Chapter 255: Bai Qian’s Wish

Chapter 255 -192: Bai Qian’s Wish

Time flew, and soon it was Monday.

At two o’clock in the afternoon, in the reception area of Yanhai Company.

Chen Lin sat on the sofa. Across from him was a young man who looked to be in his mid-twenties, with a slender build and a pair of black-framed glasses.

This was Bai Qian, a Postdoctoral Researcher from the National Academy of Sciences Automation Institute.

At this moment, Bai Qian was wearing a slightly loose-fitting plaid shirt, jeans, and a pair of sneakers that looked like they had been worn for several years.

His entire demeanor perfectly matched all of Chen Lin’s stereotypes of a “nerdy academic.”

When they met, Bai Qian seemed rather quiet and reserved.

Chen Lin took the initiative to extend his hand, smiling as he said, “Dr. Bai, hello. I’m Chen Lin.”

Bai Qian paused for a moment, then quickly extended his own hand to shake Chen Lin’s. His voice was a little quiet. “Hello, President Chen.”

As they shook hands, Chen Lin could clearly feel that Bai Qian’s palm was a bit damp, obviously from nervous sweat.

Chen Lin smiled and gestured for Bai Qian to sit on the sofa.

Yunqing came over with two cups of tea. Beaming, she handed one to Bai Qian. “Dr. Bai, please have some tea.”

The moment Bai Qian took the teacup, his cheeks flushed slightly and his voice grew even quieter. “Thank you.”

Seeing this, Yunqing couldn’t help but chuckle softly before turning and leaving the reception area.

Chen Lin didn’t pay any mind to these details.

He cut straight to the chase. “Dr. Bai, let’s not beat around the bush. Let’s get right down to business.”

“Elder Xiang has probably already given you a general idea of my situation, right?”

Bai Qian nodded and pushed up his glasses. “Yes, Elder Xiang’s assistant told me. You’re planning to develop a brand new large AI model yourself, is that right?”

“That’s right,” Chen Lin confirmed with a nod. “And not just by fine-tuning an existing open-source model. We’re going to train a completely new model from scratch.”

“The goal is to comprehensively surpass all current mainstream large models on the market in terms of content generation quality, logical reasoning ability, and control over the hallucination rate.”

Chen Lin paused here before adding:

“Of course, I know this goal sounds exaggerated and very difficult. But I believe that as long as we have the right direction, and the funding and talent are in place, it can definitely be done.”

After hearing this, the reserved expression on Bai Qian’s face gradually turned serious.

He pondered for a moment before speaking slowly.

“President Chen, may I ask if you have any specific ideas about the model’s architecture?”

“For example, are you planning to use a Transformer, or an MoE (Mixture of Experts) model, or do you have another approach in mind?”

Hearing this question, Chen Lin shook his head frankly.

“To be honest, Dr. Bai, my knowledge in fields like computer science, programming, and artificial intelligence is still at a rather elementary level.”

“The only thing I’m an expert in is mathematics.”

“So, regarding the choice of model architecture, I’d like to hear your professional opinion.”

Upon hearing this, Bai Qian’s eyes immediately lit up.

Those eyes, which had seemed somewhat evasive before, now seemed to be sparkling with light.

His entire demeanor changed in an instant.

“President Chen, since you’ve asked, I’ll be direct.”

Bai Qian pushed up his glasses, his speaking pace quickening noticeably.

“I personally believe that the two current mainstream architectures for large models, Transformer and MoE, are still far from optimal.”

“While Transformers have advantages in parallel computing and capturing long-range dependencies, their parameter utilization efficiency is actually very low.”

“And while MoE increases model capacity by introducing multiple expert modules, its training difficulty and inference costs are extremely high.”

“More importantly,” Bai Qian’s tone grew more and more excited, “I think the basic training paradigm of ‘pre-training plus fine-tuning’ that’s commonly used in the industry also has significant room for improvement.”

“In the pre-training phase, the model learns a general language representation ability through self-supervised learning on massive amounts of unlabeled data.”

“In the fine-tuning phase, it undergoes supervised learning on labeled data for a specific task, allowing the model to adapt to particular application scenarios.”

“However,” Bai Qian paused here, looking at Chen Lin with a burning gaze, “this paradigm has one major problem—”

“It assumes that the knowledge learned during the pre-training phase can be seamlessly transferred to the fine-tuning phase.”

“But in reality, there’s often a gap between pre-training and fine-tuning.”

“This leads to situations where, even if you invest astronomical amounts of computing power and data in the pre-training phase, the model’s performance on certain specific tasks is still less than satisfactory.”

Hearing this, although Chen Lin didn’t fully understand many of the technical terms, he could tell that Bai Qian had a very deep understanding of this field.

He nodded slightly, gesturing for Bai Qian to continue.

When Bai Qian saw that Chen Lin wasn’t interrupting him and was instead listening intently, he became even more excited.

His speech grew faster and faster as he gesticulated animatedly.

“So, for the past few years, I’ve been thinking about one question—”

“Is it possible to design a completely new model architecture, and a completely new training paradigm, that could fundamentally solve these problems?”

“I’ve done some preliminary research and have had some results.”

“However,” Bai Qian’s tone suddenly became tinged with helplessness, “at the Automation Institute, my research direction isn’t mainstream.”

“Most of the institute’s resources are allocated to projects that follow the industry’s mainstream paths.”

“I can’t get the computing power or the data that I need.”

“So, many of my ideas can only remain in the theoretical stage, with no way to actually validate them.”

At this point, Bai Qian suddenly looked up, his eyes flashing with an almost fanatical light.

“But, President Chen, after I saw your demonstration at the mathematics conference…”

“I suddenly felt that with you, I might have a chance to realize my dream.”

“To find a better large model architecture!”

Chen Lin was moved by these words.

He could sense that Bai Qian had an almost obsessive passion for the field of artificial intelligence.

People like that often have the most potential to make breakthrough achievements.

Chen Lin pondered for a moment before speaking slowly.

“Dr. Bai, could you briefly tell me about the research results you just mentioned?”

Upon hearing this, Bai Qian immediately nodded, the light in his eyes shining even brighter.

“Of course!”

He pulled a tablet out of his backpack and quickly opened a folder.

“President Chen, take a look…”

For the next half hour, Bai Qian talked nonstop, presenting his research findings to Chen Lin.

“Dynamic sparse activation mechanisms,” “hierarchical knowledge distillation,” “multimodal alignment pre-training…”

A whole slew of technical terms poured out of Bai Qian’s mouth, leaving Chen Lin completely bewildered.

Although Chen Lin couldn’t understand many of the details, he could sense that what Bai Qian was talking about had incredible depth.

So, Chen Lin maintained an attentive posture the entire time, occasionally looking at Bai Qian with an appreciative gaze, nodding frequently to offer him emotional validation.

Finally, Bai Qian finished speaking.

He seemed to realize he had said a bit too much and scratched his head, looking a little embarrassed.

“Sorry, President Chen. Once I get started on this topic, I have a hard time stopping…”