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 148

Wait Until I Give It a Better Brain (Bonus chapter for alliance leader [Spellcaster])

2,832 words 9 min read 1 view 1mo ago Source ​
Skip to the end of the chapter
Prev
Next

Jiang Yixin had just put her bag away and placed it on the table when she looked up and saw Su Niannian walking quickly towards her from the other end of the corridor, holding her phone in her hand with the screen unlocked.

"What was the video you just sent me made of? Was it all made using our toolchain?"

"Hmm." Jiang Yixin nodded, "It's a combination of Heavenly Craft, BugKiller, and Mercury's API, with a monitoring process added in between." "Let me see the situation."

Upon hearing Jiang Yixin's heart skipped a beat.

She had already posted a 10-second accelerated screen recording with three-segment compressed text on Lark. According to normal procedure, the leader's evaluation after reading it would have several levels. "Received," the leader said at Monday's meeting, "I've read this, and I think it's good."

"Received." I also looked at this one, and it's just so-so.

"Has XX finished?" Instead of commenting on the message itself, the person asked something else, which suggests they either hadn't seen it, or they had and disliked it. President Su someone who has seen a lot.

Now she's actually rushed over, wanting to see it again in person.

Is this thing bigger than I imagined?

"Okay." Jiang Yixin took out her laptop. "I'll set up the environment again."

Jiang Yixin turned on the screen and unlocked it. The screen still showed the previous demo page, and the log window was stuck at the last line. "Preview address generated."

She cleared the Heavenly Craft session, re-pasted the Mercury API key, rolled back all the previous changes, and finally restarted the listening process. In short, she tried to restore the system to its initial state.

"Would you like to enter a task?"

Su Niannian had already moved a chair and sat down next to her, her eyes fixed on the screen. "Reproduce what you just did."

Jiang Yixin typed out that line of text.

"Add One dark mode toggle to the settings page."

The log has started scrolling.

Su Niannian didn't speak, but leaned forward slightly, her eyes fixed on him.

Thirty-one seconds later, the preview link appeared.

The "Dark Mode" switch lies obediently in the upper right corner of the settings page. Tap it to switch the entire project to the night theme, and tap it again to switch back. Jiang Yixin glanced at Su Niannian.

Su Niannian didn't look at her.

After a while, Su Niannian asked, "Is there a risk of data leakage with the current architecture?"

"No," Jiang Yixin answered quickly, having thought it through carefully beforehand. "All the code analysis was done using Heavenly Craft; Mercury was only used for task planning."

Su Niannian glanced at Jiang Yixin; the little girl has a good sense of safety.

For many companies, the biggest problem with using AI is the risk of data leakage.

They can promise not to look at it, but once the data is handed over, whether or not you look at it is no longer up to you.

So the Source Code is actually still in a pretty good position, at least Heavenly Craft can use it.

"Then let's switch to a different codebase."

"Change it to what?"

"Open Things)."

Jiang Yixin's fingers stopped on the keyboard.

Open Things the flagship product of Source Code Technology, with over 100,000 lines of code; it's not a toy project for practice.

"Should Open Things directly?"

"Hmm, which core module?" Su Niannian turned to look at her, "Just pick one from those branches that are under development."

Jiang Yixin didn't double-check and pulled up the code.

She skillfully switched to a developing branch of the front-end development Open Things in the workspace, then pasted the path into the monitoring process's configuration. "I'll feed the task."

Jiang Yixin was nervous for the AI after seeing the line of text she typed:

"The queue logic of the project, task, and member modules is abstracted into a general queue component, which can be reused and integrated in the three modules without affecting the existing behavioral logic."

Jiang Yixin gasped.

When this kind of task is assigned to a team, even senior developers will first hold a meeting with several people to discuss for half a day, focusing on the differences in the existing logic of the three modules, how to design the common interface, the steps involved in the migration, and how to handle testing.

The code currently in use is the hardest to modify.

Because you don't know how it will break if you touch it.

Is it really possible to make it?

Let's leave Jiang Yixin's feelings aside for now.

The log has started scrolling.

Scan the project structure.

Locate relevant files: Read services/approval, services/task, services/epense. Planned changes:...

The log entry stops here; it seems to be stuck.

Su Niannian and Jiang Yixin stared at the screen together, and just as they were about to lose patience, a new log appeared.

Planning changes: Generate general state model, generate migration scripts, generate abstract service layer, generate adapters for all scenarios...

Start coding.

Test component execution failed (133).

Re-encoded: 1/10.

Test component failure: 117.

The log started repeating.

Re-encoded: 10/10.

Test component execution failed (117).

Once the maximum number of retries is reached, the program will exit.

Jiang Yixin stared at the screen, her heart slowly sinking.

A dozen minutes later, the monitoring process stopped and retried, Jiang Yixin's heart, which had been hanging in suspense, finally died.

Su Niannian didn't speak, Jiang Yixin whispered first, "It can only do simple things; it can't figure out anything too complicated."

It sounded like he was talking to Su Niannian, but it also sounded like he was talking to himself.

I thought I had created something amazing, but I got excited too soon.

Su Niannian took out her phone, but after thinking for a moment, she put it down again.

"Let's go, take the laptop." Su Niannian stood up and said to Jiang Yixin.

"Huh? Where to?" Jiang Yixin didn't react.

"Go find Boss Han."

After saying that, Su Niannian walked towards Han Luyi's office.

Jiang Yixin quickly grabbed her laptop and followed.

Han Luyi was still in his office.

Standard data.

Nuclear-powered cattle and horses can be described as such.

Sight's energy level is still holding up, but his own "energy level" is starting to give out.

It's really unrealistic for one person to do this level of labeling; three days and eight thousand labels have almost worn him out.

He slowed down the annotation process while considering how to scale this annotation capability.

There's already some progress.

Just then, there was a knock on the door, and he looked up.

It was Su Niannian's voice. "Boss Han, it's me and Xiao Jiang."

"What is it?" Han Luyi stood up and went to open the door.

"Xiao Jiang made something new." Su Niannian stood outside the door, followed by Jiang Yixin carrying a laptop. "You need to take a look." "What is it?"

Su Niannian forwarded Jiang Yixin's Lark to Han Luyi when she received it, but he hadn't had a chance to read it yet.

It's okay. Su Niannian asked Han Luyi to open Lark and quickly read through the message. Then she asked Jiang Yixin to turn on the computer and describe the complex test scenario from earlier.

In just a few words.

Han Luyi has finished drawing the mental framework.

He was a little surprised.

Unexpectedly, while I was busy making models and annotating these past few days and didn't pay attention to company matters, my employees started something big on their own.

It wasn't just surprise, it was delightful delight.

Ryan a new paradigm, a new direction that various companies in Silicon Valley are focusing on, Jiang Yixin has created a half-finished product herself. "Did you make this yourself?" Han Luyi asked.

Jiang Yixin was a little embarrassed. "No, there's also Qian Xiaole."

Han Luyi moment and vaguely remembered him; he was also a young man.

Young people have fresher minds.

Show me the log from the last round.

Jiang Yixin quickly retrieved the previous record and placed the computer in front of Han Luyi.

Han Luyi opened his Sight.

It's been a long time since I've scanned any code.

The log unfolded before my eyes, line by line, with a transparent layer of analysis superimposed on the Sight, and varying shades of color appearing on it.

Han Luyi's gaze jumped to the large, recurring patch of red in the middle section.

Plan once, fail to execute, roll back. Plan again, fail again. Plan again, fail again.

Ten times in total.

He scrolled back to see the differences between each plan.

The first few times I followed the same path, only changing some details; then one time I "remembered" that something was wrong and added a layer; then I forgot what I had figured out before and started all over again.

Pressing down one gourd only resulted in another.

What you think through at the beginning, you forget later when you write it down.

I flipped through it section by section, from beginning to end, without ever truly "thinking it through before taking action."

He glanced at the architecture as well.

The problem isn't with the architecture.

Han Luyi saw the root cause.

The model is not smart enough.

This architecture is complex and not easy to implement.

But it's also very simple in a way.

The Large Model is like the human brain.

Traditional chat-based interaction methods only equip this brain with reading and writing abilities, such as ChatGPT and Google's Gemini, where the only interaction method is chatting.

As the capabilities of models have become more and more advanced, people have begun to explore whether they can do more, by adding hands and feet.

For example, Open Things) equips models with the ability to call pre-made component libraries, thus enabling the creation of all sorts of apps.

For example, the automated workflow that Neus is developing, and the new software that Jiang Yixin just demonstrated, equip the model with eyes that can directly read code on the computer and hands that can modify code, so that it can work like a development engineer.

There's actually a lot of room for imagination here, because you can continue to give it other different hands and feet.

For example, if you enable it to send and receive text messages, it can send good morning, good afternoon, and good night messages to your crush every day for you.

Even opening an online store or trading stocks is not a problem.

But why hasn't such a product appeared yet?

The industry as a whole is still in the early stages of exploration, that's one aspect.

But the root cause is still the same as I said before.

The model is not smart enough.

Tell it to keep an eye on positive news for new energy, and to buy more shares as soon as the news comes out.

Before the market closed, it dug up a news article in the financial news section about CATL reaching a deep cooperation agreement with Apple, and then immediately gambled away the two million in your account.

The next day you open your account: it's hit the daily limit down.

Upon further investigation, it turns out that the "deep cooperation" report wasn't actually from official media, but rather a piece of fake news spread by some shady self-media outlet to grab attention. If you ask them, they might even meticulously explain the intricate logical relationship between this news and the purchase decision.

A model that frequently makes silly mistakes, once given hands and feet, can cause far more harm than a model that only has a mouth.

There are many misunderstandings among people outside the industry about what constitutes a smart model.

Some people say, "Isn't a so-called Large Model just a probability generator that guesses what the next word will be? How can you expect something like that to have intelligence?"

This statement is actually correct, but it is easily misunderstood.

Because this probability is not looked up in a flat table.

It is calculated in a huge multidimensional vector space, where each semantic segment and each logical relationship corresponds to a direction in this space.

The model inputs the text you need, compares and integrates it with everything it has seen in this space, and finally focuses on a single point.

The more input, the more precise the target; the higher the dimension of space, the clearer the distinction between subtle differences.

What do we use to build dimensions? We use layers. The deeper the layers, the more the final result is refined before it's even released, and each iteration better aligns with our true intentions. This is the core reason why Large Model are smarter than small models.

From this perspective, it wouldn't be surprising if the model did indeed develop "intelligence," since when the human brain thinks, it's just the movement of electrical potentials in neurons.

The connections between neurons in the brain are in the trillions, while the largest AI models currently have only a trillion parameters, several orders of magnitude less.

Moreover, the power consumption of the brain is less than one ten-thousandth of that of an AI model.

Therefore, in the past few years, AI companies around the world have been spending money like crazy to make their models bigger and deeper.

However, once Computing Power reaches a certain level, the improvement brought by each Add One begins to decrease, while the cost becomes increasingly expensive.

This is where the value of labeling comes in.

Good labeling is like telling the model in advance, "For this type of problem, you don't need to blindly search the entire space; you only need to search within this small area." Once the distribution space is narrowed, a limited number of layers can produce more accurate results.

The repeated red marks in front of Han Luyi are precisely the opposite of this matter.

To take this task specifically as an example, we need to extract the queue logic that is running in the three existing online modules into a set of general components, and ensure that no existing behavior changes after the migration.

The Model Training data contains code for a single queue and general components that are abstracted at the business layer. However, the requirement to "perform lossless reconstruction across three already running modules" is not seen enough, and the annotations are too sparse.

It also needs to accommodate the implementation of three modules, test cases, and "do not modify behavior" constraints all at once, which makes the context window insufficient. The space is large, sparse, and inaccurate.

The model can only guess blindly segment by segment. After trying for so long, I still haven't found that narrow correct solution.

But this doesn't mean the product is headed in the wrong direction.

It simply doesn't do enough optimization based on the existing model capabilities.

But this is the direction in which AI will develop in the future.

Empower AI tools so that they can take on more tasks.

Han Luyi was really looking forward to it.

"Xiao Jiang," Han Luyi said.

"Hmm?" Jiang Yixin cautiously.

"The product is very well made. Find a few people to refine it and use it internally first." Han Luyi said, closing his laptop and handing it to Jiang Yixin. "But the scene that President Su just tested..." Jiang Yixin hesitated.

Su Niannian smiled slightly, "All tools have problems when they first come out; some things they can do, and some things they can't. Let's use them as internal tools first and collect user feedback." Han Luyi added, "Especially the examples of failures, they're even more useful to me."

Jiang Yixin suddenly remembered something and asked, "What about the cooperation with Neus?"

Su Niannian glanced at Han Luyi, then turned to Jiang Yixin and said, "You'll be in charge of both, but prioritize this one. If you can't handle the cooperation side, you can come to me."

"Uh-huh!"

"Let's go back."

Jiang Yixin, carrying her laptop, walked to the door. She glanced back; neither of the two bosses had moved.

She closed the door from the outside.

Only Han Luyi and Su Niannian remained in the conference room.

"Nian Nian, I suddenly have a feeling."

"What?"

"Once you hire the right people, good things just keep popping up on their own."

"That feels so good."

Han Luyi smiled as soon as he finished speaking.

Is this what happiness is for capitalists?

Su Niannian smiled with satisfaction, "Xiao Jiang is indeed a talent."

The meeting room door closed.

Han Luyi while.

The night outside the window had deepened, and the reflection of the river in Qiantan swayed on the glass.

He recalled the surprise Jiang Yixin had brought him today.

The biggest limitation now stems from the capabilities of the model.

Since that's the case, I'll give it a good brain.

Han Luyi took out his phone, which showed WeChat Cheng Yuan had sent that afternoon:

"Boss Han, we were very interested in the prototype you brought last time. We'd like to know your intentions for cooperation. Shall we connect first?" It took you two days to reply. I really thought you weren't in a hurry.

Han Luyi pulled out the draft of the terms that he and Gu Siyue had discussed earlier.

As Lawyer, Gu Siyue's assessment of this clause was. "Isn't that a bit too harsh?"

Han Luyi's assessment was no.

In business negotiations, whoever controls the unique resources wins.

The advantage lies with me.

Send.

End of Chapter
Prev
NovelSummaries
Next

Explore the wiki

Han Luyi
Su Niannian
Jiang Yixin
Gu Siyue

Comments

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