The two arrived at Han Luyi's office, ready to discuss strategies for dealing with the video conference they had just attended.
Just then, there was a rapid knocking at the door.
"Boss Han, are you there?"
It was Zhao Wenyuan's voice.
Han Luyi and Su Niannian exchanged a glance.
Han Luyi walked over and opened the door. Zhao Wenyuan was holding a laptop and looked excited.
"Boss Han! You're amazing!" he said as he walked in, only to suddenly notice that Su Niannian was also in the room, and he was taken aback.
His expression immediately relaxed, and he placed his laptop on the desk.
"Perfect timing, President Su, you're here too. Let's take a look together," Zhao Wenyuan said. "The results of the first round of training are in."
He flipped open the screen, his fingers swiping rapidly across the touchpad to bring up a table.
"Let's look at the basic data first." Zhao Wenyuan switched to the first page. "The 7B's open-source foundation has been fine-tuned, with less than one-twentieth the number of parameters of GPT-4. Knowledge-based question answering, logical reasoning, mathematical calculations, and code generation—the standard metrics are above average for models of the same scale, nothing surprising. 7B is just 7B; it can't compare to the hundreds of bytes of big stuff."
He glanced at Han Luyi, anticipating his reaction.
The important part is on the next page.
Zhao Wenyuan turned to the second page.
The screen displays a bar chart. The horizontal axis lists the names of more than a dozen models, and the vertical axis represents the scores.
Han Luyi recognized several: GPT, Mercury, and Gemini.
The pillars next to these names are about the same height, and the scores are concentrated between 71 and 78.
On the far right, there is a single light blue pillar, which is noticeably taller.
The model name is PX-7B-post-trained-v0.1.
The number at the top is 83.7.
"The evaluation results for the intent understanding dimension." Zhao Wenyuan's voice was low, but his speech was rapid. "The evaluation set uses the IFEval standard framework plus my own designed deep intent reconstruction test, totaling 1,200 questions, covering sixteen vertical scenarios."
He pointed to the light blue pillar with his finger.
"83.7, leading by a significant margin."
Second place, GPT, 78.2.
Third place, Gemini, 76.9.
Han Luyi glanced at the Mercury 3.1—the one Ryan had just shown—and saw a score of 74.8. The office fell silent for three seconds.
Han Luyi didn't speak. Neither did Su Niannian.
"I ran the validation three times," Zhao Wenyuan said. "The first time, I thought the evaluation code was wrong. The second time, I changed the evaluation framework, the scores were different, but the results were sorted the same. The third time, I shuffled the evaluation set and did cross-validation, and the error was within 0.3 points."
He turned to the next page. The screen displayed specific case comparisons—the left column showed user input, the middle column showed the understanding results of each model, and the right column showed the labeled true intent.
Zhao Wenyuan flipped through several pages of case studies, not explaining each one in detail, but deliberately spending a longer time on each page, as if to check if Han Luyi and Su Niannian were keeping up.
Han Luyi caught up.
Intention understanding, simply put, is whether a model can understand what a user truly wants. The content a user types is the surface; beneath the surface lies the situation, the constraints, and the unspoken real needs. Most models simply understand the literal meaning and then provide a correct but general answer. What PX-7B does is read the specific person behind the literal meaning from the same input.
This is an extremely difficult task. Even models with ten times more parameters may not perform well, because it tests not knowledge reserves, but understanding of human beings.
Zhao Wenyuan closed the case study page and returned to the bar chart.
Zhao Wenyuan uttered these words solemnly.
"Boss Han, I've been in this industry for so many years, and I've never seen a 7B chip outperform GPT-4 in any single dimension, never." "And this is only the first round, just a few epochs, the hyperparameters aren't even optimal yet. With a few more rounds of adjustments, there's still room for improvement."
Su Niannian hadn't spoken until now, when she softly said, "Intentional understanding... isn't that what Open Things has always wanted to do?" Han Luyi nodded, seemingly relieved.
Brother Sight, you're awesome!
After a moment's thought, Han Luyi asked a question.
"What if we had 100,000 data points with this level of precision?"
Zhao Wenyuan looked up at the ceiling and thought for a while.
"One hundred thousand annotations of this precision," he said slowly, "Are theoretically enough to push intent understanding capabilities above the current industry ceiling on a 70B or even larger platform."
"This isn't just about improving by a few percentage points; you can imagine it's a generational leap from GPT-3 to GPT-4." "And it's not just about intent understanding. Intent understanding is a fundamental capability of the model. Once the model truly learns to understand human language, the performance of all downstream tasks will improve accordingly: answers will be more accurate, code will be more tailored to needs, content will be more targeted, and user satisfaction will increase significantly."
He stopped, realizing he had gone too far.
"That's in theory," he said, taking a step back. "In practice, we also need to consider data distribution, training stability, and generalization performance. We can't just rely on simple linear reasoning, but the direction is definitely correct."
He continued, "With this level of quality Annotated Data, any Large Model company could use it for alignment training and it would significantly improve intent understanding."
"Boss Han, I don't have much experience in fundraising, but I can assure you that we won't lack funding."
Han Luyi and Su Niannian exchanged a glance, both thrilled by the good news. The pressure from NeusAl had eased somewhat. "After you've finished adjusting the parameters and created an optimal prototype, how much more time do you need?" Han Luyi asked.
"About a week," Zhao Wenyuan said, his face showing reluctance. "But it will take up work time."
This is the gray area that Han Luyi and Zhao Wenyuan discussed before: Source Code Technology can claim ownership of the results made using work resources during working hours. If Hongyuan really falls out with Han Luyi, this could be a hidden danger.
After thinking for a while, Zhao Wenyuan said, "I'll take a week of unpaid leave first, to finish the project."
Han Luyi nodded. "No need to go through Shen Congyun, I'll approve it myself."
Zhao Wenyuan was about to get up and take his leave Han Luyi called him back. Han Lu turned to look at Su Niannian before saying, "Since the training results are good, I have something to say while you two are both here."
The two of them looked at Han Luyi.
"I've thought about the financing carefully. Since what we lack most is Computing Power, there's no need to go through the long process of investors; we can directly find people with Computing Power invest in us."
"Among the major players in China right now, Dingsheng is still the most powerful in this field. There was some unpleasantness between Source Code and Dingsheng in the past, but I hope you won't hold any grudges."
Han Luyi finished speaking, he glanced at the two of them and added, "Of course, this is just my initial idea. You are welcome to offer any suggestions." Su Niannian spoke first, "I have no problem with it. Business is business. As long as we can make the model as quickly as possible, it doesn't matter who we cooperate with, and..."
She glanced at Zhao Wenyuan: "You don't actually need to ask my opinion. Regardless of who you collaborate with, maintaining control is the most important thing." After hearing Han Luyi's question, Zhao Wenyuan remained silent for a longer time. Finally, he spoke, "Does it have to be done domestically? I'm using Google..." Han Luyi interrupted him, "Wenyuan, the training of Large Model must be done domestically!"
Zhao Wenyuan looked up at him and found that he was also staring at him. The two looked at each other for a moment, and then Zhao Wenyuan nodded. "... I understand, I'm fine."
"In that case, I'll handle the rest," Han Luyi said.
After seeing the two out of the office, Han Luyi took out his phone and found Cheng Yuan's WeChat.
After they met and added each other on WeChat, Cheng Yuan sent a greeting. "Boss Han, I'm Dingsheng Cheng Yuan Dingsheng. We hit it off immediately today. Let's keep in touch, and let me know as soon as possible if there's any potential for cooperation."
Han Luyi has not replied.
More than half a month later, he took the initiative to send a message.
Is it embarrassing?
It's not embarrassing.
What's so awkward about doing business?
It's all about self-interest.
"Mr. Cheng, I'm sorry, I just saw this. I have an idea for a new project, would you like to discuss it sometime?"
Send.