Chapter 9: The First Group Meeting

When Zhou Yun woke up the next morning, he felt a dull ache in his lower back.

"I shouldn't have pushed myself so hard last night."

In an effort to maintain a healthy physique, Zhou Yun had gone for a run after dinner the previous evening. Having run five kilometers—perhaps because he hadn't run in so long—he hadn't felt much at the time, but now that he was getting out of bed, the soreness was quite intense.

The monitor he had ordered arrived, so after a quick breakfast, he picked it up. Looking at the 27-inch screen, Zhou Yun nodded with satisfaction.

"Now, that’s more like it."

Deng Yonghua had already sent him the server’s account information and password the day before. He first downloaded Termius; although Windows comes with a built-in SSH connection tool, he preferred Termius. Mostly, it was because the interface looked more pleasant—young people always have a soft spot for flashy UIs.

After a few quick steps, he successfully connected to the server. The server ran on a Linux system, which typically lacks a GUI (Graphical User Interface) to minimize resource consumption and enhance security. It can feel awkward at first, but once you get used to it, you realize that in the field of scientific research, Linux is far more convenient than Windows.

He searched for the official Anaconda website on his own computer, copied the link for the latest version, and used `wget` to begin the download directly on the server. Anaconda is the most commonly used, and arguably essential, tool in the AI field. The server’s network speed was quite impressive, and the download finished in no time.

The subsequent workflow was standard: creating virtual environments, installing PyTorch, and setting up the various necessary packages. Setting up the environment alone occupied his entire morning. He took a moment to check the server status; as expected, with eight 80GB H100s providing 640GB of VRAM in total, even with others using the machine, over half the capacity remained—more than enough for his needs.

He had already clearly mapped out the workflow for "AgileEdge" in his mind, so writing the code was a matter of course. Plus, with AI assistance, writing the several thousand lines of code for the core framework took less than an hour, leaving plenty of time for self-review and debugging. By 5:00 PM, he had finished the code for the core framework—a speed that would have been unimaginable just a few years ago.

However, writing code is the simplest part of the entire research process. The real torture was about to begin. Running experiments followed by debugging—this tedious, repetitive cycle was the part that truly drained one's spirit.

Zhou Yun found a large language model with about 100 billion parameters online. Originally, he had planned to settle for one with a few dozen billion, but since his mentor was so generous with resources, he didn't need to be so stingy. AgileEdge’s primary function was to compress large models into smaller ones for use on edge devices. Because it was a fully automated AI framework, the operation was incredibly straightforward: once an edge device was connected to the computer, the program would automatically detect the hardware's specific parameters, then use AgileEdge to compress and continuously optimize the model until a functional model file was produced.

For the edge device, Zhou Yun grabbed an old smartphone with only 32GB of memory and plugged it in. He ran the program. A progress bar appeared at the bottom of the screen:

[Estimated time: 17h]

"I figured as much."

Whenever large models are involved, experiments are never quick, though seventeen hours was actually quite fast. In his previous life, Zhou Yun had once run an experiment that took two weeks, mostly because the servers back then were truly garbage.

Three days later, on Friday: After a few days of experimentation, Zhou Yun’s AgileEdge was functional, but it was still falling short of his expected performance. Furthermore, there was a significant disparity in performance between different large models after compression. His original idea was that he could compress a model without losing performance, but it now seemed that model compression inevitably led to a decline in quality. It appeared to be an unsolvable problem.

Zhou Yun was certain that his AgileEdge framework had no flaws—otherwise, it wouldn't be able to achieve compression at all—so the problem must lie with the models themselves. The unique architectures of these large models limited the extent to which their parameters could be reduced. In simple terms, no matter how you compressed them, there would always be a lower limit, and what Zhou Yun needed to do was break through that limit. This meant he had to modify the model architecture itself.

"It seems I underestimated the difficulty."

Intelligently modifying the architecture of large models was a massive undertaking; this research alone was enough to publish at an A-list conference.

Just as Zhou Yun was deep in thought, a message popped up in the group chat:

[Click to join meeting]

It was time for the group meeting. After joining, Zhou Yun immediately muted his microphone. Once everyone had arrived, the meeting began. In reality, only three people were attending online; the others were in the physical conference room. Being a new student who hadn't started yet was the only reason he had the privilege of joining remotely.

"Can you hear me?" Deng Yonghua’s voice sounded.

"Yes," came the reply from the three online students.

"Let me introduce our new team members: Zhou Yun, Shao Shi, and Chang Chengxuan. Everyone, please feel free to interact with them when you have time. Without further ado, let's begin. Qiu Yan, would you like to go first?"

What followed was a grueling three hours. In a research group, everyone’s focus is rarely the same, which meant that for many of the presentations, the rest of the group—and sometimes even the professor—didn't really understand the specific research being discussed. This was precisely why Zhou Yun detested group meetings. Being forced to sit through three hours of content he didn't want to hear and couldn't understand was hardly an enjoyable experience.

On top of that, he had to endure the professor’s occasional "bizarre questions." For instance: "I don't understand your research field, but I want to ask what your innovation is." This kind of question is enough to make anyone's blood boil, but since it's your advisor, you have to explain it to them with a smile.

Deng Yonghua, however, wasn't too bad. Based on Zhou Yun's three hours of observation, the professor generally didn't ask questions about research fields he didn't grasp. He primarily cared about the students' research progress; if there was progress and the field was unfamiliar, he was actually quite laid back. But if there was no progress and he *did* happen to understand your field? Well, then, you were in for a rough time.

After all the senior students had finished their reports, it was finally the turn of the three new students.

"You three new students, you're prepared, right? Zhou Yun, why don't you start? How is that research you mentioned last time coming along? Or are you planning to present a paper?"

Deng Yonghua had some expectations for Zhou Yun, but he knew it would be difficult to have any substantial progress in such a short amount of time.

"Certainly, Professor. Shall I report on my research progress then?"

In the conference room at Qiantang University, the senior students in the back row exchanged glances and eventually looked at Qiu Yan.

"What's the deal with this new junior?"