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 145

Poach His Data Supplier

1,607 words 5 min read 2 views 3mo ago Source ​
Skip to the end of the chapter
Prev
Next

As soon as Zheng Xiaobo said those words, Liu Dahai knew he didn't need to say anything more.

Zheng Xiaobo truly understands.

To lead a giant like Dingsheng, one certainly needs to have some skills; sufficient knowledge can significantly reduce communication costs.

Moreover, the AI field is booming, and no major company dares to fall behind. Their attention to the industry is no less than that of investors and practitioners in the field. Scale AI, one of the hottest unicorns in Silicon Valley, does not make models or products, but focuses on only one thing - data annotation. They can be said to be the shovel sellers of the A1 era.

Based on this single business, Scale AI's valuation has reached tens of billions of US dollars.

What does this mean? It's almost at its Dingsheng.

So Liu Dahai flipped through several pages of the document, skipping the prepared introduction, and went straight to a comparison chart.

In June 2023, Microsoft Research published a paper called "Tetbooks Are All You Need," which used carefully selected textbook-level data to train a small model, Phi-1, with 1.3 billion parameters. This model outperformed Large Model on the market with ten times the number of parameters and a hundred times the amount of training data in code generation tasks.

Cheng Yuan interjected, "Dr. Liu, what does this mean?"

"At the time, no model except GPT could outperform it," Liu Dahai said. "It only had 1.3 billion parameters, and it took four days to train with just eight A100 maps—that was the cost. This paper proved one thing: if the data quality is high enough, a small model can beat Large Model."

He turned to Zheng Xiaobo.

"This perfectly matches the situation with Glutinous Rice Ball. 7B, which is seven billion parameters and ten thousand training data points, aims to understand how it defeated Kun Yuan. The key is not the size of the model, but the quality of the data."

In recent years, this has almost become a consensus in the AI industry: the higher the quality of the annotations, the stronger the model. The architecture doesn't change much, and the marginal benefits of Computing Power diminish, but there is still a long way to go in improving high-quality data annotations.

The problem is that everyone knows data labeling is important, but no one knows how good the data labeling needs to be to be considered sufficient.

OpenAI, Anthropic, and Meta invest billions of dollars in data annotation every year, and while their models are indeed getting better with each generation, the progress is slowing down. Nobody knows when they will hit a wall, and nobody knows where the limits are.

No matter how hot the capital speculation gets, everyone in this industry is groping their way forward on a vast, boundless sea.

It wasn't until he saw the Glutinous Rice Ball that Liu Dahai was certain he had seen a beacon of the future.

Zheng Xiaobo remained silent.

Liu Dahai slowed down his speech and lowered his voice by half an octave.

"Judging from the model's performance, their annotation accuracy is probably the highest level I could imagine in all my years in the industry. Not just a little higher, but an entire order of magnitude higher."

The office fell silent again.

Zheng Xiaobo tapped his fingers lightly twice on the armrest.

"Where did his data come from?"

As soon as the question was raised, Liu Dahai and Cheng Yuan both looked at him.

Zheng Xiaobo's expression remained unchanged.

Liu Dahai spoke first:

"Boss Zheng, let me do some math for you. The top-tier annotation service in the industry right now, Scale AI's expert-level annotation, costs about $70 or $80 per line. But that level of precision is definitely a step below Glutinous Rice Ball's data. If you were to buy data at Glutinous Rice Ball's level of precision, there simply isn't such a service on the market. If I had to give a price estimate, each line would have to be at least $100. They're talking about less than 10,000 lines of data. Even if we say 10,000, the data cost alone would be a million dollars."

"This level of precision in annotation couldn't possibly be done manually by one or two people. It would require either a professional annotation team, a proprietary data source, or collaboration with a large research institution."

Zheng Xiaobo looked at Cheng Yuan and gave the order:

"Have the business intelligence team investigate Source Code Technology's business registration changes and publicly disclosed partnerships to see if they've had any dealings with annotation companies or data agencies in the past six months. Also, check their job postings to see if they've hired anyone related to annotation."

Cheng Yuan nodded and quickly jotted down a few notes in the blank space inside the folder.

"Find his data supplier," Zheng Xiaobo said, "And poach him directly."

Liu Dahai added, "If we can find a standard annotation document, that's more valuable than the data itself. With a standard, at least we know where we're headed, and we can try to do it ourselves."

Zheng Xiaobo glanced at him and nodded slightly.

"The second thing," he said, turning his gaze back to Cheng Yuan, "Is that we can start discussing Computing Power cooperation."

Cheng Yuan looked up, as if he hadn't quite figured it out himself. "Let's talk about cooperation."

"However, there's a condition: the cooperation framework must include a clause for one-to-one sharing of training data."

"What if Han Luyi doesn't agree?" Cheng Yuan asked.

"Let's talk, but don't put on too high a pedestal." Zheng Xiaobo's voice was strong. "He can get Computing Power from others, but we can't find another company that can provide this data for the time being."

"Once we find his supplier, there's no need to negotiate anymore."

Cheng Yuan looked enlightened and quickly memorized the sentence.

Liu Dahai watched the whole thing from the side, thinking that the two of them were chatting like they were performing a charade. This skill of flattery is harder to learn than machine learning! Zheng Xiaobo swiveled his chair to face the floor-to-ceiling window.

The Huangpu River makes a bend below, and the office buildings on the opposite bank are faintly visible in the winter mist.

He turned his back to the two people and asked one last question.

"Ocean, tell me the truth."

Liu Dahai watched his retreating figure.

"If we get their data, how much can Kun Yuan improve?"

Liu Dahai did not answer immediately.

He thought for a few seconds.

As someone who has been immersed in the machine learning industry for over a decade, he Supreme Purity what 85.2 means. Kun Yuan's interpretation is that 78 is already among the top tier in China. In other words, it takes over 200 people in the Dingsheng model team half a year to get to the point where 78 and 85 are the seven points that they couldn't even get close to. But those two people, oh right, Han Luyi, did it with just 10,000 data points.

The key is not how good his model is, but how good those 10,000 data points are.

The more he thought about it, the more excited he became.

"Boss Zheng, if we could get his Annotated Data with that level of precision, even just 50,000 records would be greatly appreciated."

He paused for a moment.

"Kun Yuan can break through the ceiling of this industry. Then our Dingsheng model will be number one in the world." Zheng Xiaobo turned around, the gray-white light from the window outlining his silhouette, a faint smile on his face.

"Then I'll wait and see."

"Cheng Yuan, go and handle it." Zheng Xiaobo picked up the bottled water on the table and took a sip. "Don't give Han Luyi a reply about the Computing Power cooperation yet; tell him to wait a couple of days."

Cheng Yuan was taken aback for a moment. "Boss Zheng, this is..."

"Remember, at the negotiating table, the one who speaks first loses."

Meanwhile, Source Code Technology.

Han Luyi sat in his office, opened the Sight, and spent twenty minutes annotating more than three hundred data points.

He glanced at his energy level and yawned.

Just then, Han Luyi's phone rang. He picked it up and saw it was a message from the [Dorm 404] group chat; someone had tagged him.

A while ago, Ma Xiaofei and Zhang Haoran were spamming the group chat with messages about watching the game, but he had set his messages to "Do Not Disturb" and hadn't checked them for a few days.

Han Luyi through the chat history.

[Ma Xiaofei]: What are you all doing, sons? It's been so long since we last saw each other, wanna get together?

[Zhang Haoran]: Half-dead, barely alive.

[Zhang Haoran]: You media people are so energetic!

[Ma Xiaofei]: What's wrong? A secure job like this can't possibly make you tired?

[Zhang Haoran]: Just kidding. The bank is now encouraging everyone to attract deposits, and my performance evaluation for this quarter is still pending.

[Ma Xiaofei]: They're even targeting you technical staff?!

[Zhang Haoran]: You've come at the right time! Help me out, bro! Do you have any wealthy connections?

[Ma Xiaofei]: I'll manually find you a rich guy, Han Luyi

[Ma Xiaofei]: Lao Han, what kind of group nickname is that? Are you trying to take advantage of me?

Han Luyi finished climbing the stairs and slapped his forehead.

I've been so busy with work lately, I completely forgot to take my sons flying! I'm so awful.

[My Father Han Luyi]: How about this weekend? See you at the usual place?

[Zhang Haoran]: I would come if I weren't working overtime.

[My father Han Luyi]: Sigh, I happen to have some spare money that I want to save...

[Zhang Haoran]: Godfather! See you there!

End of Chapter
Prev
NovelSummaries
Next

Explore the wiki

Han Luyi

Comments

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