Silicon Valley, West Coast, United States.
At 5:57 PM, dozens of employees were sitting in the Main Conference Room at Ne. usAI headquarters.
At this time, most employees should be leaving get off work.
But today is different. At 10:00 AM Beijing time, Prosperity Group is about to hold a press conference.
Converted to the time on the United States West Coast, that's 6 PM.
CEO Ryan had his assistant order dozens of pizzas to organize this event where everyone could watch the live stream together.
Dingsheng is a traditional internet company. In Ryan's mind, it can be seen as a China version of Amazon, with e-commerce, cloud computing, and other businesses penetrating various fields.
However, they are not leaders in AI.
So why did he organize the entire company to attend this press conference?
Because a new model recently appeared on the leaderboard of Large Model -
KunYuan-1.0
This new model achieved good results on various test sets, and it was also specifically tested on a Chinese-specific test set. Industry insiders speculate that this is the new achievement of a research team in China.
Ryan also had his own guesses.
He searched through various Large Model available on the market, including the code-based Large Model previously released by Dingsheng.
QianYuan – this name clearly indicates a shared identity.
Considering Dingsheng's recent high-profile launch event, it's highly likely that KunYuan is Dingsheng's new model.
Over the past decade or so, Dingsheng has been one of the leading representatives of China's internet and cloud services industry.
In the era of Large Model, although they started a little late and only QianYuan model is well-known so far, their foundation is there, and no one will underestimate them.
Just like Meta, Amazon, Apple, they haven't produced anything in the AI era yet, but that doesn't stop people from believing that they will eventually achieve something.
Especially this time.
The Dingsheng hype was excessive.
Over the past week, the press conference has been discussed on the Chinese internet, in industry media, in developer communities, and among investors, and it even reached Ryan's ears.
The test data shows that KunYuan is a general-purpose Large Model, and it has approached the level of the top closed-source models on the market.
What worries Ryan the most isn't just KunYuan.
Instead, it's Source Code Technology's Riding the Wind.
Riding the Wind has demonstrated exceptional engineering capabilities and is at the forefront of the industry in the field of Agent. The only limitation now is that the China company that created Riding the Wind, Source Code Technology, does not yet have its own Large Model.
Without Large Model, there are no trump cards, and for the time being, there is no threat.
However, another China company happened to release a general-purpose Large Model at this time.
If this Large Model called KunYuan can truly achieve the level demonstrated in the tests, these two China companies could definitely collaborate.
This is what Ryan is most worried about.
That's why he wanted to see for himself what kind of opponent he was about to face.
As a traditional major player in China's internet industry, what kind of products can Dingsheng produce in the AI era?
The large screen at the front of the conference room has been connected to the live broadcast room, and the screen is currently showing the countdown to the press conference. Against the black background, an orange light flows slowly, like some kind of neural network being lit up.
There was only one line of text in the middle. On the original invitation, two words were blurred by Gaussian blur, and the degree of blurring changed as the countdown got smaller.
The text above reads "Let AI Understand China," with an English translation below. "Let AI Understand China."
"Ne. usAI's marketing manager Michael, muttered, "They're very good at marketing; this topic is very eye-catching."
CTO John glanced down at the computer screen resting on his lap. "This model named KunYuan not only excels in Chinese, but also has impressive overall abilities."
Michael glanced at John and asked, "How does it compare to Mercury?"
John replied, "Each has its own strengths, but Venus is about to appear; they can't compare to Venus."
Venus is a secret project that Neus is currently training on, and he is almost finished with the training. His goal for Venus is to become the world's number one.
Michael felt reassured and turned his attention back to the press conference.
The screen lit up at that moment.
A middle-aged man walked onto the stage.
Ryan whispered, "Zheng Xiaobo."
The person who took the stage was Zheng Xiaobo, the CEO of Prosperity Group.
He stood in the center of the frame, waiting for the applause to subside, and slowly began to speak, "Over the past two years, Large Model have changed the world."
The people in the conference room couldn't understand Chinese, but there were English subtitles at the bottom of the screen—generated in real time using Large Model.
But we also discovered a problem.
"Many models can answer questions in Chinese, but they may not truly understand the expressions used in the Chinese-speaking world."
They know the meaning of words, but not the weight of the scene.
"They know what users say, but they don't know what users really want."
Sarah, the product manager, looked at the subtitles and felt a deep resonance.
This is a question that product managers of all AI products are pondering.
AI is powerful, but it doesn't understand users well enough, especially when users don't provide enough context.
Zheng Xiaobo continued, "So today, Dingsheng brings you a brand new universal Large Model - Kun Yuan."
Two huge Chinese characters appeared on the screen.
Kun Yuan.
The audience erupted in enthusiastic applause.
No one spoke in Ne. usAI's conference room.
Following the usual practice of Large Model launch events, the first part should include model parameters, training scale, benchmark evaluation inference capabilities, coding capabilities, and mathematical capabilities.
In fact, these scores had already appeared on the list a week earlier, and the moment Kun Yuan's name was revealed, those who cared had already matched his score to the score.
However, Zheng Xiaobo did not demonstrate these.
He started with a business case: during the Dingsheng e-commerce event, Double Eleven' Day, many merchants had a lot of problems with event preparation, and the operations team simply couldn't handle it. However, the Kun Yuan model didn't require any adjustments or special customer service programs; it was simply provided to merchants directly and could offer excellent solutions.
He demonstrated this example on the spot, and used it as a comparison with another Large Model, which is currently the world's number one, OpenAI.
The GPT model.
Michael whispered to John beside him, "Double Eleven' Day, November 11th, is Black Friday for China people."
0
United States hold large-scale promotional events across various industries and categories on the Friday before Thanksgiving each year; this is also known as Black Friday.
John understood immediately.
These are things that are difficult for a model to know, but the Kun Yuan model does.
"This is a great example," Michael said, adding that it's very well integrated with real-world scenarios, especially their own.
The demonstration ended, and the audience erupted in enthusiastic applause once again.
"It's possible they've been manipulated at the rule level," John said. "They've written down the clues beforehand, so it's impossible to tell whether they truly understood the question or just had a pre-existing understanding of it."
"Let's just pretend they don't have a rule layer," Ryan responded with his assumption.
He did not hesitate to exaggerate his opponent's strength, but he worried that arrogance might lead him to underestimate his opponent.
The press conference is still ongoing.
The second demonstration is a government affairs scenario.
Zheng Xiaobo typed a conversational question on the spot, "Our district wants to implement a support policy for micro and small enterprises, but some previous subsidy programs haven't been very effective. This time, we want to avoid being exploited and also make the application process less complicated for businesses. Can you help me think of a design?"
Kun Yuan's answer was very organized, and he explained the reasons for each one. There were three actionable steps: clarify the beneficiaries and verify with official data; instead of cash subsidies, subsidies should be reduced or waived in a digital form; and reduce application materials, streamline the process, and automatically cross-verify with the government database.
These actions were all effective.
Michael remarked, "This is so China."
Michael himself is a China expert, having previously worked in the Asia-Pacific market and spent many years in China. Noticing John's puzzled look, he explained the differences China and United States administrative units.
But the more I explained, the more I realized that there seemed to be no difference; both were inefficient.
The fact that this scenario was specifically brought up reveals Zheng Xiaobo's ambition – he wants not only enterprise-level procurement but also government procurement, which has extremely high compliance requirements.
After several demonstrations, Zheng Xiaobo immediately announced the next major piece of news.
Kun Yuan will be the first to integrate with Dingsheng Cloud, Dingsheng Office Suite, e-commerce operation platform, government and enterprise service platform, customer service system and developer platform.
In short, as long as you use Dingsheng, you can immediately use Kun Yuan.
He wants AI capabilities to immediately feed back into existing businesses through traffic.
Finally, three rows of sliding partner logos appeared on the big screen.
Finance.
Manufacture.
E-commerce.
Educate.
Medical.
All industries are included.
Ryan took a deep breath upon seeing this.
"Their distribution channels are too strong," he said. "Once they have a model, they can immediately create a scenario."
Is this the dominance of Big Mac? Ryan thought of Google and Amazon, and felt the same sense of powerlessness.
NeusAI wanted to challenge such Big Mac, like a tiny knight charging towards a dragon.
As he drifted into these melancholy thoughts, the Dingsheng Kun Yuan press conference came to an end.
Ryan could foresee the huge sensation this press conference would cause in China.
China, this mighty lion, has finally entered the first tier in the competition for Large Model.
At the press conference, Lin Shaofeng stood in the shadows behind the stage curtain, watching Zheng Xiaobo receive applause on stage.
Kun Yuan's name has been revealed, and the partner's logo is still scrolling on the big screen. The atmosphere at the scene is so good that he himself is a little dazed.
This was something his team had created, and he was filled with excitement.
But for some reason, his gaze fell on Zheng Xiaobo's back on the stage, and Liu Dahai's words suddenly came to mind.
"I think there's something wrong with the data that came in the day before yesterday. I suggest we pause the training for now."
"I'm not talking about quality issues, I just feel that the logic and direction in it seem a bit too heavy."
What problems could there be with the data?
Lin Shaofeng suppressed the thought and focused his attention back on the stage.
The applause continued.