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 186

Handled, Clean

1,638 words 5 min read 3 views 3mo ago Source ​
Skip to the end of the chapter
Prev
Next

Monday morning at 10:00 AM, 32nd floor of Dingsheng Tower.

Zhang Chi's office faces south, and on a clear day, he can see the outline of Lujiazui. Today it's foggy, and the view outside the window is a uniform gray-white.

When Liu Yaguang came in, Zhang Chi was reading a weekly report. Without looking up, he said, "Speak."

"The Source Code integration is complete." Liu Yaguang sat down on the sofa, his voice low. "We can get a copy of the data they send to the Computing Power cluster in real time."

Zhang Chi put down his phone, glanced at him, and asked, "What's coming in now?"

"Pre-training data," Liu Yaguang said, "Is basically 'washed' corpora."

Zhang Chi frowned. "What do you mean by 'washed' data? Their model has a strong understanding of intent; is this the data they used?"

Liu Yaguang straightened up and shook his head. "Training Large Model isn't something that can be done in one step."

He paused, considering how to explain, "Let me explain. This process involves several stages. The first step is called pre-training. In this stage, the model is fed massive amounts of raw text—web pages, books, papers, forum posts—the more diverse the sources, the better, and the larger the volume, the better. A good model might be fed trillions of words in this stage. What the model does in this stage is very simple: it repeatedly guesses what the next word will be. If you show it 'Today's weather,' it guesses 'very good'; if you show it 'the essence of deep learning is,' it guesses 'optimize.' If it guesses correctly, it moves forward; if it guesses incorrectly, it adjusts the parameters. After repeating this hundreds of billions of times, the model gradually learns the rules of language and a vast amount of knowledge and common sense about the world."

Zhang Chi nodded.

Liu Yaguang continued, "The data at this stage doesn't need manual annotation; we just feed it whatever text we have. But we need to wash it first—filter out garbled text, duplicate content, and low-quality junk. Otherwise, the model will learn a lot of noise, which will be harmful. Washing data sounds simple, but how well it's done directly affects how solid the foundation of the pre-trained model is."

"And what's the second step?"

"The second step is to make the model truly intelligent," Liu Yaguang said. "To enable it to understand human intent and know what the user truly wants behind the same sentence, we need another batch of data—those specially labeled data, telling the model, one by one, 'This is the correct answer in this scenario.' This batch of data is much smaller, but each piece requires manual judgment, making mass production difficult. As you told me before, the core reason why the Source Code model has such strong intent understanding is this batch of Annotated Data."

After listening, Zhang Chi summarized, "What they've received so far is the data from the first step, but what we really want is the batch from the second step."

"Yes," Liu Yaguang affirmed.

Zhang Chi looked out the window and thought for a few seconds before asking, "Can we technically confirm that we can get all the data they send to the cloud?"

"It should be possible," Liu Yaguang said. "As long as the data enters the Computing Power nodes they use, we have a complete copy on our side, and they won't be able to tell."

Zhang Chi leaned back in his chair, his tone calm. "Then let's export this batch of pre-trained data for use."

Liu Yaguang was somewhat surprised. "This batch of data is unlabeled and has limited value. We also have our own corpus--"

"I know," Zhang Chi said. "I'm not doing this for the data itself. I mainly want to check if this channel is stable and if it can be safely exported. Let's run it once. If there's no problem on their end and the path is fine, then things will be easier."

He gave a half-smile and added, "Since there's no harm in using it, let's send it to Kun Yuan. Whether they need it or not is up to them."

Liu Yaguang nodded and said, "Understood, Boss Zhang."

After leaving Zhang Chi's office, Liu Yaguang walked out of Dingsheng Tower and boarded a bus to Dingsheng Cloud Park.

Liu Yaguang's workstation is in the cloud computing park. He came here today specifically to report to Zhang Chi in person.

At 2 PM, Liu Yaguang sent Zhang Chi a message from his workstation. "Boss Zhang, the data is ready."

Zhang Chi replied a few minutes later, "Is it clean?"

Liu Yaguang: "It's been cleaned up."

Zhang Chi: "Send it to the data team."

Liu Yaguang put down his phone, opened the internal messaging system on his computer, and sent a message to Wu Yinghao: "Are you there?"

Then he picked up his phone and sent Wu Yinghao a WeChat: "New data."

A short while later, Wu Yinghao sent back an emoji of someone wearing a green helmet.

Liu Yaguang sent a link to an encrypted compressed file via internal communication.

Then he sent Wu Yinghao a message on WeChat: "Same password as before, follow the data entry process, write 'external purchase' as the source, I've already generated the batch number, just fill it in."

Wu Yinghao replied to the WeChat message quickly, "How's the quality?"

"It's been washed, so it's usable."

"Good."

This wasn't the first time Wu Yinghao had obtained data from Liu Yaguang; it wasn't a clean business, and the two had a tacit understanding.

The data team receives these batches every now and then, from all sorts of sources: some are scraped, some are purchased, and some are "shared with partners." Filling in "external procurement" is the cleanest way to write it, because the procurement sources are so diverse that it's not easy for Auditing to find out.

This isn't the first time this has happened, and it won't be the last.

Three days later, the Kun Yuan project team.

Liu Dahai stared at a curve on the screen for about three minutes without moving.

This is Kun Yuan's loss curve for this round of pre-training.

Loss is the loss value – it can be understood as the degree to which the model makes mistakes; the lower the number, the better the model has learned. The training process involves letting this curve continue to decline.

The curve is indeed trending downwards, but faster than he expected.

Liu Dahai stretched out the timeline and overlaid the curves from the previous round for comparison. The difference was obvious; it didn't seem like fluctuations within the margin of error, but rather a systemic improvement.

He mentally ruled out several possible causes: the learning rate hadn't been changed, the model architecture hadn't been altered, and the Computing Power configuration hadn't changed. Could there be something special about this batch of data?

Liu Dahai immediately stood up and went to the data team.

The head of the data team is Wu Yinghao, a man in his early thirties who wears glasses. The main task of the data team is to collect and clean data.

This is a dirty and tiring job, and it's at the bottom of the Large Model development industry chain.

When Liu Dahai arrived, he was checking an inbound log.

"Yinghao, did we use any new data in this round of training?" Liu Dahai patted the back of his chair.

Wu Yinghao turned around and greeted him. He then flipped through the log. "Let me see the recent new arrivals... A batch was purchased externally a few days ago and put in for use."

Liu Dahai nodded and said, "The quality of this batch of data is excellent. We've come down from the loss, and it's even faster than the last round. This batch of data should definitely be credited for that."

Wu Yinghao paused for a moment, then stood up and leaned close to Liu Dahai's ear, his tone filled with a cautious eagerness. "Teacher Dahai, this batch of data was very noisy when it came in. We spent time processing it ourselves; several colleagues worked overtime to run the cleaning process. Do you think the results were good?"

Wu Yinghao was certain that Liu Dahai wouldn't inquire about the specific source of the data, because he never did, and it wasn't within his job scope.

"Great job washing!" Liu Dahai said earnestly. "This batch is better than the previous ones. Keep it up!"

Wu Yinghao smiled and nodded repeatedly, quickly glancing around to make sure no colleagues were watching their conversation.

Liu Dahai returned to his workstation, opened this week's training report, and added a few lines to the progress section:

"The data quality in this round of pre-training has significantly improved compared to the previous round, with the loss convergence speed increasing by approximately 9%. Preliminary assessment suggests this is related to the quality of the newly added corpus, and the data cleaning work in this round was thorough. Kun Yuan's pre-training phase is progressing smoothly and according to plan."

He read it once, changed two words, and submitted.

The weekly report was sent out and went into Lin Shaofeng's inbox. Lin Shaofeng is the vice president in charge of algorithms and AI. Every Friday afternoon, he would summarize the weekly reports from each group and forward them to those who needed to know. This was a fixed procedure, four or five pages long. Today's content was quite extensive. Kun Yuan's section was hidden in the middle of the third page. It was short and not highlighted in red or bolded.

That evening, Zhang Chi received a summary weekly report forwarded by Lin Shaofeng on his phone.

He scrolled down quickly, looking for the section related to Computing Power. Kun Yuan's progress was on the third page. He scrolled over it, his eyes lingering on it for less than two seconds before continuing to look down. Nothing special; the pre-training was running, the loss was decreasing, everything was normal.

Everyone stands on their own piece of the puzzle, seeing only what is real, but no one knows what the puzzle will ultimately become.

Meanwhile, Han Luyi, sitting in his office marking data, looked at the Sight with a puzzled expression: Has my experience point increased even faster lately?

End of Chapter
Prev
NovelSummaries
Next

Explore the wiki

Han Luyi

Comments

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