Zhao Wenyuan had no clue what to do when he heard the question.
He wasn't actually sure if Boss Han's figure was just something he made up.
From a computer system design perspective, 200 milliseconds is not a very fast number; many high-availability systems even require a response time of less than 50 milliseconds.
Just like the Douzhi Model Zhao Wenyuan is working on halfway through, which is the model based on System 1, he has already boasted to Han Luyi and Jiang Songran that he will reduce the single response time to ten milliseconds.
However, Brain-Computer Interfaces are a completely different field.
However, there is a process to follow when diagnosing system bottlenecks. The first step is to find out where the bottleneck is.
There's an old saying that goes: You can't solve a problem you can't see.
The first step to solving a problem is to see it.
So Zhao Wenyuan asked, "President Meng, how is the processing time of the Mengnao device distributed per cycle? I mean, where is it slow?"
Meng Yao gave a wry smile and said, "From my perspective, nothing is slower, it's just that there's too much to do. Invasive brain-computer interfaces can achieve 100 milliseconds, and I've made some speculations about that. I guess they can directly collect electrical signals by attaching electrodes directly to the outside of neurons, making analysis simple. But our collection method involves the scalp, skull, dura mater, and cerebrospinal fluid in between, so the signal is already weak, and we still have to convert the potential signal into a digital signal."
"After acquiring the signal, most of it is noise. We need to filter it first, and then perform a Fourier transform—just accumulating enough sampling windows for one Fourier transform takes three to four hundred milliseconds. And that's not all; afterwards comes our algorithm model, classifier, and pattern matching. At the current eight hundred milliseconds, I don't see any room for optimization." After listening, Zhao Wenyuan had a general understanding of the Mengnao's computational process.
He almost blurted out, "That's why we need self-supervised training. Transfrmer-based models inherently possess powerful high-dimensional latent spaces. Simply put, we don't need to understand brainwave signals at all."
Meng Yao was stunned after hearing this.
Zhao Wenyuan was still standing in front of the whiteboard, and he used a marker to write down the process that Meng Yao had just described.
Sampling, filtering, Fourier transform, feature extraction, classifier, action.
Meng Yao stared at the whiteboard in surprise. She hadn't expected Zhao Wenyuan to understand the flowchart so well after just one listen. And this wasn't even a rigorous explanation; it was just something she said casually.
After finishing the drawing, Zhao Wenyuan said, "This is a completely handmade concept."
Zhan Xiang frowned, unable to resist retorting, "President Zhao, this is the most mature practice in the field of brainwaves."
"Is that so?" Zhao Wenyuan said. "But in the field of machine learning, this is no longer the mainstream practice."
These words left Zhan Xiang speechless for a moment, but he had nothing more to say.
After all, Zhao Wenyuan is indeed a top expert in the field of machine learning.
"Take Glutinous Rice Balls as an example. When we train them, we don't tell them which features are important. As long as there is enough training data and it's accurate enough, after these signals enter the high-dimensional space, the useful patterns will naturally have a larger proportion. After receiving the brainwaves of this pattern again, regardless of whether it's a signal or noise, the whole thing can be recognized by the general model, and the signal part can also be effectively separated," Zhao Wenyuan continued to explain. "But even if we can get the signal, what does it represent?" Zhan Xiang asked.
In their current traditional approach, this requires one-to-one training and feature extraction on the device user so that the model can understand the intent expressed by an EEG signal.
"What other good methods are there for machine learning?" Zhan Xiang for a moment.
"This is where post-training fine-tuning comes in. We collect EEG data of several typical user movements on-site, and then fine-tune it based on the general model. This allows us to determine which waveform corresponds to which movement."
"How long will this step take?" Meng Yao asked.
Their initial customization phase might take dozens of hours, and since most users cannot use it continuously at a high intensity, it could take several weeks.
Zhao Wenyuan thought for a moment and said, "Using cloud Computing Power, it should take a few minutes to a dozen minutes." How much longer would it take to fine-tune the 0.5b model?
"This fine-tuned model can incorporate most of the above steps, resulting in the new architecture."
Zhao Wenyuan drew a new architecture diagram directly below the previous one:
Sampling, model, action.
Meng Yao's heart skipped a beat, and she asked, "How long does one model calculation take?"
"A few hundred milliseconds, I guess," Zhao Wenyuan said, then before Meng Yao and Zhan Xiang be disappointed, he added, "But we happen to be developing a new model architecture recently, called Douzhi." "Douzhi?" Meng Yao and Zhan Xiang couldn't help but repeat it.
Both of them had lived in the Beijing for many years, so of course they knew what it was.
Naming the models after food items—this President Zhao has quite a twisted sense of humor.
Zhao Wenyuan was unaware that the two senior executives of Mengnao Technology had pinned the blame on him, even though the name was clearly given by Boss Han.
He continued, "This architecture was inspired by the theories of System 1 and System 2. Its purpose is to allow the model to operate in the System 1 mode. If conditions are right, it should be possible to control the single operation time to within fifty milliseconds."
Fifty milliseconds, fifty milliseconds.
Meng Yao couldn't help but repeat it to herself twice.
If it's at this speed, plus the time for sampling and the electrical signal to reach the device, it might actually be possible to achieve 200 milliseconds.
"Then let's give it a try!" Meng Yao said.
Zhao Wenyuan nodded. "Once the first step of training is in place, we can proceed to this step."
After discussing the implementation plans for both phases, and with both sides having certain expectations for the work ahead, Zhao Wenyuan stood up and shook hands with Meng Yao and Zhan Xiang, saying, "I look forward to our friends from Mengnao joining us next week."
After meeting Han Luyi and Zhao Wenyuan again—especially Zhao Wenyuan — Meng Yao gained a new understanding of Origin's scientific research capabilities. Perhaps with their help, Mengnao Technology could truly break through its current bottleneck and advance its research to the next stage.
At this point, Meng Yao remained wary of the possibility that Origin might force a takeover of Mengnao Technology or affect its R&D and implementation strategies. However, she also realized that this investment would bring Mengnao Technology far more than just 40 million yuan in R&D and operating funds.
The research and development capabilities and product implementation capabilities that a large Artificial Intelligence company can bring may be far more powerful than Meng Yao initially imagined.
Just as Meng Yao was about to take her leave, Zhao Wenyuan suddenly said, "President Meng, I have one last question I'd like to ask."
"Please go ahead, President Zhao."
Zhao Wenyuan said, "Where is your data currently stored? If you can share the data securely, I'd like to start the first round of training as soon as possible. If we're lucky, we should have results by next Monday."
"Next Monday?" Meng Yao exclaimed in surprise.
Zhao Wenyuan explained, "What we were discussing earlier was that we'd run a model with 500 million parameters in the first round. If it's that many parameters, we'll run it once using Glutinous Rice Ball's training framework. We also have a Thousand-Card Cluster we can use flexibly. The results for the first round of pre-training should be available in two or three days, which means next Monday or Tuesday."
Meng Yao and Zhan Xiang exchanged a glance, neither of them expecting Zhao Wenyuan to be in such a hurry.
Moreover, the fact that Zhao Wenyuan could say "Thousand-Card Cluster" so easily gave them another glimpse into the power of Origin.
Zhao Wenyuan smiled and said, "It's Friday now, and we need to enjoy the weekend. I can't bear to let the computing cluster rest, so it's best to assign them tasks as soon as possible."
Actually, that was a joke on his part. With the Computing Power of Origin so strained right now, the computing cluster couldn't possibly rest. Even without work from Mengnao Technology, there were other tasks to complete. "Then let's make a trip there later and bring the data hard drive and backups over. We were going to move them over on Monday anyway," Meng Yao said.
Zhao Wenyuan looked at her with a smile and said, "Then I'll really trouble you President Meng." His tone was full of expectation for the data, without a trace of apology.
The plane landed in Haicheng around 6 p. m.
Just after turning off airplane mode, Han Luyi received a message from Zhao Wenyuan: "Boss Han, the first round of brainwave signal training has begun. I will report the results to you as soon as possible."
Han Luyi looked at the message and shook his head.
Wenyuan, you're really pushing yourself too far. No wonder Dr. Zhang forced you to live in his office.
No, I need to talk to him about this privately when I get back.
A balance between work and rest is essential for maintaining the team's combat effectiveness. Even if overtime is necessary, the overtime pay must be adequate.
As Han Luyi, Zhang Biao, and Liu Yu walked towards the arrival hall, Han Luyi said, "Biao, let's take a taxi. Liu Yu, you should go back and get some rest." Actually, Han Luyi hadn't planned to stay away for so long this time, but the opportunity with Mengnao Technology was rare, so he didn't return to Haicheng Han Luyi the investment was finalized and their launch meeting was held.
Liu Yu said, "Boss Han, it seems someone is answering your call."
Upon hearing Han Luyi's heart stirred, and he turned around to see Gu Siyue standing in the distance.
She wore a light-colored dress, black stockings, and high heels, making her stand out remarkably tall in the crowd. Combined with her delicate yet icy expression, she naturally exuded an aura of unapproachability. Many passersby couldn't help but steal glances, but no one dared to approach her. Han Luyi turned to look at Zhang Biao, then at Liu Yu, and said, "You can all go home now. See you Monday."
Then he quickly walked to Gu Siyue's side, ignoring the stares of those around him, and embraced the Lawyer sister.
Gu Siyue showed neither shyness nor resistance, and even gently wrapped her arms around Han Luyi's back.
They hugged quietly for a while before gently separating.
Han Luyi smiled and asked, "Why did you suddenly come to pick me up without telling me beforehand?"
He had previously told Gu Siyue his flight information, but the Lawyer sister had never mentioned coming to pick him up at the airport.
Gu Siyue glanced at him, her eyes filled with complex emotions, and said, "Have you eaten on the plane? I'm going to take you out for a social event."
Han Luyi eyebrow and asked, "Socializing? Who?"
"One of my clients."