Upon hearing this question, Lin Shaofeng took a deep breath, smiled slightly, and said, "Of course."
Live demonstrations are risky because you can always discover things that you can't test during actual testing.
Unless one has sufficient confidence in their product, they generally won't do a live demonstration.
However, on the other hand, this is an internal meeting. Even if there's a problem during the demonstration, it's just an internal embarrassment. But if there's a problem during a live demonstration at a press conference, it will make the news and trend on social media.
Lin Shaofeng glanced at Lü Yun and Zheng Xiaobo again; both of them were staring intently at him. He exhaled the breath he had just taken and opened Kun Yuan's interactive interface.
Since a live demonstration was to be conducted, he had obviously prepared the materials privately many times beforehand. However, the characteristic of Large Model is that they are unpredictable, as those who know, know.
Input: [It's almost Double Eleven' Day, could you please go through this inventory preparation plan for me and check for any problems?]
Then I uploaded an attachment, which was a stock preparation plan document, including inventory plans and warehouse allocations for dozens of SKUs.
All the key logistics timelines are included.
After Kun Yuan processed the data, not all risks were listed in the output.
Instead, only three items were listed.
Based on last year's sales curve, the inventory of top-selling SKUs in the main warehouse is estimated to be insufficient to prevent stockouts during peak periods. The logistics timeline from the South China warehouse to the main distribution area falls on the evening of the first day, which may lead to delays. Additionally, there are loopholes in the pricing of two promotional packages, resulting in some items being sold below cost price after additional discounts.
Each item is followed by repair suggestions, all of which are specific actions that can be completed within a few days.
In contrast, Lin Shaofeng sent the same attachment and the same prompt to ChatGPT.
ChatGPT's reply was extremely detailed—from warehouse layout to long-term supplier relationships, from traffic prediction models to customer service contingency plans, it was a lengthy document with over a dozen points. Putting aside whether the answers were correct, the sheer volume of text was enough to make one's scalp tingle and not want to read it.
Lin Shaofeng placed the two outputs side by side on the large screen.
"These are real case studies that our researchers obtained from the e-commerce department. After reviewing the output, they felt that Kun Yuan's output was better." Lin Shaofeng then turned to the VP of the e-commerce department, "Mr. Xu, what do you think?"
Mr. Xu was taken aback when he was called out, but he reacted quickly, smiled, and replied, "Of course, from my perspective, Kun Yuan's answer is clearly more practical. I hope to arrange it for all my colleagues in the operations team right now."
A soft chuckle filled the conference room, and even Lü Yun had a slight smile on his face.
"Of course, GPT can do this too."
Lin Shaofeng changed the subject, saying, "But to achieve this effect, the prompts have to be written like this."
He switched screens to another document, and a prompt appeared on the screen:
Please analyze the following inventory preparation plan, requirements:
List only the problems that can be resolved within three days;
Ranked from highest to lowest based on their impact on peak sales during Double Eleven' Day;
Each question is followed by specific actionable suggestions;
[Ignoring structural problems that require long-term optimization.]
"Following this prompt, GPT will most likely produce similar results."
Lin Shaofeng paused for a moment and looked around at the people present.
"But the problem is, this prompt was written by researchers and operations specialists in twenty minutes," he said. "Most users are just thinking: 'Double Eleven is coming soon, please check if there are any problems.' They don't know to say 'Only check those that can be resolved within three days,' 'Sort by impact,' or 'Ignore structural problems.' They just want a useful answer directly."
The meeting room fell silent, and some people's expressions clearly showed they were deep in thought.
The user said, "It's almost Double Eleven, please go through this for me." He didn't specify that he should only look at issues that could be resolved within a few days, only those affecting peak sales, or that the conclusions should be actionable. But Kun Yuan understood what "almost" meant in this time window, and what kind of question "are there any problems?" was asking in the context of Singles' Day preparation and stocking up.
"This is an understanding of Chinese business scenarios, an understanding that breaks free from the limitations of prompt words," Lin Shaofeng summarized.
In this comparison, even if both are equally capable of solving complex problems, the Kun Yuan model has already won in the area of interaction.
If this were a public speech, there might be applause, but this is an internal meeting, so no one is clapping.
After the demonstration, the marketing manager did not raise any further questions.
He was already thinking about what kind of catchy slogans to use.
This time, there's something real.
Lin Shaofeng had actually prepared several other demonstration examples, but since no one suggested them, he didn't use them. It seems that the previous demonstration was enough to prove Kun Yuan's value.
After a pause, Lin Shaofeng added another sentence.
"There's one more thing." He glanced at the VPs of various business units and heads of business lines present, "Kun Yuan is independently developed and deployed, and the data is all on Dingsheng's own Server, which solves the biggest hidden danger. Are you comfortable transferring the annual Double Eleven inventory preparation plan and pricing strategy to external AI?"
Hearing this, Lü Yun nodded slightly, then finally asked his first question, which was unrelated to the demonstration. "Who is the technical lead for this project?"
Lin Shaofeng hesitated for a moment. The direction and progress of the technology were determined by Liu Dahai, but in the final stage of training, the two disagreed on the judgment of data quality, so he reported to Zheng Xiaobo and was transferred to another project.
Although Lü Yun wouldn't pass through the approval process for this level of change, Lin Shaofeng didn't believe he knew nothing about it.
"The technical architecture and direction of Kun Yuan were mainly developed by Chief Scientist Dr. Liu Dahai and his team," he added. "The entire team has been working very hard recently to bring Kun Yuan to fruition."
"Hmm, then why wasn't he there?" Lü Yun said, glancing at Zheng Xiaobo.
Lü Yun's voice wasn't harsh, but Lin Shaofeng's heart skipped a beat.
Before Lin Shaofeng could answer, Zheng Xiaobo spoke first, "Dr. Liu has recently been transferred to another project to tackle the challenges. Kun Yuan is nearing completion, and there are many trivial matters to attend to, but the overall progress of the technical team has not stopped."
Lü Yun added, "Next time this kind of presentation is given, the technical lead should still be present."
Without waiting for a reply, he changed the subject. "The completion rate is acceptable, let's talk about the release plan."
Zheng Xiaobo received the signal, stood up, and walked to the screen.
Lin Shaofeng consciously stepped aside and used his computer to bring up the slideshow that had just displayed the data.
Pointing to the evaluation data on the screen, Zheng Xiaobo said, "This press conference shouldn't be treated as a regular technology release. Everyone is welcome to share their thoughts."
Then he turned to the head of the marketing department. "What are your thoughts on the marketing side?"
The head of the marketing department stood up. "Boss Zheng, according to international practice, we'll release the technical report and evaluation data first, then hold an online live-streamed press conference, arrange for media to provide in-depth analysis, and invite ecosystem partners to participate."
After listening, Zheng Xiaobo shook his head. "Too mediocre."
He glanced at Lü Yun, then addressed everyone. "There are no Large Model in China that can stand in the first tier right now. The fact that Dingsheng can bring out Kun Yuan today is a declaration: Dingsheng was the first in the Internet era, and Dingsheng will still be the first in the AI era."
Everyone understood that Zheng Xiaobo had high hopes for Kun Yuan, intending to use it to put Dingsheng at the center of the global AI narrative, on par with OpenAI and Google.
Moreover, with Lü Yun present, Zheng Xiaobo's words in front of him indicate that this was also Lü Yun's intention.
"We'll be holding a press conference in a week. Let's start building hype now, releasing Kun Yuan's test data, but don't reveal that Kun Yuan is our model. We'll announce it on the day of the press conference," Zheng Xiaobo said. "On the day of the press conference, we'll ensure the media and partner companies are all there at the highest level; we want to tell a compelling story."
At this point, Zheng Xiaobo's expression turned serious. "In the past, Large Model could understand instructions from the English-speaking world, but not the complex expressions of the Chinese-speaking world. Kun Yuan aims to solve this problem."
He wants to connect modeling capabilities with national sentiment, industry sentiment, and workplace pain points.
The theme of the press conference was "Let AI Understand China."
The room fell silent upon hearing this.
Lin Shaofeng was also shocked after hearing it; the narrative was too powerful.
At the same time, he was also a little excited - the China's first and world-leading Large Model, with the world's best Chinese intention understanding ability, was created by the team he led. This idea made him excited.
Lin Shaofeng glanced at Cheng Yuan, who was sitting in the conference room.
The key Annotated Data used to improve the model's capabilities was shared by Source Intelligence Technology. Are they also training their own Large Model?
On second thought, Model Training involves resources and capabilities from various aspects, not just Annotated Data. Even if everyone's labels are the same, the upper limit of the model's capabilities will be different. Kun Yuan has devoted its Dingsheng strength to this for more than half a year, and he has confidence in Kun Yuan.
Thinking of this, his earlier excitement overshadowed the slight worry lingering in his heart.