reviewed his background and would like to schedule a time to discuss it.
Zhao Wenyuan did not reply.
But he followed the link in the message to Neus AI's official website.
Team composition: Over forty researchers, with core members from OpenAI, DeepMind, and MetaAI Lab. Technical goal: To create the best general-purpose Large Model. Currently hiring for the position of base Model Training.
Every single one of them hit the spot that itchies most in his heart.
He certainly didn't want to switch jobs, but he really envied those resources. He had a dozen or so people under him, and every day they could only grind for a little improvement under the ceiling of post-training.
Footsteps sounded at the end of the corridor, and Zhao Wenyuan quickly closed the page.
At 9:55, the thirteenth floor was already half full.
Little Niu, carrying a cup of coffee, turned around and said, "Brother Zhao, the ablation experiment from last night is finished, and the results are on the shared drive."
"I've seen it," Zhao Wenyuan said.
Little Niu paused for a moment, then asked, "What time did you watch it?"
"Morning."
Little Niu opened his mouth, but didn't say anything, and turned back to his workstation.
Xiao Guo, who was adjusting the data cleaning script next to him, turned to ask how to handle token splitting involving multiple languages. Zhao Wenyuan walked over, glanced at the screen, and pointed to the third line. "Let's change the mode here and run a version to see the effect."
Xiao Guo nodded and began making changes.
Zhao Wenyuan stood in the middle of the aisle between workstations and scanned the area.
Thirteen people, most of whom he personally interviewed and selected. Little Niu was his junior from Jiaotong University, and Xiao Guo was the only one who dared to argue with him about the sampling strategy face-to-face during the interview. In the corner, two new data engineers were annotating on the screen without even looking up.
8:53.
Zhao Wenyuan closed his laptop, picked up his phone, and went downstairs.
When I knocked on the door, Han Luyi was already there.
There was an unfinished cup of coffee on the table, and the laptop was on but the screen was off. Han Luyi's eyes lit up.
"Sit down." Han Luyi closed the door.
Zhao Wenyuan sat down opposite the desk.
There was no small talk.
"I want to create a general-purpose basic model," Han Luyi, getting straight to the point.
Zhao Wenyuan showed no surprise.
He had been waiting for similar words. Heavenly Craft's training curve had been flat for almost two weeks, and the things he could do after training had reached their peak, but... Zhao Wenyuan remained silent.
He was thinking about how to start the conversation.
"Boss Han, the general basic model is just burning money."
"I've seen it, and I've done it."
He didn't elaborate, Han Luyi knew his resume; he had worked at Google.
"It's not that I don't want to do it," Zhao Wenyuan said.
He thought about it.
But I know the price.
Han Luyi did not rush to respond.
"Where will the money come from? What will we do when it runs out?" Zhao Wenyuan said, "The Source Code finally has a positive cash flow."
"So I plan to set up a new company, without touching the Source Code," Han Luyi said. "An independent entity, an independent team, and independent financing." Zhao Wenyuan looked at him in surprise.
Han Luyi started by explaining each point one by one.
"Building a model boils down to three things: data, algorithms, and Computing Power."
"Algorithms, you take the job. This is your forte, and something you've always wanted to do."
Zhao Wenyuan did not deny it.
"You don't need to worry about the data; I'm confident I can get the best data."
"I'll handle the Computing Power and the money too."
How do we handle this?
"Tell a story, but the story isn't about ' Source Code money to build a model'—investors might not buy into that." Han Luyi moment.
"It's about having a team with real-world data and the ability to implement solutions, aiming to create a useful, general-purpose model."
"The two stories faced different levels of difficulty in securing funding."
Zhao Wenyuan remained silent.
He had considered this logic, but as a technologist, he was thinking about the underlying architecture and training strategies, not how to package the technology into a story.
Han Luyi open the notebook, turned it around, and pushed it toward him.
The screen displays an email in English.
From: NeusAl.
Zhao Wenyuan's fingers twitched slightly.
"I received it a few days ago," Han Luyi said. "It was written to me by their CEO. The core idea is one thing: autonomous workflow. AI autonomously breaks down tasks, autonomously calls tools, and autonomously verifies results."
"This is the next paradigm, and they're already doing it."
Zhao Wenyuan stared at the screen for more than ten seconds.
Autonomous workflow.
He didn't see this on NeusAl's official website this morning.
He didn't say it.
But a quick comparison formed in my mind: Han Luyi and Neus AI were looking in the same direction. The difference was that one had a model but no product, while the other had a product but no model.
"They lack product implementation, we lack a universal model."
Han Luyi's voice rang out from the other end, almost in sync with the judgment in his mind.
But we have something they don't.
Zhao Wenyuan pushed the notebook back.
The office was quiet for a long time.
Zhao Wenyuan rationally made a list in his mind: new company, zero funds, starting from scratch, hellish difficulty.
But there was a sound coming from the other side.
This is what he wants to do: a self-training base, instead of painstakingly grinding it down under the ceiling of post-training, and not relying on other people's model bases. He thought of what had happened before.
Han Luyi took the initiative to talk to him about Wang Zhiyuan. It was Han Luyi pushed for Wang Zhiyuan's resignation. Han Luyi could have kept quiet and let the matter fester over time, which would have been the safest option for him, but he did.
I did the deed, took the consequences, and even let you know.
Sometimes, choosing the right people is more important than what you do.
Zhao Wenyuan spoke.
How much time do you want to think about it?
Han Luyi looked at him.
How long will it take you?
Zhao Wenyuan paused for two seconds.
"No need, I'll do it with you!"
Zhao Wenyuan came out from the twelfth floor and went back up to the thirteenth floor.
I stopped by the tea room, grabbed a glass of iced water, and went back to my workstation to sit down.
Open the notebook.
The training panel is still there, and the loss curve is still flat, exactly the same as it was an hour ago.
But he feels differently about this line now.
This is the end point of post-training.
It is also the Starting Point of another path.
He opened LinkedIn and found the message.
Delete.
Close the page.
After Zhao Wenyuan left, Han Luyi sat alone in the office for a while.
The hardest step has been taken.
Zhao Wenyuan agreed, and the core of the model company was now complete—an algorithm lead, a team framework, a technical roadmap, and most importantly, a trustworthy person.
He went through it in his mind.
First, fly to Beijing.
He Yunshen invested in Source Code Technology. Whether the new company can secure support from Hongyuan is uncertain. However, He Yunshen is the investor who best understands the value of Source Code's data assets, and his attitude will influence the judgments of everyone else.
If He Yunshen doesn't support it, then find a new one. He already has a complete story: people, achievements, and direction.
The internal workings also need to be well-ventilated; the Source Code cannot fall apart.
Han Luyi picked up her phone and scrolled to He Yunshen's WeChat.
"Boss He, when are you free to meet and have a chat?"
After typing it, he looked at it for two seconds, then deleted the line. It was too direct.
"Boss He, I have some personal business to attend to in Beijing tomorrow, and I'll take the opportunity to visit you. When are you free to meet and chat?"
Send.