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 162

You know nothing

3,109 words 10 min read 1 view 1mo ago Source ​
Skip to the end of the chapter
Prev
Next

The moment the Sight opened, the scrolling log on the screen changed.

Every piece of data from the cleaning pipeline appeared with a thin layer of color in Han Luyi.

Clean, informative data glows with a pale blue light, flowing smoothly through the pipeline like water; junk data that has been correctly intercepted by the rules is gray, cleaned up at a certain node, and no longer processed by the following stages; while problematic data that has not been intercepted glows with a glaring orange-red light, standing out conspicuously amidst the blue water.

Han Luyi also noticed another color: occasionally a few data points flashed with dark blue, but they stood out against the gray data. They were mistakenly judged as garbage and thrown out by the rules, but they were actually valuable content.

Just like standing next to an assembly line, others can only see the products moving on the conveyor belt, but Han Luyi's eyes can perform quality inspection on every product, and even on every conveyor belt itself.

Han Luyi stared at the screen for about ten minutes, starting with the most obvious orange-red color and gradually figuring out the pattern.

There are mainly three issues.

The most common problem is that pure spam hasn't been cleaned up. This includes low-level spam like "spinach GG," "sexy traffic generator," and keyword stuffing. The existing rules only match keywords, but there are too many variations.

Using homophones, replacing words with emojis, or splitting sensitive words into two parts and stuffing them into the surrounding text—simple regular expressions simply can't prevent this. This type of data is highlighted in the brightest orange-red in the Sight, and there are quite a few of them.

Then there's semantic redundancy. Two paragraphs use completely different words, but express almost the same meaning, which cannot be identified based on keyword and format rules. The last one is the most insidious: the disguise of low-quality content. Some content has the characteristics of high-quality content in terms of format, length, and keyword distribution, but it is actually plagiarism or machine-generated filler, and many of the contents contain factual errors.

These three combined account for nearly 70% of all the data.

Han Luyi exited the Sight and began revising the script.

Using traditional garbage sorting rules to achieve full coverage would be a huge undertaking, so Han Luyi directly changed his approach.

The model for Glutinous Rice Ball is still there, right? Just send the data to Glutinous Rice Ball for testing and intent recognition. No matter how you vary it, how you break down words, or how you use emojis, as long as the final content is "guided to click" or "induces payment," it will be flagged.

Similarly, using Glutinous Rice Ball to simplify the data makes merging much easier.

The last step, identifying the disguised content, was a little more difficult, but it wasn't a challenge for Sight. With Sight's hints, Han Luyi added a layer of filtering based on the degree of confusion.

For texts with genuine information, language models have a higher degree of uncertainty in predicting the next word; while plagiarized and filler content, due to their formulaic nature, has a much lower level of confusion.

After making the changes, I ran it again.

Look at it again using your Sight.

The orange-red color almost disappeared, leaving only a few scattered data points outside the boundary conditions faintly visible in the blue water. The dark blue false positive data also decreased significantly, with only one or two occasionally flashing by. Han Luyi noted them down and then fine-tuned the parameters again.

After the third round of running, all that remained in my Sight was a clean blue.

It was already 2 a. m. when I turned off the Sight.

He leaned back in his chair, looking at the final cleaning report.

The effective retention rate of the original data dropped from 41% as previously noted by Zhao Wenyuan to 29%—but this 29% is truly clean. At the same time, the false positive rate dropped from 12% to less than 3%, and a batch of good data that had been mistakenly killed by the previous pipeline was also saved.

While improving cleaning quality, the processing speed has also increased significantly, taking only about half the time compared to before.

Han Luyi sent a message to Zhao Wenyuan, sending him the revised script and the cleaning report together.

"Wenyuan, try this version tomorrow. I've made three changes: the garbage sorter, semantic deduplication, and perplexity filtering."

Zhao Wenyuan replied instantly, "You're still awake?"

The second message came right away. "Let me see."

Half an hour later, Zhao Wenyuan sent a long message:

"How did you manage to create this in one night? Most teams would spend two or three months just iterating through the cleaning pipeline. The one you just sent me has better efficiency and false positive rate than the industry benchmark! At this efficiency and accuracy, full data cleaning should be completed in about a week! Pre-training can begin after a week! We should see the first version of the model in two months!"

It was rare for Zhao Wenyuan to post so many exclamation marks. With such high-quality raw and Annotated Data, he was really looking forward to seeing the training results.

However, Han Luyi was already asleep at this time.

Rewind time by a dozen hours, to the other side of the Pacific Ocean.

Pacific Ocean Time (United States)

December 27th, Sunday, afternoon.

John's detached house is located in Los Altos Hills.

John Snow is the CTO of NeusAl.

Unlike CEO Ryan, who has a PhD in Artificial Intelligence and has worked in AI Research Institute, Ryan has a PhD in artificial intelligence.

John only has a bachelor's degree. He started his entrepreneurial journey before graduating from Stanford. He founded or co-founded a total of five companies, two of which were acquired by Google and Meta respectively. He left those companies a few months after the acquisitions to continue his entrepreneurial ventures.

John has extensive connections in Silicon Valley's venture capital and tech circles, and he's also an investor himself.

Someone like him, who has already achieved financial freedom, no longer works for money. The reason he accepted Ryan's invitation to join Neus Al is because he is optimistic about this direction and also has faith in Ryan himself.

John's house has a driveway that goes uphill from the main gate, with four or five cars parked there. You can only see the yard after you walk to the back of the main building. The backyard is very spacious, with an excessively wide oak tree whose shade covers almost half of the lawn.

A barbecue party is being held here.

The party started at 2 p. m., and by 3 p. m., the guests' glasses had been emptied two or three times.

Christmas lights still adorned the pillars and a few trees. In the bright daylight, the lights were on, but barely visible; a gust of wind caused them to sway slightly. Over at the grill, a professional chef was still turning the ribs, the grease dripping into the charcoal with a sizzling sound.

In California in December, the temperature is around 18 degrees Celsius. The weather is sunny with a slight breeze. It's not cold, but not cold enough to warrant taking off your coat.

John hosts a small party like this every year after Christmas, inviting people he knows well in the tech industry, and invitees can bring friends.

This event is considered a small salon for Silicon Valley's tech and venture capital circles.

When Ryan arrived, John was by the grill discussing the cooking time of the ribs with the chef. When he saw Ryan, John raised the tongs in his hand to greet him, then returned the tongs to the chef and walked over.

Ryan and John are old friends who have known each other since their days at Google.

"You always bring this." John glanced at the OpusOne in Ryan's hand.

"Your drinks are never good." Ryan chuckled and handed him the bottle. "Open it for me."

John took it, found a corkscrew on the nearby table, skillfully unscrewed it, poured the wine into a glass decanter, and placed it in a cool, dark place. "Okay, this one's definitely better than mine," John said, "But it'll take at least half an hour."

"I know." Ryan grabbed a bottle of beer from the ice bucket. "Just wait."

The two walked around the yard and greeted several people they knew.

John's social style is different from Ryan's.

Ryan blended in naturally with the crowd, charming and sociable, able to chat with anyone for three to five minutes before gracefully withdrawing, leaving no one feeling neglected. John, as the host, was more casual; sometimes he would chat with someone for a long time, sometimes he would simply nod and walk by, entirely depending on his mood and not caring what the other person thought.

When asked about Neus's recent activities, Ryan would always smile and say "not bad," nothing more, nothing less. John would listen in the background, occasionally adding a trivial comment; the two worked together seamlessly.

Ryan then found a chair near the grill and sat down, while John went to attend to the other guests.

Not far away, a middle-aged man wearing glasses was telling two people next to him about a logistics AI startup he had recently come across, "I asked them to show me a prototype demo, and guess what? The CEO opened his computer, showed me a slideshow, and said, 'This is our vision.'"

The people around burst into laughter.

On the other side, a blonde, Pangpang woman in a dark long dress leaned against the wooden railing, looking exhausted. She complained to her friend about not being able to find a technical partner, "I've received dozens of resumes, but none of them are suitable. Either they lack experience, or they're not satisfied with the salary. These people are too demanding these days. They only offer technical skills, what more do they want?"

John walked through the crowd and put his hand on her shoulder.

"Don't talk about this at the party."

"So, what should we talk about?"

"Let's talk about food." He gestured with his chin towards the grill.

She glanced up at him timidly, raised her glass, took a sip, and changed the subject.

John's yellow Labrador was lying on the edge of the lawn, belly to the ground, chin resting on its front paws, eyes closed, only its tail occasionally swishing.

Two children were playing a game of some kind, chasing each other across the lawn, and after a while, the ball disappeared.

An investor he'd chatted with briefly before approached Ryan and extended his hand. "Ryan, Long Time No See. How's Neus doing lately? I heard Mercury Model 3.1 is doing research on inference?"

"Not bad." Ryan gave a handshake. "We've gotten some results on a few vertical scenarios."

When are you planning to publish the paper?

"It might not be published," Ryan said. "It's mainly product-level validation; it might not be worth publishing a paper on."

The other person nodded, didn't ask any further questions, and was pulled away by another person next to them.

John walked back, sat down on the chair next to him, took the bottle from Ryan, poured himself a glass, and put the bottle on the floor. "What you said earlier was true?"

Which sentence?

"The part about not publishing papers."

Ryan raised his glass. "What's the point of publishing papers? Recruiting people, boosting rankings, getting citations from peers—we don't lack those things now." John didn't speak, but also raised his glass.

Their glasses clinked together.

John watched Ryan drink and recalled the first time he met him.

It was at a small technical sharing session within Google. Ryan was a researcher at the time, and he was talking about a paper he had just published on optimizing attention mechanisms.

The presentation was excellent, but what impressed John most was not the content itself, but the way Ryan answered questions after he finished speaking.

Ryan never uses platitudes to buy time for thinking, like "good question" or "let me think about it." He tells you directly where your question is right and where it's wrong, and then gives his answer.

That sense of certainty is rare in academia and also in industry.

As the two grew closer, John realized that it might not have stemmed from their shared knowledge of the field.

Ryan, however, doesn't really doubt himself.

This trait made him an outstanding CEO.

Makes decisions quickly, has a clear direction, and can always lead the team forward.

But John sometimes thinks that just because someone doesn't doubt themselves doesn't mean their decisions are right.

Around 5 p. m., the sun was almost setting behind the mountain ridge, and the temperature had cooled down a little.

People came over one after another to greet John, thank him for his hospitality, and say they wanted to leave a good impression.

After most of the people had left, John and Ryan moved two chairs to the poolside and each grabbed a bottle of beer from the ice bucket.

The pool is unused this season; the water is calm, with only the occasional fallen leaf drifting in.

The setting sun painted the water a pale orange hue, while the ridgeline still shone with its last rays of light.

Ryan took a sip and pulled out his phone.

"I'll show you something." He turned the screen to John, "This is the API call analysis from the China company we've been working with for over a month. "John took it and flipped through it.

"At the task planning level, we follow the Mercury Model entirely," Ryan said. "In their product, every time they receive a user instruction, the subsequent thinking part—intent understanding, task breakdown—is entirely ours. They handle the code execution themselves." John slowly swiped the screen without speaking.

"The product experience is great," Ryan continued. "It's very smooth to use, user satisfaction is high, and the data we've collected is also of high quality. But what makes this product smart is the Mercury Model; their own model is essentially a code executor."

John handed the phone back to him. "Their code generation capabilities aren't bad either."

"Not bad," Ryan took it. "But that's useless. At this level of parameters, the difference in code generation is already very small. It's hard to widen the gap further. Intent understanding is Mercury Model's trump card. They know this themselves, which is why they accepted our cooperation." He exited the report and reopened a file.

We are now in the first phase.

His tone was calm, as if he were reiterating a logic he had already run through many times:

"At this stage, let them develop the product, accumulate users, and refine the user experience. We collect revenue while simultaneously gathering data from all real-world task scenarios. This data is the core training material for the next-generation Venus model's Agent capabilities; user scenarios are far more valuable than the baselines we construct ourselves."

John drank a beer.

"By the time the contract expires, their user base will have been built up, and the entire intent layer will be running on the Mercury Model. At that point, we'll increase the revenue share and start charging API fees." He paused. "What are their options? Migrate to another model, then modify the architecture, retrain, validate, and regress the product, or accept the new terms?"

John did not answer.

"They had no choice. Their user base was built on the Mercury Model, and the migration cost was too high, plus they didn't have the time."

"After the Venus model," Ryan put his phone on his lap, "Our Agentic capabilities are backed by real data, and we've thoroughly understood user scenarios. At that time, we can create our own developer tool, using the data we've accumulated during this period, to make a better product than theirs." John looked up at the sky for a while, then shook the bottle of wine in his hand.

Will they find out?

"Yes," Ryan said. "By then, they'll have no other choice."

"I mean..." John looked away, "Do they know now?"

"They probably sense it," Ryan said. "Their CEO isn't stupid; he should be able to see this logic. But they need us. Without the intent-based understanding provided by the Mercury Model, their product won't work. They probably want to build their competitive advantage before the contract expires and then develop their own model."

Is it possible?

"How long does it take to develop a model? A year? At the very least, it'll take at least six months. Our contract expires in April."

The implication is that it's impossible to create a Large Model that can replace the Mercury Model in four months.

Once their contracts expire, they'll be nothing more than fish on a chopping board.

John finished the rest of his beer before asking:

"Once we have enough data, we'll kick them out?"

John remained silent for a moment.

The pool water had darkened; the light orange hue had vanished, replaced by a somber blue. A last ray of light remained behind the ridge, slowly fading away.

The sound of a Labrador barking came from the other side of the lawn. It had probably been startled by something; it barked twice and then fell silent.

"Ryan, have you considered this possibility?" John placed the empty bottle at his feet, "Their CEO, if he's really as clever as you say, might already be working on a model."

Ryan glanced at him. "What are they going to do it with? They have no Computing Power, no data team, no research background. Building a model that can replace Mercury isn't just a matter of writing a few lines of code."

"I didn't say replace Mercury," John said. "What I said is, if he only makes a usable model, not necessarily better than Mercury, just good enough to take over the understanding of intent from us, then your whole plan falls apart."

Ryan didn't answer immediately, but looked down and swirled the last bit of beer in his glass.

"Four months," he concluded. "From scratch, four months to build a usable base model? To build a model that can replace Mercury at the level of intent understanding and mission planning? Do you think that's possible?"

John did not answer the question.

He genuinely felt it was unlikely.

But then he remembered the video he had seen before, the video of the CEO of that China company giving a speech at GAIS.

John also gave a speech at GAIS, albeit remotely.

He felt that person was not simple.

"Actually, we can do it ourselves. I'm confident I can do it better than them," John said sullenly.

"Our resources are limited, and we need to focus on the most important things," Ryan said. "Why not use this free and readily available labor?"

At this point, a smile appeared on Ryan's face:

"You know nothing, John Snow."

"Damn it, how many times have I told you, my name is John, not Jon," John cursed. He hated Game of Thrones.

And that George Martin.

"You're the one who knows nothing, Ryan Miller."

Ryan simply smiled at this childish retort.

He didn't say anything.

End of Chapter
Prev
NovelSummaries
Next

Explore the wiki

Han Luyi
Zhao Wenyuan

Comments

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