Lin Yizhou sat in his office and watched the entire Kun Yuan press conference.
Of course, during work hours, he didn't dare to go too far. He would turn the screen down and place it in a corner, keeping a close eye on what the manager was doing. If there was any disturbance, he would quickly use the code editor interface to block the video.
Wearing headphones is even less of a problem.
Which programmer doesn't wear headphones while working?
Lin Yizhou is not an employee of Dingsheng, nor has he bought Dingsheng stock. He is at most a user, which is only because the company uses Dingsheng Cloud.
However, Lin Yizhou is a fervent advocate for AI.
While everyone else was still hesitating about whether to learn it or give it a try, Lin Yizhou had already resolutely embraced AI.
Although he himself does not have an academic background in Artificial Intelligence and is just an ordinary backend development engineer, he actively follows news in the AI industry and spends time studying tools on his own.
When Riding the Wind was released, he registered immediately.
Now that Dingsheng Sheng is heavily promoting their first general-purpose Large Model, Lin Yizhou naturally wants to watch the live stream.
The press conference was truly stunning, but having seen so many press conferences, he still had doubts about the actual effectiveness.
Big companies always have to hype things up at their product launches.
It's unclear how much of what's been released was pre-programmed, or if it's all of it.
As soon as the press conference ended and the public beta link was released, Lin Yizhou registered.
Work-related content was definitely not allowed to be uploaded, so he tried it out using a few scenarios he frequently used on ChatGPT and Mercury, and was quickly impressed by Kun Yuan.
Kun Yuan really has something!
He soon couldn't resist sharing it with others online.
[@A Leaf on a Flat Boat: Brothers, you know what, Dingsheng really came through this time. The stuff they made is actually pretty good. Kun Yuan]
He also included several examples of his own practical use.
At first, there weren't many replies, but the buzz quickly grew. The last time Lin Yizhou saw 99+ replies was when he was complaining about his company not allowing him to use Riding the Wind.
This time, even more people replied, and Lin Yizhou couldn't keep up with reading them all.
More and more people are posting screenshots of their Q&A sessions using the Kun Yuan model in the comments.
The questions asked were varied and numerous:
Someone doing e-commerce operations threw in a messy campaign review, and Kun Yuan directly extracted the three issues that truly affected the conversion rate for him.
Some customer service quality inspectors pasted a bunch of customer complaint records into the system. Kun Yuan didn't summarize emotions like other models, but instead directly pointed out which point in the after-sales process was most likely to cause a second complaint.
Some people even included their own online novels in the comments. Unlike other models who would shower the author with compliments like "Lu Xun reborn, Ernest Hemingway second," Kun Yuan directly commented. "Your three golden chapters are lacking. Go back and think carefully about what the satisfying moments you're providing for the readers are."
The final conclusion is that the Kun Yuan model is too clever.
After getting home from work, Lin Yizhou turned on his computer and chatted with Kun Yuan for a few rounds, but soon reached the chat limit and was Rate Limiting.
The API isn't open yet, otherwise Lin Yizhou, who's currently having a great time chatting with it, might have to spend some money to use it and enjoy it before making any further comments.
Lying in bed, Lin Yizhou was still thinking about what the CEO of Dingsheng had said, "Let AI Understand China."
That's awesome.
With this thought in mind, Lin Yizhou fell asleep.
The next morning, Lin Yizhou groggily opened his eyes as his phone alarm rang. He picked up his phone from beside his pillow and started scrolling through his WeChat Moments. Soon, he saw an article that caught his eye.
The title is: <People smarter than you are even more cunning? Dingsheng Kun Yuan releases version 1.1 of its Large Model overnight>
Lin Yizhou rubbed his eyes.
Yesterday they held a press conference and launched public beta 1.0, and today they've already updated to 1.1? That's incredibly fast iteration! Is this how big companies operate?
I'll try it when I get to the company.
Thinking this, Lin Yizhou lay in bed scrolling through his phone for another half hour. When the second alarm went off, he dragged his tired body out of bed, ready to start his grueling day.
Lin Yizhou entered the company with the soy milk and fried dough sticks he bought on the way. When he arrived at his desk, he found a task assigned by the manager already on the table.
The ongoing project has to be postponed due to insufficient resources. A project postponement explanation needs to be written and shared with upstream and downstream teams.
This was originally the manager's job, but the manager refused to do it and instead passed it on to Lin Yizhou, so he had no choice but to bite the bullet and do it.
Thankfully, there's AI.
Lin Yizhou opened Kun Yuan's webpage and, sure enough, saw that yesterday's model number 1.0 had changed to 1.1 today.
Version 1.1 is naturally better than 1.0, and he didn't care, so he directly entered his requirements.
The company doesn't allow uploading code to AI tools, so I asked him to write a report for me, is that alright?
Soon, Kun Yuan output. "The explanation for the project delay should be elaborated from the following aspects:"
I. Project Background (Checkmark)
Introduce the project's background, development goals, and current progress.
II. Reasons for the postponement (checkmark)
This can be explained from aspects such as changes in demand, resource coordination, technical difficulties, and testing and verification.
III. Follow-up Plans (Checkmark)
Clearly define the work arrangements, responsible persons, and timelines for the next stage.
IV. Safeguard Measures (Checkmark)
Strengthen project management, optimize communication mechanisms, improve collaboration efficiency, and ensure the smooth progress of the project.
Lin Yizhou was completely baffled by this answer.
Oh no, did I use the wrong model? This answer is even worse than ChatGPT's! And what's with all this emoji stuff on your screen? Where's my understanding, resourceful Kun Yuan? Where's my big Kun Yuan from yesterday?
Lin Yizhou tried several different prompts, even asking the same question he had asked yesterday, but the output was completely different from yesterday's.
Wait, yesterday's usage records are gone too.
Finally, Lin Yizhou had to come to a conclusion.
Kun Yuan 1.1 has a significantly lower intelligence level compared to Kun Yuan 1.0.
The difference is too obvious.
Lin Yizhou took out his phone, opened a small, unofficial AI community chat group, and found that people were already discussing this matter.
Did you guys try again today? Why do you seem dumber?
[It's not just a feeling, I've really become dumber.]
Has Kun Yuan changed its model?
[The page is now showing version 1.1.]
[No way, version 1.0 only lasted one night?]
[Last night I asked it to analyze a growth plan for me, and it immediately pointed out a flaw in the design. Today, with the same question, it gave me a generic, all-purpose reply.]
The person who spoke last posted two screenshots.
Yesterday, after reviewing a group-buying growth plan, the first thing Kun Yuan 1.0 said was:
The biggest problem with this plan isn't the lack of growth strategies, but rather the assumption that fulfillment capabilities can withstand growth—this premise is flawed. The faster the growth, the greater the pressure to fulfill obligations, the worse the user experience, and ultimately, the faster churn.
Asking the same question again today, Kun Yuan answered:
[Community group buying growth can be achieved through multiple aspects such as user acquisition, event operation, supply chain optimization, and fulfillment experience. Hope this helps! (smiley face)]
People in the group started frantically scrolling through the history.
Those who used Kun Yuan yesterday have been posting screenshots of their saved conversations.
After several hundred more posts in the group chat, a comment suddenly caught Lin Yizhou's attention.
[Has Dingsheng taken down the special edition of the press conference?]
Strangely, many people are discussing this issue on WeChat, but it's not on Weibo's trending topics list; even yesterday's trending topics [Kun Yuan][# Let AI Understand China] are gone.
Lin Yizhou even felt that he had briefly hallucinated yesterday, fantasizing that a useful AI had been developed domestically.
Kun Yuan 1.0 thus became another urban legend circulating on the Internet.
In Lü Yun's office at Dingsheng Tower, Lü Yun sat behind his desk, while Liu Dahai sat opposite him in the chair, feeling somewhat uneasy.
Last night, he was urgently called back to the lab from home. He selected a "clean" model from the past two or three months that performed best, named it Kun Yuan 1.1, worked overtime to go through the basic tests, and then launched it online to replace Kun Yuan 1.0.
The term "clean" means that the high-quality annotated versions have not yet been used.
Although no one told him explicitly, Liu Dahai roughly guessed that there was indeed a problem with those labels, and now that it had been exposed, such urgent action was necessary.
He'd seen this kind of thing before. When several Large Model Distillation each other's products, sometimes the cleaning wasn't done properly, resulting in situations where model A was asked "Who are you?" and the answer was "I'm model B." These kinds of incidents always caused a stir in the industry, and everyone was used to it.
--
However, he carefully reviewed the data during the cleaning process and shouldn't have made such a basic mistake.
"Dahai," Lü Yun spoke, interrupting Liu Dahai's wandering thoughts, "You've been at Dingsheng for two years now?"
"It's been a year and a half, Boss Lü," Liu Dahai quickly replied, snapping out of his daze.
Lü Yun nodded slightly, a smile appearing on his face, "That's right. From the moment you came in, Qianyuan has been entrusted to you, Kun Yuan is also under your control."
Liu Dahai's forehead began to sweat. Indeed, he was the one who set the technical direction for the development of Dingsheng's Large Model. Now that Kun Yuan had probably run into trouble and emergency measures had just been taken, was Boss Lü going to settle accounts with him?
He tried to defend himself, saying that he had raised different opinions before the release, but Lin Shaofeng had silenced him.
But he was afraid that if he said this, he would be seen as making excuses.
Sigh, Liu Dahai was even nervous when he met Zheng Xiaobo, so meeting the legendary Boss Lü today put him under a lot of pressure.
Lü Yun didn't let Liu Dahai overthink for long and quickly got to the point. "Dahai, the reason I called you here today is to give you some more responsibility."
Liu Dahai was taken aback. What did that mean?
"I've thought about it myself. For Large Model research to succeed, it still needs to be led by the technology side. We need to change the structure of the Research Institute. Your job title will be raised one level, but you'll still be called Chief Scientist. However, you'll be in charge of the Research Institute."
Liu Dahai paused for a moment, before the joy could even fully surface, and his first question was, "What about General Manager Lin?"
The Chief Scientist oversees the Research Institute; what was the VP who originally managed the Research Institute doing?
Lü Yun hadn't expected Liu Dahai to ask so directly, but he still replied, "He'll be transferred to Dingsheng Cloud, and the two of you will need to work together from now on."
Liu Dahai pondered for a moment: wasn't Dingsheng Cloud originally managed by Zhang Chi? So where is Zhang Chi now?
But that was too far removed from his life, so he didn't ask any more questions.
Regardless, a promotion is a good thing.
Liu Dahai thanked the big boss for his appreciation, and Lü Yun then asked a few technical questions.
How different is this current version from the one shown at the original launch event?
This was the issue that Lü Yun was most concerned about.
"The gap in intent understanding, especially in a Chinese context, is quite large, dropping from over 90 to less than 80."
In this respect, Liu Dahai is actually the one who feels most wronged.
All the models in the Kun Yuan series were developed step by step with a lot of effort by him. The version presented at the press conference used those "golden annotations" for post-training, which greatly improved the semantic understanding ability of Chinese.
However, Kun Yuan's own knowledge, logic, and mathematical abilities are also among the top tier of closed-source models.
The semantic understanding ability and the ability to connect with real-world situations, which were emphasized at the press conference, have all disappeared after the regression. Now, all the public opinion he sees is saying that Kun Yuan has become stupid.
This is still public opinion that has been managed through Public Relations efforts. If netizens were really allowed to discuss it freely online, who knows how much criticism Kun Yuan would receive.
How damn wronged I am!
After listening to Liu Dahai's answer, Lü Yun slowly nodded and asked another question, "Do you have a direction for achieving the same effect as at the press conference?"
Liu Dahai was now confident. "Yes, Boss Lü, yes. We've already verified that the standard for annotation improves intent understanding. Next, as long as we increase investment in this area and build a high-caliber annotation team—not just ordinary annotators, but top students in the humanities, like those studying philosophy, psychology, or linguistics—I'm confident that in six months, we can achieve the same level of quality as at the press conference."
This brought some comfort to Lü Yun.
It's good to be one step ahead, but if you can't be fast, then at least do it well.
A few days later, in NeusAI's office on the West Coast of the United States.
CE0 Ryan reviewed the test report of Kun Yuan 1.1 again.
This is the third report.
The first document came from the product team, the second from the research team, and the third from a data compilation from an external community.
The three reports reached essentially the same conclusion.
Kun Yuan 1.1 has decent overall capabilities, but it falls far short of the imposing presence it projected at the launch event.
Especially in complex Chinese business scenarios, it still has advantages, but these advantages are not insurmountable.
John CT0 sat on the other side of the conference table, his expression visibly relaxed.
"So, what exactly did they show at the press conference?" Ryan asked.
John shrugged. "It could be a special version, it could be overfitting to the testing and demo scenarios, or they might have implemented a safety measure after release. In short, the version we can test right now isn't worth changing Venus's release schedule."
He said at the time that it was probably tuned at the rule level, not a native capability of the model.
This can now be confirmed.
Marketing manager Michael also breathed a sigh of relief.
He was genuinely a little nervous after watching the Dingsheng press conference a few days ago.
If a major China company truly develops a model that thoroughly understands Chinese business scenarios, coupled with its Dingsheng channels and customer resources, it will put significant pressure on NeusAI's overseas expansion and its assessment of the China market. Add to that the rapidly growing Riding the Wind and Kaiwu Overseas Edition, and NeusAI's situation will become very precarious.
But now it seems that things are not as terrible as they seem.
The public opinion on the China internet has already completed the first round of questioning for them.
Once the label of "score fraud" is attached, it's very difficult to remove it.
Even if Dingsheng does create an even better model in the future, everyone will still ask: Is this for real?
Ryan also let out a soft sigh of relief.
He's not afraid of having one more competitor, but what's wrong with having one less?
With the peak of his Dingsheng over, Ryan looked up and asked the team another question.
"How's the preparation going with Source Code Technology? We've finished collecting all the data. After March, when the contract expires, we'll revoke their access."
The day after Kun Yuan's press conference, and after signing a settlement agreement with Dingsheng, Han Luyi stopped paying attention to Dingsheng.
He has more important things to do.
It's already mid-March. Glutinous Rice Ball's training is complete, the scores have been measured, and the application scenarios have all been tested. The results are even better than expected.
However, the deployment in Source Intelligence's own Data Center is still in full swing.
Therefore, Han Luyi did not rush to make a public announcement.
With the current usage of Source Code Technology, the combined Daily Active Users of Open Things and Riding the Wind have already exceeded 100,000. After Glutinous Rice Ball is released, the AI robot on the webpage will inevitably bring another wave of traffic.
It's fair to say that the company will face traffic pressure as soon as it's released.
Han Luyi didn't want to deploy Glutinous Rice Ball on any cloud platform, so he could only make use of the existing 500 cards and the 2,000 cards that came from Dingsheng as soon as possible.
The technical difficulty was greater than previously anticipated.
Although it only uses domestically produced cards for inference and does not involve the training part, it is still the first company to do this, and it is all trial and error.
In addition, the volume in Zhangjiakou has increased significantly, so a dedicated operations and maintenance team needs to be set up.
Han Luyi estimated that, putting aside whether or not to go to Beijing, he should first open a branch office in Zhangjiakou.
Since the end of the year, Zhang Biao has been stationed in Zhangjiakou, busy with the Data Center. Han Luyi thought about it and felt that he still had to go there in person and bring Zhao Wenyuan with him.
After arranging the itinerary, Liu Huo booked plane tickets for the three people for March 15th.