Chapter 658 [Fang Hong's Suggestion]

The meeting lasted for about 40 minutes, and Chen Yu assigned tasks to the major development teams.

There are also many project teams in the company's technical department, involving natural language processing, recommendation systems, image generation, deep reinforcement learning, computer vision, and robotics.

Quantitative Capital, the outside world regards the company as an investment management company, but it is actually a high-tech company disguised as an investment management company.

As the meeting ended, everyone left the meeting room, leaving only Fang Hong and Chen Yu.

"I don't think there is any problem with the development route you set." Fang Hong said with a smile: "But it can't be slowed down. I have reliable news. Google's artificial intelligence laboratory GCP proposed in an internal meeting. An idea, to enable machines to 'learn' language and grammar, is also working on language models like you."

Chen Yu was surprised: "Google is also doing it? What level of technology has it reached?"

As the real behind-the-scenes boss of Qunxing Capital, Chen Yu believed that he should have obtained reliable information when he said so.

Fang Hong said calmly: "Not as advanced as you, I can be sure of that."

According to the memory of previous lives, around October this year, Google should release a GPT-based language model, which learns language by encoding the input, mainly a pre-trained language model, but Google has not developed it before. This language model.

Hearing what Fang Hong said, Chen Yu immediately felt relieved. In this field, one step ahead is often one step ahead, and the development speed of AI often develops rapidly at an exponential rate.

After a while, Chen Yu looked at Fang Hong and said, "Our AI large model, the first generation of iterative research and development task is to realize the ability to generate human text, supported by a scale of tens of billions of parameters, and has now entered the second generation of research and development. Larger and stronger than the first generation, it can generate coherent and attractive responses to various cues, with a parameter scale of the order of 50 billion."

"As for the third-generation model, it should need 200 billion parameter support. The third-generation goal is to be larger, more powerful, and more advanced, and to achieve amazing performance on multiple natural language processing tasks."

"For the upgrade of AI large models, it needs the support of strong underlying capabilities to play well. To break through the bottleneck of AI through large computing power and large-scale training data, the operating cost is also very large. Now the operating cost of one month exceeds The scale of 300 million yuan."

Hearing this, Fang Hong said succinctly: "Money is not a problem, I will take care of enough."

Chen Yu said with a smile: "It's really not enough, I will ask you for it."

Its underlying capabilities include computing power resources, data storage, power supply, network bandwidth, etc.

These are very expensive things.

AI training requires the use of large-scale computing resources, including GPUs, GPUs, etc. Cloud computing can provide powerful and elastically scalable computing resources.

The underlying data center of cloud computing can provide high-performance computers and hardware devices, and can be optimized to improve the utilization of computing power resources, thereby accelerating the training and generation speed of AI, and can also provide powerful parallel computing capabilities. Computing tasks are assigned to different computing nodes for computing at the same time, improving training efficiency and speed.

In addition, training AI requires network requirements. One is to achieve perfect link load balancing, and the other is to quickly recover when the network fails.

Because of the "barrel effect", as long as one link has uneven load and causes network congestion, which becomes a shortcoming of the barrel, then even if the other two links are unimpeded, the collective communication time will still increase significantly, thereby affecting training efficiency.

It is similar to the one-core and multi-core onlookers.

As the scale of the cluster increases, link failures are inevitable. A similar link failure will cause the entire communication to stagnate, which in turn will cause the model training to stagnate. Therefore, it is necessary to quickly repair the fault.

Fang Hong crossed Erlang's legs and leaned back on the chair, leaning on the armrest of the chair with one hand, and said thoughtfully: "Your AI is still missing one element."

Hearing this, Chen Yu asked curiously, "What's the difference?"

Fang Hong said succinctly: "Image."

Chen Yu was slightly taken aback: "Image?"

Fang Hong immediately said: "How to make the product accepted by the public more quickly? How to form user stickiness, especially to capture the younger generation. This AI must be visible, not a series of 0s and 1s, so AI should be given A character image, that is, anthropomorphic, must capture the younger generation, especially the post-90s and post-00s generation, and furthermore, AI should be given a two-dimensional character image."

Chen Yu couldn't help but nodded when he heard this: "It makes sense."

After a while, he looked at Fang Hong and said, "I know how to do it. If we want to do it, we should make it big. Later, I will set up an additional department to create a virtual image brand operation department for AI and create an exclusive IP image. It enriches its characters and has a complete story background."

"Including hiring some professional animation production teams and screenwriters to produce some excellent story fragments to enrich the image of its characters, and it can be called the first aborigine in the 'virtual world' in publicity, etc."

Fang Hongru said: "It is not easy to create an IP from scratch until it is accepted by the public and deeply rooted in the hearts of the people. It takes effort and cost, and it also takes time. The sooner you do it, the better."

Chen Yu nodded, and after Fang Hong's little advice, he immediately understood the importance of AI visualization and anthropomorphism.

And it is very promising. After the launch of APP, it can become a virtual assistant that users must install, including the Internet of Everything in the smart home. In the future, various home appliances and smart devices can be hosted by the virtual AI assistant.

Now that this IP image is established and accompanied by the growth of the younger generation, there is a lot of room for imagination in the field of smart home appliances alone. For air conditioners, washing machines, computers, mobile phones and other equipment, these manufacturers are unable to engage in AI. It is only necessary to access the AI ​​assistant of quantitative capital.

In the future, if it cannot support access to this virtual AI assistant, users will not even want to use products such as washing machines or air conditioners of this brand.

After all, there are many brands of washing machines and air conditioners, but there is only one such powerful AI assistant, and the IP virtual image created has grown with the growth of a generation, and has formed user habits. As long as there are other air conditioners and washing machines that can support this AI , users will have a great tendency.

At that time, those manufacturers will not be able to connect, and it is impossible for quantitative capital to make air conditioners and washing machines by themselves. The construction of AI and ecosystems only needs to make their products well and refined.

...

Fang Hong stayed at Quantitative Capital for an hour and a half before leaving. These tracks were handed over to Chen Yu, and he had great confidence in him.

Back at Jingxinju Villa, Fang Hong went to the study, and the fame system data panel appeared in front of the screen, and then directly consumed 100,000 achievement points to exchange for an item card to check the character attributes of Chen Yu.

[Target: Chen Yu]

[Character: Excellent]

[Reputation: 85105209]

[Health: Excellent]

[IQ: 145, genius]

[Coordination: Excellent]

【Wealth: RMB 17.6 billion】

...

"These hexagonal data are not bad." Fang Hong was very satisfied when he saw the data of the six attributes, especially the IQ value reached 145. As long as it reaches a value above 140, it is considered a genius.

The most important thing is that Chen Yu's data are all initial data, and have not been strengthened by Fang Hong.

Especially the IQ level, which is an excellent level between 120 and 139, and the level of most senior talents is around 130 to 139. Those who have been promoted to the leading talent level by Fang Hong have an IQ level of over 140 .

Chen Yu's initial data level surpassed those promoted leading talents, which shows that it is very rare.

In the end, Fang Hong decided to raise Chen Yu's IQ level to the 150 level, and the further he raised it by 1 unit, the greater the consumption of corresponding prop cards would be.

Chen Yu's initial data was so good, which indirectly saved Fang Hong a lot of achievement points.

After a while, Fang Hong exchanged a large number of R\u0026D prop cards, all of which were designated to improve Chen Yu's R\u0026D ability, and the data showed the improvement of his IQ level.

After raising Chen Yu's IQ level to 150 before stopping, Fang Hong glanced at the achievement points consumed and couldn't help muttering: "It actually consumed 12.6 million achievement points..."

It takes 100,000 achievement points to redeem an item card. To increase Chen Yu's IQ level from 145 to 150, it increased by 5 units, but it consumed 126 item cards.

However, Fang Hong doesn't feel distressed. The greater Chen Yu's fame will be in the future, his prestige and social influence will be greatly improved, and Fang Hong will also be able to accumulate fame and achievement points indirectly. The more than 12 million achievement points consumed by him It will definitely pay back, and even make a lot of money.

...