Shockingly, a counterfeit factory has produced a 3-nanometer chip.

Chapter 545 Zhang Lei: So this is what it feels like to have cheat codes

A sense of despair washed over everyone in the conference room.

After all, the core challenge of edge voice assistants lies in their massive computing power and size.

The reason why Xiaoyin Assistant is so smart today is because it is supported by a huge server cluster.

After all, we have our own computing power cards, so we don't need to be as stingy as outsiders. But this leads to... cloud AI is a very large model.

Compressing this model into a mobile phone chip while maintaining speed and accuracy is akin to performing a ritual inside a screw casing.

Simply put, it's impossible, or rather, it's simply impossible with current technology.

Most cloud assistants in the industry today are cloud-based transmitters, with most of the calculations done on the server, and only simple wake-up and command uploading performed locally.

If it were completely severed, the already artificially disabled model would become even more...

Even more intellectually disabled.

"Try again, prune the wake word model by another 20%."

Zhang Lei gritted his teeth and decided to give it another try. "If the accuracy drops a bit, so be it. Let's solve the latency problem first."

"But after editing, the false wake-up rate skyrockets..."

"Let's test it first. After all, even if our model's wake-up rate skyrockets, it's still better than theirs."

Just as everyone was preparing to act, the laboratory door was pushed open.

When Shen Fei walked in, the host immediately stopped what he was doing, after all, the boss had come to inspect the work again.

"President Shen?" Zhang Lei was slightly flustered and quickly went to greet him. "What brings you here?"

At this moment, he had already mentally rehearsed the entire research and development process over this period of time, and thought that the boss was dissatisfied with the progress here, so he came in person...

Accountability!

At this moment, a little cold sweat began to appear on his forehead, after all, the salary and benefits of Transsion Company were really too high.

"Don't panic, I just came to check on the progress and see if there's anything I can do to help."

Shen Fei glanced at the test curves on the screen, but instead of pointing out their problem, he smiled and asked:

"I've hit a bottleneck."

Zhang Lei's face flushed, and he was a little embarrassed: "Yes... the model compression has hit a ceiling."

With limited local computing power, the latency cannot be reduced in order to ensure the recognition rate.

To reduce latency, accuracy drops drastically.

We tried more than a dozen pruning methods, but none of them achieved the project's goals.

"Or... we can mainly promote the Yunnan model, and make a simplified version locally that can achieve basic functions."

I believe most people will not encounter such an extreme situation.

In fact, Zhang Lei had been thinking about a question: Xiaoyin's cloud model was already ahead of other peers, and as for local models, other companies hadn't even started a project with a PPT.

Transsion is already far ahead in terms of being able to produce this, so there's no need to pursue an overly extreme response time, such as 0.3 seconds.

Shen Fei shook his head; there was no room for discussion on this issue.

"no.

The core of edge-side voice communication is local response. If we still rely on the cloud, what's the point of this technological iteration?

"Moreover, everyone is working on cloud computing, so the competition is about who has the best servers and the most resources."

Although algorithms can create a significant gap, this arms race is still too small. Old Huang's sales of power cards are still not good; we need to put more pressure on them.

Zhang Lei gasped instantly; he finally understood why his boss had done this.

This isn't about developing a new model; it's clearly about misleading competitors to make it easier to sell computing cards.

At this moment, Zhang Lei suddenly muttered in a low voice: "Boss, isn't what we're doing a bit unethical?"

After all, what we can make, they may not be able to make.

At this moment, some of the resentment in Zhang Lei's heart completely disappeared. After all, people from his Zhang family were best at helping others...

and so……

At this moment, the other members of our group also showed subtle expressions.

Why are you all making that face? I didn't say anything.

Seeing that everyone's morale had instantly soared, Shen Fei also started to cheat.

"I was deaf just now, what did you say?"

With one person taking the lead, others followed suit. After all, the engineers from other companies had suffered a lot during this period, and they wanted to make sure that the engineers from other companies also suffered a bit.

[Engineers from other competitors: ...]

"What's the big deal? It's just a minor hiccup, but we still need to work on the local model properly, since there are some areas where we still need the edge model."

"Please show me your model architecture diagram."

When Zhang Lei saw that his boss wanted to give instructions, he quickly projected the complex model diagram onto the big screen.

After scanning for a while, Shen Fei had a general idea of ​​what was going on.

The team's approach was sound; it was the traditional model pruning and quantization method, and they had basically exhausted all the potential.

If they rely on themselves to slowly work their way up, they might be able to get the hang of it in another six months, but that would take too long, and he can't afford to wait.

He's going to cheat.

So he reached into his pocket, took out a USB drive, and placed it on the worktable.

"Here is a new speech model architecture, along with a corresponding compression algorithm."

You should modify and retrain the model according to this plan.

"New architecture."

Although Zhang Lei was a little confused, he quickly took the USB drive from Shen Fei. After all, the sound transmission experiment was the first law, and the boss's ability should not be questioned.

"right.

This model uses a lightweight dynamic network structure, eliminating the need for brute-force pruning.

It can compress the model size to one-tenth of its original size while maintaining a 98% recognition rate.

One-tenth.

Several engineers nearby gathered around, their faces full of curiosity.

After all, their model has already been compressed to 8GB, so the compressed result would be 800MB?

A model of this size achieving a 0.3-second response time sounds like a fantasy.

"Don't just stand there like an idiot, try to get in."

Shen Fei pointed at them.

Zhang Lei turned around and quickly plugged the USB drive into the computer.

Open the folder, and you'll find not only complete architecture code, but also detailed training guidelines and optimization notes, even clearly marking common pitfalls.

He became more and more excited as he looked at it, after all, this architecture used a method they had never thought of to resolve the contradiction between model size and accuracy.

It's like someone opening a new door in a dead end.

"This...this is incredible..."

"Dynamic routing assignment example, using small models to create the effect of a large model... How come we didn't think of that before?"

Zhang Lei was dumbfounded. Sure enough, with the guidance of a famous teacher, one can instantly gain enlightenment. The ancients were right.

"It's not too late to think about it now." Shen Fei smiled. "I'll give you a week to fine-tune the model. Can you do that?"

"able."

Zhang Lei and the group of engineers nodded repeatedly.

"Don't even mention a week, just give us two days, or even just a press conference, that's fine too."

"No need to rush, this is to generate buzz while the product is in its normal sales period."

However, you can leak the information in advance to attract the attention of engineers in the same field.