Rebirth: Starting from Lighting Up the Tech Tree

Chapter 117_Encountering a Bottleneck Again

Chapter 117 Encountering Another Bottleneck

The air in the conference room seemed to freeze.

The whiteboard was covered with formulas and model parameters, all three erasers had been used, and crumpled draft papers were scattered in the corner. It was the fourth day of continuous work. Shen Yiming and Fang Ze were each guarding two laptops, with the thirteenth optimization plan running on the screen.

When the system log results popped up, Shen Yiming glanced down at them and fell silent.

"89.3%." He read the number aloud, his voice as calm as if he were reading an insignificant weather forecast.

Fang Ze looked up from the other computer: "The same as last time?"

"It's a difference of 0.01 percentage points." Shen Yiming rubbed his temples. "We're going in circles."

Thirteen plans, thirteen failures. Each failure was graceful, with the accuracy rate remaining stable between 89.2% and 89.4%, like an invisible glass ceiling that you can push hard but won't break, yet won't fall.

This feeling is more agonizing than simply returning to zero.

Fang Ze closed his laptop, leaned back in his chair, and stared at the ceiling: "This isn't a problem that can be solved by hard work. Compressing the model to 1/30th is already the theoretical limit; compressing it further would sacrifice information. With less information, accuracy naturally suffers."

"I know," Shen Yiming said.

"Then why are you still running?"

"I don't know what else to do."

This statement isn't meant to be self-pitying; it's simply the honest confession of a technical person facing a theoretical bottleneck. Shen Yiming is an academic; he can clearly tell you where the wall is and what its material is, but that doesn't mean he has a way to break through it.

There are only three days left until the technical solution is submitted.

Zuo Cheng sat by the window, a cup of coffee that had gone cold beside him. He didn't participate in Fang Ze and Shen Yiming's discussion; his eyes were looking out the window, but his thoughts had already wandered elsewhere.

He's calculating.

There's still one card left in the integration of technology.

Of the leaves he currently possesses, hyperparameter auto-optimization is a newly copied version from Chapter 115, model compression optimization was obtained during AI branch activation, and edge AI inference is also one of the branches. If these three leaves are merged, theoretically, a new leaf with stronger overall capabilities can be generated, resulting in a qualitative change in performance compared to the original.

But integration comes at a price.

Consuming 5 points is a minor issue; the problem lies in the uncertain outcome. The system will automatically calculate the fusion direction. Success will result in a new blade, while failure, though extremely rare, does occur. Furthermore, the fusion's technological fusion capabilities require a 48-hour cooling-off period after fusion.

The proposal will be submitted in three days, which is just enough time even under the most optimistic calculation.

If you gamble and lose, you lose everything.

Zuo Cheng tapped his fingers lightly on the table, his rhythm even, like an unconscious countdown.

Fang Ze noticed this action but remained silent. Having known Zuo Cheng for so many years, he knew that this movement meant Zuo Cheng was making a significant decision.

Shen Yiming began organizing the drafts on the whiteboard, recording each crossed-out plan in a document to explain why those approaches were unworkable in the future. This was a habit ingrained from his academic training: even failures must be documented, for failure itself is data.

"How many quantitative strategies did you try in total?" Zuo Cheng suddenly asked.

Shen Yiming turned around: "Seven types. The one with dynamic precision is the sixth type, and it works the best, stopping at 89.3%."

Have you tried mixed precision?

"I've tried it, and it's worse than dynamic accuracy; the accuracy dropped to 87.6%."

Zuo Cheng nodded and didn't ask any further questions.

This wasn't the question he really wanted to ask; he was just using questions to steady his thoughts and keep his brain working, rather than getting caught up in that gamble.

Car headlights swept past the window, casting an arc of light across the conference room glass before vanishing in an instant.

"Brother Cheng," Shen Yiming said.

Zuo Cheng snapped out of his daze and looked at him.

"What are you thinking about?" Shen Yiming's eyes held a hint of inquiry, but more so, weariness. "I've seen you staring out the window for almost twenty minutes."

Fang Ze also turned his head, his waiting gaze calmer than Shen Yiming's, but he was also waiting.

A moment of silence fell over the meeting room. Outside the window, the city lights of Hangzhou stretched as far as the eye could see, and atop a distant office building, a red signal light slowly rotated, flashing rhythmically and silently every few seconds.

Zuo Cheng withdrew his gaze.

"I'm thinking of a bet," he said.

Shen Yiming didn't press further. He could roughly sense the weight in Zuo Cheng's words; they were the kind of things that, once spoken, meant you had to actually take the gamble, not just casual complaints.

Fang Ze turned back and restarted his computer, but instead of running any new tests, he just stared blankly at the 89.3% figure.

Zuo Cheng stood up, walked to the whiteboard, picked up a marker, and wrote a line in the upper right corner of the blank space:

Target: 95%

Then it added below: "Current: 89.3%"

He stared at the two lines of numbers for a long time, as if he were negotiating with them.

The difference is 5.7 percentage points.

Conventional methods have been exhausted; only two paths remain. One is to lower the target, accept 89.3%, and then compensate with other strengths at the review meeting. The other is to play our trump card and take a gamble.

Zuo Cheng capped the marker, put it back in the whiteboard tray, and turned around.

"Yiming, continue the test run tomorrow," he said. "I'll think about it again tonight."

Shen Yiming nodded without asking anything.

Fang Ze tidied up his computer, casually tossing the empty coffee cup into the trash can, the motion so practiced it seemed like he'd done it countless times. And it probably had.

Zuo Cheng was the last to leave the meeting room.

The only sound in the corridor was the hum of the lights. He didn't return to his seat immediately, but stood in the corridor for a while. His points were 262, the technology fusion cooling-off period had long since reached zero, everything was ready, only his approval was needed.

He recalled Zhou Henian's words: "If the proposal doesn't work, the six-to-one result at the review meeting might be reversed."

The Sky Dome Phase IV project is no ordinary project. The autonomous coordination of 480 satellites is something no commercial project globally has ever accomplished. If 402 can secure this deal, it won't just be a single contract, but rather establishing an unshakeable position in the entire edge AI field. If they fail, Lin Jianhua's team will be closely monitoring every detail, looking for an opportunity to turn the tide.

In his previous life, he had experienced the feeling of being trampled underfoot and unable to get up. It was a memory he never wanted to relive.

89.3% is not enough.

Zuo Cheng stood in the corridor for a few more minutes, waiting for the last few mathematical steps in his mind to be clear. Three blades, automatic hyperparameter optimization plus model compression optimization plus edge AI inference, a three-way fusion—this was the strongest combination he could currently control. The direction of the fusion was uncertain, but it was definitely centered around the core goal of "improving the efficiency of edge AI."

This judgment is not blind; it is the most reasonable deduction after the three branches intersect.

Zuo Cheng took a deep breath, found the entry point for technology integration in the system panel, hovered his fingertip over the selection interface for a second, and then moved it away.

Not now. Tonight, he needs to calmly and carefully consider this decision, not because he's still hesitant, but because this stake is too high to be taken lightly.

The stakes are clear; all that's needed is a quiet night and the move made when you're awake tomorrow.