Three days later.
Prosperity Group, Large Model Research Institute, AI Lab.
Liu Dahai had been in the lab for over twenty hours, working through the night, his eyes bloodshot and red.
He's not young anymore, and staying up so late is taking a toll on him.
But I was still very excited.
The time he met Han Luyi, the model called "Glutinous Rice Ball" that Han Luyi had fine-tuned using a 7B miniature model and 10,000 data points greatly impressed him.
Later, he was called to the office of Zheng Xiaobo, the CEO of Dingsheng, to ask about the model. He told the truth that the Annotated Data must be good.
Zheng Xiaobo asked him what achievements he could make if he could obtain data of the same quality.
Liu Dahai said he could make a hole in the sky.
That's an exaggeration.
But to his surprise, the data actually ended up in his hands; Cheng Yuan, the VP of the Strategic Investment Department, personally delivered it to him via USB drive.
And it doesn't contain 10,000 entries, it contains 80,000.
It is said that Dingsheng and Source Intelligence Technology signed a cooperation agreement, and the other party shared the information.
After a quick glance, Liu Dahai was able to confirm that this was the data from the "Glutinous Rice Ball" fine-tuning. With the data labeled to this extent, no wonder it could fine-tune a small 7B model so effectively.
Eager to get started, Liu Dahai pulled his team together for a new round of post-training, going straight to the maximum computing resources.
We examined the results round by round, and the effects were better than we had imagined.
Before I knew it, I'd stayed up all night.
Many people believe that Large Model is a good job, as it's lucrative, highly sought after by major companies, and attracts widespread attention.
But only those who actually work in this field know that it's not an easy job at all. Many times, you have to figure out how to reduce losses, how to converge the reward model, how to balance various indicators, and how to improve coding skills, but then your mathematical reasoning skills drop.
Some can be explained by theories, while others cannot be explained.
Working in this line of work is like sailing a ship on a pitch-black sea.
You can see neither the end nor your surroundings; you only know that you are moving forward, but in which direction are you moving?
Have no idea.
I don't know if you'll reach land first or crash into an iceberg first.
No wonder model makers jokingly call themselves alchemists; before you start the furnace, you don't know if it will succeed.
They look so alike.
Knowing how difficult it is to improve a model's capabilities even slightly, Liu Dahai was so excited to suddenly obtain so much high-quality Annotated Data and improve the model's capabilities so much.
But after the initial excitement wore off, he suddenly sensed that something was wrong.
With such high-quality Source Intelligence Technology, why don't they use them to train their own models? Even if they don't have the resources to start from scratch, they could still achieve amazing results by fine-tuning existing open-source base models.
Why is there absolutely no sound?
After noticing this slight discrepancy, Liu Dahai carefully examined the shared Annotated Data again.
Something felt off.
He flipped through them one by one.
Is the quality high? Yes.
Is the logical chain tight? Yes.
But the more Liu Dahai looked, the more his brows furrowed.
A team member passing by noticed his expression and couldn't help but ask, "Dr. Liu, what's wrong? The results are better than expected, why do you look so pale?"
Actually, this team member is also a PhD himself. It would be strange if someone working in Large Model group wasn't a PhD.
But at least in this group, the only person who can be called Dr. Liu is Liu Dahai; everyone else is addressed by their given names.
This is naturally because Liu Dahai holds the highest position, Chief Scientist of Dingsheng Artificial Intelligence.
Liu Dahai looked at his team members, pointed to a label on the screen, and asked, "What do you think of this?"
The group member was completely confused. After looking at it, he said, "It's pretty good."
Liu Dahai then pointed to another one and asked, "What about this one?"
The group member read it carefully once, and then said, "This one is pretty good too, isn't it?"
He was starting to lose confidence after Liu Dahai's questioning.
Liu Dahai then asked, "What do these two annotations have in common?"
The team member thought for a moment, then tentatively asked, "High quality?"
Liu Dahai was taken aback. That was true, but clearly useless information. "Don't you think that the semantics of these annotations place too much emphasis on guiding data compliance?"
Huh? The team member was taken aback. The Annotated Data was guiding the model's behavioral patterns, or rather, its personality.
It's normal to have compliance requirements, isn't that what alignment training is all about?
Looking at Liu Dahai, the group members thought he was being a bit of an unfounded worryer.
The quality of these labels is exceptionally high, astonishingly high.
If you go out and tell others, "I now have a batch of high-quality Annotated Data, the quality is so high, it also includes compliance guidance."
That won't be seen as complaining; it will only be seen as showing off.
That kind of smug, obnoxious bragging.
It's important to know that the biggest problem for the groups at Dingsheng Research Institute Large Model lately has been the lack of high-quality Annotated Data.
Whether it's an outsourced annotation team or Dingsheng's internal annotation team, there's quantity, but the quality is hard to come by. Sometimes, adding annotations actually reduces the model's capabilities and comprehension.
These scientists probably spend more time selecting annotations than the annotation staff spend on actual annotations.
The team member didn't say anything more and went back to doing their own work.
With the data in hand, everyone was excited and worked non-stop.
Seeing that no one responded to his suspicions, Liu Dahai shook his head and decided to report to his superiors.
Large Model are something that cannot be taken lightly.
"You guys can stop the next round of training for now, we'll continue when I get back," Liu Dahai called out, leaving the lab and heading straight for Lin Shaofeng's office.
Lin Shaofeng is the VP of Dingsheng Research Institute and Liu Dahai's direct supervisor.
Although the two have the same job title, Liu Dahai is an IC (Institutional Controller) and does not manage people, while Lin Shaofeng is in a management position. Liu Dahai's performance evaluation is determined by Lin Shaofeng, so there is naturally a difference in their positions.
Liu Dahai arrived unexpectedly. He told his secretary at the door, and after a while, the secretary invited him in.
"President Lin." As soon as he entered the room, Liu Dahai got straight to the point, "I think there's something wrong with the data that came in the day before yesterday. I suggest we pause the training for now."
"Oh?" Lin Shaofeng raised an eyebrow. "Didn't you tell me yesterday that the data quality was very high and the training effect was very good? How come you changed your mind in just one day?"
"No, nothing's changed," Liu Dahai said. "I'm not talking about quality issues; I just feel the logic and direction in the story are a bit too heavy-handed."
"What impact will it have?" Lin Shaofeng asked.
"It's hard to say yet, I need to think about it some more," Liu Dahai said honestly.
Lin Shaofeng frowned. Kun Yuan was the focus of the entire group, and he was under a lot of pressure. Even a one-day break would have a significant impact.
"How much time do you need?" he asked patiently.
Liu Dahai lowered his head and thought for a moment, "One week."
"What?" Lin Shaofeng thought he had misheard, and composed himself. "I remember you said that a complete round of post-training is two weeks, right? You're going to spend a week investigating Annotated Data?"
"Yes, the logic behind this label is a bit off," Liu Dahai replied, not finding anything wrong with it.
"No, we have to start training now," Lin Shaofeng said decisively, "And we need results within two weeks."
Liu Dahai's face flushed red. He hadn't expected Lin Shaofeng to refuse so decisively, and he was also a little agitated. "No, we can't train. I said no, and that's final. I'm the Chief Scientist; I'm responsible for the final model."
Lin Shaofeng thought to himself: Scientists can be incomprehensible at times.
He looked at Liu Dahai and slowly said, "Dr. Liu, you seem to be working too hard lately."
Liu Dahai was taken aback. Why was he suddenly concerned about my health halfway through the fight?
"You've been working so long straight, your judgment might be affected by fatigue. How about this, you go and rest in another group for a bit, and let your deputy take over the training."
Liu Dahai's expression changed.
"Lin Shaofeng, what do you mean? You think you can transfer me?"
"I'll consult with Boss Zheng. You can go back and wait for further instructions," Lin Shaofeng said calmly.
"You still need to consult Boss Zheng? Let me tell you, I'll consult Boss Zheng too. We'll see about this." Liu Dahai stormed out.
Lin Shaofeng rubbed his temples helplessly. He wasn't angry, just a little helpless.
Scientists are really hard to manage.
This matter only considers the company's interests. The entire company is monitoring Kun Yuan's progress. Why bother taking it to Boss Zheng, or even Boss Lü?
It just made me look bad.
Then he turned on his computer and began editing a report for Zheng Xiaobo.
After finishing the report, he asked his secretary to call over Liu Dahai's deputy, ready to start assigning tasks immediately. He knew this deputy well; he was competent and not as stubborn as Liu Dahai, making him relatively easy to manage—otherwise, he wouldn't have gotten the position.
I'll call him over first to arrange the work, and once I get a reply from Boss Zheng, I'll start this round of training as soon as possible. Hopefully, I can meet the requirements.
Kun Yuan's progress can no longer be delayed.
After finishing all these tasks, Lin Shaofeng made a phone call to Zhang Chi.
"Boss Zhang, are you busy?"
"No rush, no rush. What can I do for you, Mr. Lin?" Zhang Chi's tone was relaxed. After all, he had finished his work here, and the pressure was now on the Research Institute.
I've given you the data; if you still can't figure it out, it's not my fault.
"Here's the thing, Kun Yuan's training is going very smoothly, and we'd like to speed things up. Boss Zhang, do you think you can provide some support?" Lin Shaofeng asked.
"Of course." Zhang Chi chuckled. He was naturally more than happy to have the opportunity to contribute to such a big project.
After hanging up the phone, Zhang Chi immediately called Liu Yaguang.
Liu Yaguang answered the call quickly.
"Old Liu," Zhang Chi said, "How much spare GPU space do we have left? Allocate some more to the Research Institute."
Suddenly, as if remembering something, he lowered his voice and said, "Oh right, keep an eye on Han Luyi's side secretly. We're starting training together, so don't let them get ahead of us."
"Okay, I understand what you mean." Liu Yaguang also lowered his voice.
Zhang Chi was still not convinced and asked again, "How should we bet? Do you have any ideas?"
"Don't worry, I know what I'm doing," Liu Yaguang said. "A direct speed limit would be too obvious, so I've asked the dispatch team to lower their task priority, placing them after our own. Their nodes will still run normally, but when there's a competition for resources, we'll go first."
Will they be able to find out?
"They'll most likely not notice. It's load balancing, it's normal. We can see priorities here, but they can't." Liu Yaguang paused. "Also, I can arrange a routine inspection, during their peak training period, temporarily taking some nodes offline for maintenance. Maintenance is a periodic process; this is our Data Center, our Hardware, timely maintenance is essential."
"Okay, do as you say." Zhang Chi didn't ask any more questions. "If no new data comes in, be careful during this period and don't cause any trouble."
"Understood, don't worry."
After hanging up the phone, Zhang Chi leaned back in his office chair.
A faint smile lingered on his face.
On the other side, Zhao Wenyuan hurriedly made Han Luyi.
"Boss Han, we've been throttled." Zhao Wenyuan's tone was somewhat exasperated. "That Dingsheng is so unfair! Not only did they steal our data, but they also unplugged our internet cable!"
"Why is the speed limited?" Han Luyi asked as he logged into his Dingsheng Cloud account.
"The training throughput dropped significantly, but there weren't a single GPU error. We checked our own code, and it's fine," Zhao Wenyuan said. "Then I checked their maintenance notice, which was sent at 2 PM this afternoon. It said that some nodes would be taken offline for maintenance starting at 3 PM, and we should pay attention."
Han Luyi eyebrow. "Give me an hour's notice?"
"Yes, maintenance windows are usually open in the middle of the night, right? They're never open during the day," Zhao Wenyuan said. "Isn't that speed limiting?"
Han Luyi didn't say anything. At this moment, he had already logged into Dingsheng Cloud's account, his gaze fixed on the screen, and the Sight was activated.
He glanced around, his expression unchanged, and directly opened the online work order system, creating a new urgent work order.
He compared the timestamps of priority changes, the department to which the operation account belonged, and the historical execution windows of the maintenance tasks, to...
The document included screenshots of the daily throughput decline curves, clearly listing each point. The problem description stated. "The above actions have significantly slowed down our training tasks, which is inconsistent with the resource guarantee clauses in the service agreement. Please explain the reason and restore the task to normal priority; otherwise, we reserve the right to pursue breach of contract liability."
Send.
He put down his phone and said to Zhao Wenyuan, "Keep an eye on it; it should be done soon."
There was a moment of silence on the other end of the phone. Zhao Wenyuan stared at the training monitor's graph; the throughput remained stubbornly low. He waited, and after about two minutes, he couldn't help but speak again, "Boss Han, still no movement."
At the Dingsheng Cloud Operations Center, Liu Yaguang was handling other matters when a work order message popped up.
He glanced at it, then stopped using the mouse in his hand.
I watched it again.
He read the entire work order from beginning to end and felt a chill run down his spine.
Comparison of operation accounts, timestamps, and historical maintenance windows—how did they manage to examine such detailed information?
He sat in the chair for a while, without consulting Zhang Chi, and directly opened the backend, changed the priority of Source Intelligence's job back, and brought the "routine maintenance" node back online.
--
It's okay to do this secretly; the gamble is that the other party won't notice.
But if you get caught red-handed, the best course of action is to stand at attention and take the beating.
However, after all these years, this was the first time he had been caught.
Liu Yaguang replied to that work order, "After investigation, it was found that the system scheduling parameters were abnormal. It has now been fixed. Thank you for your feedback."
After closing the work order, he looked up the account at Dingsheng Research Institute and increased the Computing Power quota originally allocated to them by one level.
That's the only way.
We can only pray that the Research Institute can expedite the process.
Zhao Wenyuan looked at the work order that popped up on the screen and replied.
Thank you for your feedback.
Looking at the throughput data again, it has returned to normal levels.
He swallowed hard, looked down at his phone, and saw he was still on the phone with Han Luyi.
--
"You... is that all?"
Keep training.
99
Han Luyi hung up the phone.