It's hard to say what kind of cutting-edge technology Sight is.
However, computers were invented and manufactured by humans, and all their operations are traceable.
Han Luyi stopped using the Sight control and returned to his usual way of operating the browser, gripping the mouse and searching for information.
Then I opened the terminal and entered the following command.
"$ cat /proc/bus/input/devices"
A long string of output was displayed, including all connected devices. Han Luyi one and found nothing unusual. Then he started a listening process, opened a new document using the Sight controller, and typed "HelloWorld" into it.
The output of the monitoring process shows that for each operation, an unknown device connects to the computer for a very short time, triggers a system input, and then disconnects.
It's as if a device sends a signal at the kernel driver level, bypassing the verification that requires persistent device state.
Han Luyi wasn't a Hardware expert, but having investigated to this point, he was certain that the operation of Sight was probably very difficult to track in reality.
It's like a Ghost Device.
His investigation was not merely to satisfy his own curiosity.
It is important to ensure that operations that affect the field of Sight do not leave any traceable external traces.
Of course, this was his own computer, and it was unlocked, so it was basically unsecured. However, he hadn't used the computer's login password during the experiment.
To what extent the new capabilities of the Sight can operate on protected electronic devices remains to be further tested.
But at least with this new capability, Han Luyi can increase the speed of annotation.
Sight scanning of the corpus and generating annotations, although labor-intensive, is basically a one-second process.
The previous bottleneck was Han Luyi's output speed.
Typing or voice input, whether it's the latter or the former, is too slow for tens or even hundreds of thousands of data entries.
The new ability turned Han Luyi into a computer capable of outputting data.
He checked his energy level; the operation he had just performed had cost him about 50 points, roughly the same as the energy he had spent analyzing the corpus.
If you use up all 1000 energy points, you can output about 10,000 data entries.
It won't take more than a few minutes.
Energy levels recover slowly, and can be fully restored after a night's sleep.
The previously limited English Annotated Data is no longer a problem.
Zhao Wenyuan has been pushing hard for English annotations lately, Han Luyi quickly generated more than 7,000 annotations. Adding them to some of his own previous annotations, he made a total of 10,000 annotations and uploaded them to the company's encrypted cloud drive.
Zhao Wenyuan, who was working at his desk upstairs, saw a new upload notification, stopped what he was doing, and quickly glanced at the content.
Indeed, the quality of these English annotations is no lower than that of the previous Chinese annotations.
He nodded in satisfaction, without making any attempt to contact Han Luyi to confirm.
He's used to it.
The old man really had a way; he could get the resources he needed just by asking.
What would you like to order next time?
The Main Conference Room on the 48th floor of Prosperity Group's headquarters.
When Lin Shaofeng walked in, the conference room was already quite full.
People from the technology, product, marketing, legal, Public Relations, and even the group's strategic investment department were all present.
This is not common within Dingsheng.
Dingsheng is a giant, not a startup.
The advantage of giant companies is their abundant resources, but the disadvantage is their numerous layers of management. Most projects, once they reach a certain stage, have to go through layer after layer of reporting, making progress extremely slow.
But today is different. Sitting in the main seat in the conference room today is Lü Yun, the chairman of Prosperity Group.
He was merely an observer at the meeting, which was chaired by CEO Zheng Xiaobo.
But the mere presence of Lü Yun is enough to demonstrate the high level of this meeting.
Today's meeting is to produce results.
Lin Shaofeng was the main presenter today. He connected his laptop to the large screen in the conference room, and the screen lit up.
The slide had a large title: Kun Yuan.
Below is a line of English subtitle - KunYuan.
Zheng Xiaobo stared at the screen, his face expressionless.
In the AI field, Dingsheng is not lagging behind among the major domestic companies, and the launch of Qianyuan was also very successful. However, the current problem is that it does not yet have a product that can be presented internationally.
Regarding AI, Lü Yun has his own judgment, which he has shared with Zheng Xiaobo. He believes it's a winner-takes-all game, and in order not to be swallowed up, Dingsheng must become a winner. Dingsheng's competitors are not domestic, but international.
This also made Zheng Xiaobo realize that Lü Yun was not satisfied with his current performance.
He was also anxious; achieving results in the AI field was what the board of directors, specifically Lü Yun, expected of him.
But he knew that rushing wouldn't help; things had to be done step by step.
Kun Yuan is the first step.
The training results came out yesterday. Several public test sets were completed, but the results haven't been released online yet. Lin Shaofeng has already privately reported the results to Zheng Xiaobo.
Today's meeting is a demonstration and an information session, not a decision-making discussion.
"Boss Lü, Boss Zheng," Lin Shaofeng said bluntly, "Kun Yuan's internal assessment has been completed. Judging from its overall capabilities, Kun Yuan 1.0 has entered the top tier globally."
The opening emphasized the achievements, Lin Shaofeng glanced around at the audience, paying particular attention to Lü Yun's reaction, but couldn't discern anything.
He pressed the page turner in his hand, and a summary of data appeared on the large screen.
Mathematics, code, multilingual comprehension, long context, logical reasoning, knowledge-based question answering, intent understanding.
Each item has a score, ranking, and benchmarking model.
"If we only look at general capabilities, Kun Yuan 1.0 is currently behind GPT's latest flagship model, but the gap is not large," Lin Shaofeng said. "Compared with the latest versions of Gemini and Mercury, we have each won some projects. Especially in Chinese tasks, enterprise scenarios, and complex instruction compliance, Kun Yuan's performance is outstanding."
Then he switched to the next page.
The title of this page is: Special Project Assessment of Chinese Intent Understanding.
The screen displays an even more exaggerated set of data.
Kun Yuan 1.0 outperforms all benchmark models in all aspects, including understanding complex Chinese requirements, identifying implicit conditions, completing business processes, and judging users' true intentions. And it's not just a slight lead, but a significant leap forward.
Fault compaction.
A low murmur of discussion broke out in the conference room.
Most people know what this result means.
Large Model research is a latecomer in China. If we can create the first Large Model that reaches the top in the world in a single field, it will be of great significance to Dingsheng's industry status.
Lin Shaofeng then explained the specific evaluation criteria, including publicly available test sets, internally built question banks, third-party publicly available task sets, and anonymized data from real-world enterprise scenarios.
"The biggest problem with domestically produced Large Model is that they all look too similar. When applied to real-world business scenarios, they often fail to understand human language. If a user says, 'Help me with this,' the model doesn't know what 'this' means. If a user says, 'The boss wants to see this,' the model doesn't know whether to focus on risks, results, or reporting standards."
Many business line managers in the conference room nodded in agreement; they all had a deep understanding of these scenarios.
Most Large Model look impressive during demonstrations, but when actually integrated into business systems, they resemble a recent graduate intern.
My business acumen is far inferior to my coding skills.
This isn't the fault of Large Model; the code itself is verifiable—right is right, and wrong is wrong. But business logic isn't like that; there are many ambiguous aspects.
During the Q&A session, the head of the marketing department was the first to speak.
He knew that the purpose of today's meeting was to plan the release of Kun Yuan.
The data Lin Shaofeng just presented looks impressive, and it's one thing to show it off, but its true value is hard to say.
He needs to figure out Kun Yuan's true level before he can formulate a corresponding plan.
"Mr. Lin, this result is amazing. It's not that I don't trust evaluations, but could we do a demonstration to see the effect in a real-world scenario?"
As soon as the marketing manager asked that question, Lü Yun turned her gaze over.