Reborn in 2010, the AI ​​in my mind is too powerful.

Chen Yiran's fingers landed on the keyboard, and following Jasmine's instructions, the code began to spread out.

Although both are applications of DeepFlow, he did not refer to the ideas of Yin Qi and the other two at all.

Rewrite everything, starting with the first import.

The network structure was extended to a deeper number of layers, with batch standardization applied after each convolutional layer, and the activation function was changed from sigmoid to ReLU.

The optimizer was replaced with Nesterov momentum, the initial learning rate was set to 0.001, and a cosine annealing scheduler was added afterward, which eliminated the need for manual adjustment of the decay curve.

Three indignant Tsinghua students initially stood by, pointing and whispering angrily.

A few minutes later, they started whispering among themselves, their faces filled with shock.

Chen Yiran's subsequent speed left the onlookers, the pride of heaven, completely stunned.

The code in the console ran without any pauses. Function definitions, convolutional layers, pooling layers, fully connected layers, training loops, early stopping checks, validation set metric callbacks—all were stacked up line by line by Chen Yiran, without a single character being deleted or rewritten.

Ten minutes later, the three of them started to lose their train of thought.

Twenty minutes later, Chen Yiran had finished writing the first part of the code. After saving it, he created a new CUDA kernel file.

He added a piece of logic for manually allocating shared memory, instead of letting the compiler automatically manage the movement of convolution kernels into shared memory, he wrote the tiling logic by hand in the code.

The input data is split into blocks, and weights are inserted according to their indices, bypassing the global memory bus bottleneck. There is no ready-made example for this logic in the CUDA programming manual; it relies entirely on understanding the latency levels of L1 cache and shared memory in the SM architecture.

After Lin Xiao returned, Chen Yiran immediately opened a bottle of iced cola and started chugging it down while looking at the three people who looked a bit disheveled.

These three have such good mental fortitude. An average undergraduate would have already prostrated themselves in worship. These three haven't even knelt down yet.

Chen Yiran calmly said: "After all, they are Tsinghua students, and from the Yao Class at that. They were personally taught by academicians; they've seen it all, they're not ordinary people. Let's continue, we'll make sure they feel like they're living in the shadows for the rest of their lives."

Putting down his cola, Chen Yiran continued typing. After a little over an hour, he embedded the DeepFlow acceleration framework into the training script and clicked the run button.

The first line of timestamps and training logs popped up in the console; the lab was very quiet.

The first epoch has finished, and a series of numbers have appeared on the console: training accuracy 92.7, validation accuracy 91, and the time taken for one epoch is 37 seconds.

Chen Yiran sighed and said, "Sigh, 37 seconds, that's still a bit long. This Dell Alienware M17x R2 uses an NVIDIA GTX 480M. It has more than a hundred more cores than your GTX 285, and its memory bandwidth is also a level higher. It's only an hour faster than you guys, what a failure."

The three Tsinghua students were no longer able to react to Chen Yiran's teasing. They stood to the side, their faces pale, looking as if their resolve had been shattered.

They knew there would be a gap, but they never imagined it would be this big. All three of them were prodigies from a young age, having fought their way through various Olympic competitions to get into the Yao Class.

They knew the power of Deepflow and admired Chen Yiran to some extent; this was a normal emotion for technical personnel towards a strong individual.

But to be honest, they don't think they're that bad. They feel they're not incapable of producing something similar by chance. This is the confidence that technical people have in themselves.

But in just over an hour, all the confidence they had built up over the past decade was shattered.

Yin Qi, as the leader of the small group, said in a hoarse voice, "I admit it, indeed, the three of us are not on the same level as you."

The other two opened their mouths, but ultimately said nothing.

Chen Yiran closed the notebook, handed it to Lin Xiao, and then began to make his suggestion.

"Don't compare yourself to me. Everyone in my company knows that you shouldn't use me as a normal comparison point; it will hurt your self-esteem. My conditions remain the same. I can provide you with funding. I'll put up one million US dollars, which is about seven million RMB."

Each of you owns 5% of the shares, which is equivalent to a Series A funding round, enough for initial development. What I've written today isn't exactly mainstream; it's just showing off, and there's not much room for further development.

I'll set a goal later and write out a basic framework. The three of you will be responsible for subsequent extensions and optimizations, or training the neural network for vision tasks. Of course, if you have anything else you'd like to do, feel free to tell me, and I'll consider it.

The three of them looked at each other for a few moments, then rather comically took a few steps back and formed a circle, whispering and discussing amongst themselves.

Finally, Yin Qi, who was brought along for the negotiations, said: "How about the three of us each take 10%? I know our technology isn't as good as yours, but we originally wanted to start our own business. Sigh, we're a bit discouraged now, but only getting 5% of the shares... we feel... a bit uncomfortable."

After all, they were three recent college graduates, all science-oriented guys majoring in technical fields, and their eyes held a clear-eyed naiveté. Their negotiation tactics almost made Chen Yiran burst out laughing.

He thought for a moment and came up with a new solution.

"I can invest a million US dollars directly, and five percent of that would mean each of you would increase your net worth by fifty thousand US dollars for no reason. It's too much for you to ask for ten percent."

How about this? The current cooperation method is too troublesome. Why don't you all come and work at my company? My Shougang Wanguoyuan will set up a special project team for the three of you, and you three can be the project team leaders.

Then we'll conduct research on visual training, which is also used in the chip EDA industry. This work experience is something you'll need.

We'll sign a five-year contract first. If you want to start your own business after five years, I can let you go. And you'll have to work hard at Wanguoyuan, although I can't share some of the patent rights with you.

But when you leave to start your own business five years later, I can invest in you, from angel rounds to Series C.

After further discussion, the three of them agreed on an exceptionally high salary package offered by Chen Yiran: 30,000 yuan per person per month, plus a 16-month bonus at the end of the year. They immediately signed contracts with Shougang Wanguoyuan, becoming the initial members of the Wanguoyuan Longan project team.

Chen Yiran watched with great satisfaction as the three signed the papers, then began to urge them to recruit more people: "You're from Yao Class, so you can poach anyone talented from your class to join the project team. You three alone won't do. You should create a list of personnel needs and have Lin Xiao help you recruit."

I'll provide you with a basic framework as soon as possible; just follow my steps. If you have any technical questions, there's a dedicated Q&A page in the company system where you can leave a message for me.

Chen Yiran left the three people with Lin Xiao, telling them to go through the onboarding process, and went home.

He needs to start preparing; he's going back to the United States next week.

Catherine has already given him the schedule, and he expects to be living in the headlines for the entire month of March.