Skip to main content
  • Home
  • Library
  • Read
    • Popular
    • Charts
    • Browse
    • Categories
    • Wikis
    • Glossary
  • Request
  • Activity
  • Leaderboard
  • Help
Our Apps
  • DaoSearch
  • Narratorr
Social
  • Discord
  • Patreon
The AI Wave Tycoon's Rise
Chapter 426

Now this was an API for the new era

3,169 words 11 min read 1 view 1mo ago Source ​
Skip to the end of the chapter
Prev
Next

Han Luyi answered the phone.

"Boss Han, congratulations!" Zheng Xiaobo's hearty laughter rang out.

"Boss Zheng, what brings you this good news? Why would you take time out of your busy Boss Zheng to call me?" Han Luyi asked.

"Boss Han, the opening of the Glutinous Rice Ball API Launch is such a big deal, the whole industry is discussing it, so naturally I'm here for this," Zheng Xiaobo said. "I heard that Origin was previously troubled by Computing Power issues, but now that the API is open, this bottleneck must have been resolved, right?"

"Thank you for your concern, Boss Zheng. There has indeed been some progress," Han Luyi replied.

Zheng Xiaobo's words are all nonsense.

The lack of Computing Power at the Origin is no secret in the entire industry.

Moreover, due to the restrictions of the Entity List, Origin was impossible for Yuanchi to buy Nvidia graphics cards, and Han Luyi did not intend to buy them either. Therefore, he took a different approach and invested all his resources in the Domestic Computing Power.

No one could guarantee how long it would take to produce results in this direction until Origin suddenly emerged and, under the guise of official authority, organized Domestic Computing Power Alliance. Only then did everyone realize:

They actually made it for them.

The current Computing Power alliance includes both wholly-owned subsidiaries of Dingsheng and companies in which Dingsheng has invested, Origin from the promotion of domestically produced computing power.

The Power Alliance was established to solve the problem of insufficient Computing Power, and everyone knows this, so Zheng Xiaobo has no reason not to know.

They could even piece together the exact number of cards and Computing Power that Origin possessed from industry insiders.

So his call definitely wasn't because he was worried about Computing Power.

"Does Boss Zheng have any other suggestions?" Han Luyi asked.

Zheng Xiaobo Harbin University a few more times. "Boss Han truly perceptive. I certainly don't come here without a reason; I have a favor to ask."

"Please go ahead, Boss Zheng."

"With such a enthusiastic market response to Glutinous Rice Ball rice balls, even if Boss Han has extra Computing Power, he must be nearing his limit, right?" Zheng Xiaobo's voice became serious. "I wonder if Boss Han would consider deploying the model on Dingsheng Cloud and having us sell it? You have the model, I have the Computing Power; you have the market, I have the customer base. It's a win-win situation."

Han Luyi cold laugh.

"Boss Zheng's plan is quite good, but I'm not entirely confident about cooperating with your company. If it weren't for Boss Lü's personal intervention last time, things probably wouldn't have ended so simply, would they?"

Dingsheng really seems to have forgotten its past mistakes. Last time, they stole the Annotated Data Tangyuan 0.9's training run on Dingsheng Cloud and used it to train Kun Yuan. And now, less than two months later, they want to repeat the same trick?

Who can guarantee you won't steal it again?

By deploying the model on someone else's Computing Card, the contents of the GPU memory are readily available, and the model's parameters and weights are essentially fully accessible.

Why does Neus AI dare to put its models on Google Cloud? Why does OpenAI deploy its models on Microsoft Azure Cloud? It's because they are not only partners, but also have a direct investment relationship.

Otherwise, they're also afraid their parameters will be stolen.

To address this issue, Nvidia has introduced a technology called Memory Encryption Computing to ensure that the contents of the video memory cannot be read by outsiders.

Ultimately, if you run the model on someone else's Hardware, no amount of encryption will help. Even if you can guarantee the parameters, Distillation is still inevitable.

Given the relationship between Dingsheng and Origin, this is simply impossible.

Zheng Xiaobo explained, "Boss Han, the previous misunderstanding between our two companies was purely due to individual actions within the company, and the individuals involved have all been dismissed. Please believe me, Dingsheng Cloud prioritizes customer privacy above all else. The cloud business is worth hundreds of billions; is the AI market that big? Dingsheng would not jeopardize its reputation for a new opportunity."

Han Luyi didn't believe it at all.

The problem isn't whether it's permissible from a management perspective, but whether it's physically feasible.

Today you say it's an individual act, but what if you steal again and it's still an individual act? Why not just fire two people and call it a day?

"Boss Zheng, there's no need to discuss this anymore. The Tangyuan Model will not be deployed on Dingsheng Cloud."

Zheng Xiaobo then advised, "Boss Han, Boss Lü previously mentioned to me that if Boss Han interested, Dingsheng could issue additional shares to Origin through a private placement, and we could swap shares. What do you think, Boss Han?"

Han Luyi was startled.

They probably heard about the terms they negotiated with Chengjiang Technology, and that's why they came up with this idea.

Regardless of the valuation used for the share swap, Dingsheng definitely wants to join the board of directors. Preparations for the listing of Origin have already begun, Han Luyi doesn't want to let them in.

"Please thank Boss Lü kindness, but we won't be considering this option for the time being," Han Luyi said again, declining the offer.

"Boss Han, it seems you still have some reservations about Dingsheng. I hope it won't affect our cooperation. I noticed that the Tangyuan 0.9..."

The model interface is not open, and you can't specify a model in Tangyuan Chat and Riding the Wind. I wonder if Origin has completely given up? Could you please put Tangyuan 0.9 on Dingsheng Cloud?

Zheng Xiaobo was rejected twice, but seemed completely unconcerned, and then proposed a third option.

Han Luyi immediately realized that this was his real purpose.

Han Luyi was happy to accept the first two suggestions, but he didn't care if they were rejected, because he just wanted a connection.

This marks the second consecutive rejection.

It wasn't until he proposed a more reasonable option for the third time that Han Luyi, felt somewhat embarrassed.

Zheng Xiaobo is right. After the official release of Glutinous Rice Ball 1.0, Tangyuan 0.9 was basically completely removed from the app store.

The reason is that versions 0.9 and 1.0 actually use the same pre-trained base, only the Annotated Data used in the later training is different. The two are about the same size and consume about the same amount of Computing Power. Furthermore, version 1.0's results are superior to 0.9 in every aspect, so there is no reason to deploy version 0.9 again.

Han Luyi secretly considered that Zheng Xiaobo's desire to take over 0.9 for sale was indeed negotiable.

The Kun Yuan 2.0, developed under Liu Dahai's leadership, uses the same approach as Glutinous Rice Ball, with a cleaner dataset and more complex Annotated Data. Its capabilities have been significantly improved, and the gap with Tangyuan 0.9 is not too large. Whether they directly steal parameters or secretly Distillation, it will not have any impact on Glutinous Rice Ball 1.0.

This part utilizes Dingsheng's Computing Power, which is licensed from Origin to earn licensing fees or share profits proportionally. At the same time, it can also expand Glutinous Rice Ball's user base. Overall, the benefits definitely outweigh the drawbacks.

This third suggestion is indeed worth Han Luyi's careful consideration.

"Boss Zheng, you're amazing," Han Luyi exclaimed, without specifying what made it so impressive. "Tangyuan 0.9 is the previous generation model. Boss Zheng is also interested in it?"

Zheng Xiaobo laughed and said, "In front of Boss Han, I can't claim to be any great. Dingsheng is a comprehensive company; we don't just do AI R&D. We have to consider all our business units comprehensively. Many of Dingsheng Cloud's enterprise clients hope we can launch the Tangyuan Model, and we're doing this for their benefit."

Han Luyi felt that these words sounded somewhat familiar.

He suddenly remembered that he had said the same thing when he went to Tianshu Cloud to sell BugKiller back then.

It's not that we insisted on coming; it's just that our clients specifically requested that we list our products, so we sought cooperation. It's all for the clients' benefit.

Han Luyi amused and exasperated.

Sure enough, when it comes to doing business, all the excuses in the world are pretty much the same.

"What are Boss Zheng's thoughts on the specific terms of cooperation?" Han Luyi asked.

"One price, 1 billion, Dingsheng will buy the license for Tangyuan 0.9. How about it?"

Han Luyi.

"Boss Zheng, there's no reason for a fixed price for this kind of cooperation."

"So, how does Boss Han plan to cooperate?"

"Since Boss Zheng is so sincere, then please send someone over to discuss this further."

"Boss Han a straightforward person."

As soon as Zheng Xiaobo hung up the phone, the smile on his face disappeared.

Han Luyi has grown up too fast; Zheng Xiaobo can no longer treat him the same way he used to.

Zheng Xiaobo dared not underestimate him anymore, not because Lü Yun valued him, but because Origin Group had become an undeniable force in the Artificial Intelligence industry.

He thought for a moment, then called his secretary in, "Have Wang Hao come over. Tell him about the next phase of cooperation with Origin, and ask him to make preparations."

"Okay, Boss Zheng."

Let's rewind two hours.

Pengcheng.

Around nine o'clock, Lin Yizhou opened his eyes and stared at the bedroom ceiling.

He has moved out of his rented apartment in Longgang and rented a new Two-Bedroom Apartment in Nanshan District.

One room is used as a bedroom, and the other as a studio.

For nearly a month, Lin Yizhou lived as if in a dream.

Since the Tangyuan Protocol has been integrated with more and more Large Models and AI assistants, more and more people are buying Yuanxiao Mailbox.

In less than a month, he had already earned 100,000 yuan.

100,000 yuan!

When he was doing backend development, even the Second-Tier Mid-Sized Tech Company he switched to could only offer him an annual salary of 400,000 yuan before tax, which meant he only received a little over 20,000 yuan per month.

My current income is four times what it was back then.

Moreover, this is a monthly subscription, not a one-time purchase, which means it costs 100,000 per month. If the number of paying users continues to increase, it will rise further.

Is this what it feels like to be a pig riding the wave of opportunity?

This is so awesome!

Recently, more and more users have been purchasing Yuanxiao Mailbox accounts, and many of these users are from overseas.

Clearly, compared to the situation in China where many services are integrated into WeChat, overseas users tend to be heavy users of email.

Many companies have even inquired about the collaboration features and want to purchase them.

Lin Yizhou, who had experience in the SaaS business, immediately realized that this was a truly stable source of income. Unlike individual users, the corporate procurement process is complex, and once approved, it is not easy to interrupt.

However, one of their main needs is that they don't just want Yuanxiao Mailbox to be a tool used by the AI assistant, but rather they want Yuanxiao Mailbox to be an entry point, while also having AI functions such as natural language filtering of emails and email manipulation.

Lin Yizhou naturally wanted to continue seizing this opportunity.

However, if you want to add AI functionality to this little software, you have to call the API in the background.

This is truly developing an AI Agent.

To create a tool that an AI assistant can use, you only need to adapt to the Tangyuan Protocol, meaning you just need to make your functions usable by AI; you don't actually need any AI capabilities at all.

However, developing Agents requires some understanding of AI.

Although many of the users served were from overseas, Lin Yizhou, the developer, was based in China and could only use the APIs provided by domestic model manufacturers. Lin Yizhou had tried several different options, but the results were not satisfactory.

Take the most basic email filtering function as an example; the function that Lin Yizhou needs is actually very simple.

If a user says in Chinese or English, "Please find my emails from last week."

Lin Yizhou needs to send this input, along with the System Prompt he wrote, to the AI model's API.

As a traditional backend developer, he naturally had little experience with System Prompts and could only explore the effects on his own while searching for templates shared by others online.

The first version was written like this:

[You are an email assistant. The user describes their needs in a language. Transform these needs into one of the following commands and return a JSON document in a fixed format.]

[The JSON document format is:] is followed by a multi-level nested JSON format. Then, a parser is written in the Yuanxiao Mailbox client to convert this JSON into specific commands.

[User Input:] Then fill in the user's input message.

In theory, at this point, the API should translate the customer's request into a command that the email software can understand, and then the software can run it directly.

However, Lin Yizhou was surprised to find that such a simple architecture had so many pitfalls.

Because a client might say, "Please find the emails from last week that I need to reply to."

How do I find it?

First, you need to define the time range, then send the email content back to the API so it can determine whether a reply is needed.

However, the API will only return one type of command now, which is the one specified in the System Prompt.

Lin Yizhou was at a loss. The System Prompts became increasingly complex, and the processors receiving the replies also became more and more complex, but the functionality didn't improve much.

Last night before going to bed, I tried again but failed. Kun Yuan's API returned content that the parser could not understand, and the system reported an error.

Damn, I only realized how difficult it is to develop an Agent writing one myself.

This feature has been stuck with Lin Yizhou for over a week with little progress. As expected, he'll be spending another day in the studio battling this new function.

Lin Yizhou got out of bed and brushed his teeth while scrolling through his phone in the bathroom. He came across an AI he usually followed.

Blogger.

Although this blogger occasionally posts some irrelevant content, the style is really exaggerated, and they like to sell anxiety. Lin Yizhou reads it absentmindedly.

"Breaking news! The API for Glutinous Rice Ball 1.0 is officially open today! If you haven't heard of Glutinous Rice Ball 1.0, you're really out of touch. This is the model behind Tangyuan Chat that we use every day. To all my followers, I recommend you register an account on the Origin Developer Platform now; it will truly change your perspective--"

The blogger was still rambling on with his content, but Lin Yizhou realized what was happening.

What did he say?

Glutinous Rice Ball has launched an API?

Why haven't we heard anything at all?

Lin Yizhou rinsed his mouth, spat out the foam, and then used his phone to open the Origin Group Official Website to view the content about the Tangyuan Model.

It's actually open!

Limited availability, first come, first served!

Lin Yizhou didn't even bother to eat breakfast yet. He went straight to his computer after leaving the bathroom to buy the API quota first.

Lin Yizhou was a veteran developer who had gone through this process countless times. He quickly bound the payment method, generated the API Key, and deposited 1,000 yuan into it.

The reason he invested so much money was naturally due to his long-standing trust in the Glutinous Rice Ball Series Models.

Even if your API development doesn't go smoothly, you'll still need these credits in the future, so they won't be wasted.

On other platforms, he wouldn't dare to recharge so much; the money won't be refunded, and if you can't use it all, it's your own fault.

Then Lin Yizhou carefully examined the official tutorial on how to use the Glutinous Rice Ball 1.0 API.

It provides a detailed explanation of how to develop Agents using the Glutinous Rice Ball API, with particular emphasis on how to handle Long-Range Tasks.

Lin Yizhou's heart pounded wildly as he read this.

If a user needs to use a function related to the email content in my Yuanxiao Mailbox, wouldn't that be Long-Range Tasks?

Sure enough, the Glutinous Rice Ball API has something new for Long-Range Tasks related to Agents: Tool Calling Chains.

In other words, it's not just about telling Glutinous Rice Ball what the Yuanxiao Mailbox needs to receive, but about telling Glutinous Rice Ball everything the Yuanxiao Mailbox can do, and then Glutinous Rice Ball decides which functions to call and whether to continue communicating with Glutinous Rice Ball API.

All the complex logic that needs to be implemented on the client side doesn't need to be written anymore; it can be directly handed over to Glutinous Rice Ball to handle.

Lin Yizhou watched with growing suspicion.

It's not that I don't believe it, but isn't this written a bit too beautifully?

However, since the experiment wasn't complicated, Lin Yizhou commented out the dozens of lines of System Prompts he had written before, and then, following the official tutorial, wrote the functions provided by the Yuanxiao Mailbox into a tool format that the Glutinous Rice Ball API could accept.

It's actually quite easy to write; the format is basically the same as the Tangyuan Protocol.

Finally, I wrote only one line of parser and one loop.

Ready to test? Lin Yizhou held his breath. Wait, no, wait a minute, let's Add One limit to the number of loop iterations first. Otherwise, what if the model keeps running the loop and runs out of money?

Lin Yizhou was glad that he had quick wit and added the missing logic. The loop would execute a maximum of ten times, which meant that it would make a maximum of ten requests to the Glutinous Rice Ball API.

Then he held his breath again and typed the command that had recently frustrated him so much on the newly compiled local development version, "Find the emails from last week that I need to reply to."

Instead of waiting to see the results, he immediately opened developer mode and started looking at the local logs.

[Sent to Glutinous Rice Ball API: User: "Help me find the emails from last week that I need to reply to", Tool: --]

[Reply: Using the tool: Filter emails by date and retrieve email content]

[Content of 246 emails sent to Glutinous Rice Ball API]

[Reply: Using the tool: rendering interface, email ID*17]

After being called twice, the call chain ends.

Lin Yizhou switched back to the Yuanxiao Mailbox interface, where he saw that of the more than two hundred test emails he had deliberately generated, the seventeen that he had manually marked as needing a reply had been filtered out.

At this moment, Lin Yizhou couldn't describe how he felt.

That's incredibly awesome.

The problem that had been bothering him for over a week was solved by this simple cycle.

What APIs did I test before?

This is the API of the new era!

Lin Yizhou quickly submitted the change, then entered a new command in Yuanxiao Mailbox: "Find all emails that have inquired about the Yuanxiao Mailbox AI entry point and reply to them all: The new feature has been developed and is expected to be updated next week."

End of Chapter
Prev
NovelSummaries
Next

Explore the wiki

Han Luyi

Comments

No comments yet.
AboutStatsAPIChangelogStatusContactPrivacyTermsDMCA
© 2026 RandomTranslator·v3.26·RSS