Wednesday afternoon, Origin Group headquarters, Vice President of Product's office.
Jiang Yixin, who had just returned from lunch, sent WeChat to Liu Haoyu.
[Xiao Jiang (Wanting Hot Pot Version)]: Guess what we're having for dinner tonight?
[Big Liu (Lord)]: Sorry, my soldier, I can't take you out for hot pot today. I have to go home and play Mount & Blade after work.
[Xiao Jiang (Wanting Hot Pot Version)]:???
Jiang Yixin angrily locked her phone screen, then imagined the awkwardness of going to Haidilao alone and having a teddy bear placed across from her.
Forget it, I'll just go home and order takeout.
After finishing her slacking off, President Jiang became engrossed in her work, needing to review the call data shared by Dingsheng.
This was also the result of the negotiations. Although Dingsheng and Origin were cooperating in more and more areas, the foundation was not very solid. During the business negotiations, Jiang Yixin insisted that the Tangyuan 0.9 deployed by Dingsheng must send real-time call volume data to Origin's Server synchronously, in order to prevent Dingsheng from falsifying the call volume.
Therefore, although the call fee is settled monthly, the call volume can be viewed in real time.
From Monday morning when Tangyuan 0.9 was launched on Dingsheng Cloud's model store, to now, a little over two days later, it has been deployed on...
The total number of API calls to Dingsheng Tangyuan 0.9 has exceeded that of the official API of Origin.
Jiang Yixin was utterly astonished when she saw the data. No wonder Dingsheng was in such a hurry to launch the Tangyuan Model; how did they have so many users willing to pay a price even higher than a fully-fledged flagship product to subscribe to Tangyuan 0.9 through Dingsheng's channels?
Of course, you can ask President Su about this, but Jiang Yixin has been deliberately working on developing her ability to think independently lately.
She opened a new document on her computer and first threw her thought process into it.
First of all, is Glutinous Rice Ball 1.0 popular? Yes! It's so popular that it's hard to find Token. Glutinous Rice Ball is the most discussed topic in the entire AI community.
However, since Origin has limited Computing Power and is unwilling to deploy Glutinous Rice Ball 1.0 on other platforms, those who want to buy it cannot, but the development speed using Riding the Wind has indeed been greatly accelerated.
AI narratives are no longer optional; any tech company that hasn't announced its embrace of AI is simply seen as outdated.
Companies eager to embrace AI have quickly developed prototypes and integrated Large Model into their services through dialogue. However, only a few can develop their own models; most can only call other companies' APIs. Tangyuan 0.9 is the most cost-effective option available.
In addition, Dingsheng's cloud service ecosystem is well-developed, with all permission management functions built-in. For companies that have already chosen Dingsheng, the integration process is not complicated. Furthermore, Dingsheng has accumulated years of experience in the cloud market and has much more Computing Power than Origin, which can basically meet the usage needs of companies that want to purchase it.
This explains the strange phenomenon that the 0.9 version sold by Dingsheng actually outsold the 1.0 version from Origin.
Jiang Yixin sat there and frowned. She even thought of an advantage that Dingsheng had but Origin did not: the company's potential growth in Computing Power demand.
Everyone knows that Origin is facing a shortage of Computing Power; getting access to it is basically a matter of luck. If a company finds its usage to be exceptionally high, it will grow rapidly, but what happens when Origin can't supply it?
On the contrary, Dingsheng has a stable supply. Although the price is higher and the version is older, for many companies, a stable supply is more valuable, and they are willing to pay a premium for it.
From Dingsheng's perspective, although selling Glutinous Rice Ball) is profitable, this money still has to be shared with Yuanqi (the source of the Origin), so it's definitely not as worthwhile as selling Kun Yuan. Therefore, they deliberately set the price of Glutinous Rice Ball higher and the price of Kun Yuan lower, in order to encourage people to use Kun Yuan.
However, Glutinous Rice Ball's reputation is still too authoritative in the entire AI circle. Since people can buy Glutinous Rice Ball, they are still willing to buy Glutinous Rice Ball.
Simply put, Origin has models but lacks Computing Power, while Dingsheng has Computing Power but its models are not as attractive as Glutinous Rice Ball. All of this leads to the fact that for many enterprises, deploying Tangyuan 0.9 on Dingsheng Cloud is the best choice.
Jiang Yixin looked at her densely written document and felt that her thoughts were much clearer.
Therefore, the key to the next stage of development for Origin is to continue expanding Computing Power. Jiang Songran is already in charge of the Hardware side, but what can she do from a product perspective?
Jiang Yixin thought about it carefully. Currently, Riding the Wind, Tangyuan Chat, and Glutinous Rice Ball API all use the same Computing Power pool. If these two first-party applications can save Computing Power, wouldn't that mean that the API side can have more quota to sell?
Of course, we shouldn't reduce the intelligence of Origin's two flagship products just to save Computing Power. If we did that, it would lead to a sharp drop in reputation and a loss of user trust, which would be putting the cart before the horse.
Jiang Yixin, the vice president of product, understands the principle of prioritizing user experience.
However, after working as an AI product manager for so long, Jiang Yixin has accumulated a lot of methodologies. One of the most important ones is that the output quality of a model is not always positively correlated with the Computing Power it consumes.
Using the same model and the same problem to solve in two different ways may indeed result in one method consuming less Computing Power and yielding better results; it's possible to have the best of both worlds.
If we continue to optimize the system constraints of these two products, will we be able to improve in this direction?
"This is true cost reduction and efficiency improvement, isn't it?" Jiang Yixin thought happily.
Then she quickly organized her thoughts into a document, preparing them as project goals for the next phase.
"So our next goal is to continue optimizing the Agent." As she was writing, Jiang Yixin suddenly got stuck on this sentence.
Then she thought of a question: this kind of ability, which is independent of the model's capabilities, is crucial to the performance of the Agent.
It seems that some of them don't have a name yet.
What should we call it?
This is actually a very important issue, because without a concise name, people cannot discuss a specific matter. For example, when Agent were first invented, the concept of Agent did not yet exist; if people wanted to describe them, they could only say:
This is an Artificial Intelligence automation tool that can autonomously perceive its environment, perform reasoning and planning, invoke external tools, and generate final results.
Naming tools that meet these criteria as Agent greatly reduces the communication costs required for people to discuss them.
The concept Jiang Yixin just came up with is actually an extension of the large-scale application of Agent. You are Agent, and I am Agent, so why is it that the Agent developed by Origin Group are better than those developed by other companies, both in terms of Computing Power utilization efficiency and task execution results?
This is because the constraints outside the model's capabilities are handled better.
Jiang Yixin began to recall the history of AI tool development over the past few years. Shouldn't this comprehensive constraint on model capabilities be considered a kind of engineering capability?
So what should we name it?
Prompt Engineering? No, this is much more complex than simple cue word management. It includes not only the cue words given to the model, but also the data format that the model should return, how the tool itself processes the returned data, and even when to actively compress the context.
Similarly, even the term "context engineering" cannot fully represent all the work.
Not very confident about her name Jiang Yixin and thought for a while, and suddenly remembered the game called "Mount & Blade" that Big Liu had been obsessed with recently.
If we compare Large Model with reasoning and thinking capabilities to a thoroughbred horse capable of traveling a thousand miles a day, and the person using the model to the rider of that horse, then could this ability to provide direction and constraints for the model be called "Harness Engineering"?
Harness Engineering.
Its English name is Harness Engineering.
It's settled then!