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 442

Not enough data

1,275 words 4 min read 1 view 1mo ago Source ​
Skip to the end of the chapter
Prev
Next

It was 6:00 AM on Friday morning in the dormitory for doctoral students at Shuimu University in Beijing.

Zhang Wenduo opened his eyes as his phone alarm rang.

After turning off the alarm, getting up, washing up, and getting everything ready, Zhang Wenduo was already sitting in front of the computer about ten minutes later.

This was still his routine from when he was pursuing his doctorate. Now, he works at Source Intelligence Technology —oh no, it's now called Origin Group. Although the Research Institute is still called Source Intelligence Beijing Research Institute, Zhang Wenduo still maintains his routine of going to bed at 11 pm and waking up at 6 am every day, because he firmly believes that self-discipline is the key to success.

I used the time after waking up to read papers, because after a night's rest, my mind was at its clearest.

Why do you ask why someone who has already graduated with a PhD still needs to read dissertations?

Because graduating with a doctorate is not the end of learning, but only the Starting Point of scientific research.

Around 7:30, Zhang Wenduo would go to the cafeteria for breakfast. He would order a tea egg, two boiled eggs, and white porridge. Breakfast needed to provide enough protein to ensure efficient work in the morning. Boiled eggs were too bland, but eating only tea eggs would exceed the recommended sodium intake in the Dietary Guidelines for Chinese Residents, so a balance was just right.

Then Zhang Wenduo would ride a shared bicycle to the Research Institute's office, which was only three kilometers away from the Shuimu Forum University campus. The ride took about ten minutes, faster than taking the subway or bus, and it was also good exercise—a win-win situation.

I arrived at the office at exactly eight o'clock, and before the nine-hour workday, I could do some research that interested me.

Zhang Wenduo enjoys this regular lifestyle, but he has two worries lately.

The first issue is about accommodation. He has already graduated with his doctorate and started working. His student apartment is only available until mid-July at the latest, and he is currently...

It's already late June, and finding a new place to stay is the most urgent task.

However, no matter how hard you look, it's hard to find an option with such a high cost-performance ratio as student apartments.

What's more, moving out means there's no cafeteria anymore.

The second problem, which was related to his work, caused Zhang Wenduo even more headaches.

Since the new asynchronous algorithm architecture he proposed last time greatly improved the training efficiency of the model, Zhao Wenyuan kept his promise and approved 128 L100 training cards for him, making Zhang Wenduo's research group the richest in the entire Research Institute.

Zhao Wenyuan also assigned him two doctoral students who were interning there, Zhang Wenduo, who had just graduated, also became the group leader.

Zhang Wenduo's main research direction is World Model, which is slightly different from the research direction he conducted during his doctoral studies. At that time, he believed that expressing the world itself in textual form had limitations, and that more innovative algorithms were needed to smooth the mapping from WeChat to high-dimensional space in order to more efficiently establish the world state inside the model.

However, after arriving at the Source Intelligence Research Institute, the research based on Glutinous Rice Ball completely overturned Zhang Wenduo's previous assumptions.

As a Large Language Model, Glutinous Rice Ball does not possess native multimodal capabilities. However, by providing sufficient high-precision and high-quality Annotated Data, the internal world stability of Glutinous Rice Ball 1.0 surpasses that of many cutting-edge studies exploring World Model.

Therefore, Zhang Wenduo's research topic is to study the extent to which World Model can be constructed using Large Language Model, that is, pure text, based on Glutinous Rice Ball 1.0.

His research process wasn't particularly complex, just quite tiring. It mainly involved testing which data points were more important and crucial in the fine-tuned Annotated Data of Glutinous Rice Ball 1.0, in order to provide guidance for future data annotation directions.

To put it simply, it's still the process of alchemy. It's about adding this label or that label, and then using a probe to test the stability of the internal state after adding different labels.

These operations are all dirty and tiring work, and do not require doctoral-level knowledge. Even a junior high school student who is good at using computers can do them.

The part that really requires brainpower is designing improvement plans, proposing hypotheses, and using experiments to verify those hypotheses.

Sometimes a theory can explain several experimental results, but when a new experiment is run, the results deviate from the theory. In such cases, the theory must be adjusted, and a meaningful adjustment direction for Model Training must be designed based on the temporary theory.

This is the kind of work only a PhD can do; even AI can't do it yet.

Of course, the true AGI era will only arrive when AI can evolve on its own.

Zhang Wenduo was the first to arrive at the office again. He had just sat down when the door opened again and Zhao Wenyuan walked in.

Zhao Wenyuan wasn't surprised to see Zhang Wenduo; the two of them were always the first to arrive at the office, sometimes he was earlier, sometimes I was earlier.

"Good morning, Dr. Zhang," Zhao Wenyuan said.

"President Zhao." Zhang Wenduo nodded from his seat.

Zhao Wenyuan didn't go into his own office, but went straight to Zhang Wenduo's workstation and asked, "How's the project going lately?"

Zhang Wenduo's face darkened, and he said, "It's not going too smoothly."

This is his second problem: his research on the World Model based on Glutinous Rice Ball 1.0 has hit a wall.

Zhao Wenyuan, intrigued, pulled up a chair and sat down, asking, "I remember there was some new progress last week. What problems have you discovered this time?"

Zhang Wenduo's group is focused on long-term strategic goals and is not under pressure to produce results in the short term. Moreover, their previous research has already made considerable progress, so Zhao Wenyuan is not worried. This concern stems from the leader's care for his employees rather than his concern for the project itself.

Zhang Wenduo said, "President Zhao, the results of the recent rounds of probes show that the capabilities have not only not improved, but have actually decreased somewhat."

Zhao Wenyuan then asked, "What do you think is the reason?"

Zhang Wenduo thought for a moment and said, "I think the data is insufficient."

Zhao Wenyuan looked at him in surprise.

The number of parameters and the proportion of annotations for Glutinous Rice Ball are not low, not to mention the level of detail and quality of these annotations.

In conclusion, thank you, Boss Han.

Is Zhang Wenduo not enough?

"What data are you missing? I'll help you figure it out," Zhao Wenyuan said.

It is a leader's responsibility to help subordinates solve problems when they encounter difficulties. This applies to Zhao Wenyuan himself, and it also applies to Boss Han, right?

Anyway, I'll take on the work first, Boss Han will find a way to solve it.

Zhang Wenduo glanced at Zhao Wenyuan, wondering why he dared to make such a bold claim without even asking what the data was, showing such confidence.

But he still explained, "It's data about people, more precisely, data about interactions between people."

It would have been better if he hadn't explained; the more he explained, the less Zhao Wenyuan understood.

Current Annotated Data is mostly one-to-one input and output pairs. What does this data on interactions between people mean?

End of Chapter
Prev
NovelSummaries
Next

Explore the wiki

Zhao Wenyuan

Comments

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