The Heart of Machine Originality
Author: Egg Paste
"What is the most expensive thing in the 21st century? Talents!" Uncle Li's quotation 20 years ago does not sound outdated now.
![](https://nimg.ws.126.net/?url=http%3A%2F%2Fdingyue.ws.126.net%2F2024%2F0621%2Fe2e10b63j00sff5jg0011d000hs00f1m.jpg&thumbnail=660x2147483647&quality=80&type=jpg)
In the past two years, the big model track has been stormy. Technology companies are fully committed to the research and development of generative AI technology, looking forward to integrating new breakthroughs into their own products. OpenAI is a very successful sample, and has been widely concerned and deeply discussed by researchers, engineers and investors: Why OpenAI?
Its success comes not only from years of insistence on technology exploration and the capital continuously injected by investors, but also from the gathering of a group of top AI researchers. We can see that behind such epoch-making products as ChatGPT, Sora and GPT-4o, there is a long list of core contributors,
If a technology company wants to continue to be "great", technical talents are the key element.
Looking back at China, we can also see a group of technicians with ideals and enthusiasm: Someone worked hard for several months to create the first Chinese native DiT architecture open source model of cultural map with the team; Someone went through all the latest papers in order to reduce the delay experienced by users from 10 milliseconds to 1 millisecond; Some people have devoted themselves to the research and development of the big scientific model for many years, hoping to find the code to explain life with AI.
How do these people work? After approaching and listening to their stories, we found three common elements: I really love the track, the ultimate pursuit of cutting-edge technology, and persistent self drive.
Of course, talents also need to be adapted from the soil of encouraging innovation. One view is that technology companies should choose the person who best knows how to solve problems fairly, not just education. Another view is that we should not only focus on qualifications, but also dare to reuse new people, so that the team will always flow fresh blood. In a word, this is a problem worthy of serious consideration by any technology company.
Let the Chinese native big model lead the world
Seven years ago, when Gao Yan (not his real name) got his doctorate and embarked on the flight home, he could not imagine that in the next few years, the field of artificial intelligence would experience such a huge change.
We know that there are two ways to understand visual processing in the field of computer vision: discriminant and generative. These two methods led researchers to different paths. Before graduation from senior college, discriminant AI technology represented by "face recognition" had just experienced an explosion. From the other side of the ocean to China, the atmosphere of technology entrepreneurship is also hot, and many new applications have been born in such scenes as access control clocking, mobile phone unlocking, and smart home.
It was in this year that Gao Yan came to Tencent. This is a "two-way rush" story: the senior researchers who hold several top titles of the conference only invested in Tencent, and then successfully entered the company through Tencent's "technology giant" project.
Now, with only 7 years of work experience, Gaoyan has left its name in several key projects and papers.
But in this process, his research content is keeping pace with the times. Especially since 2022, the explosion of generative AI has almost overturned the previous research ideas, and the Wensheng map has become a new hotspot in the field of vision.
"With the continuous development of technology, many of the things learned in that year have been" eliminated ". Therefore, learning has become a part of the daily work of the advanced research institute. Every day when returning home from work, we should browse the latest technical research and track the latest papers.
As one of the earliest members of Tencent's hybrid model team, the research and development achievements of Gaoyan have been used by many people - in the past year, the cultural map ability of hybrid model has achieved "from scratch", and then more "accurate and beautiful".
Behind the strong ability of Wensheng Diagram is the months of hard work of Gaoyan and team members: the industry's first Chinese native DiT architecture Wensheng Diagram model, "hybrid DiT". Recently, Hunyuan DiT has been fully open source.
![](https://nimg.ws.126.net/?url=http%3A%2F%2Fdingyue.ws.126.net%2F2024%2F0621%2F35639c7cj00sff5jh001id000mx00dgm.jpg&thumbnail=660x2147483647&quality=80&type=jpg)
- Model address: https://huggingface.co/Tencent-Hunyuan/HunyuanDiT
- Technical report address: https://tencent.github.io/HunyuanDiT/asset/Hunyuan_DiT_Tech_Report_05140553.pdf
The visual generation effect of hybrid DiT is more than 20% higher than that of the previous generation hybrid Wenshengtu large model. It supports bilingual input and understanding in both Chinese and English, with 1.5 billion parameters. It can not only support the Wensheng map, but also serve as the basis for multimodal vision generation such as video.
After several iterations, the big model of hybrid Wenshengtu has solved the three key problems of "semantics, content, and texture", and has taken the lead in landing: Tencent has long begun to explore AI's automatic generation of advertising materials in advertising scenes, such as the product "Tencent advertising ideas" that generate product advertisements or advertising maps. In multiple rounds of evaluation under the advertising business, the case excellence rate and advertiser adoption rate of Tencent's hybrid Wenshengtu reached 86% and 26% respectively, both higher than similar models.
We often say that the behind the scenes team has made a product brilliant, but from another perspective, the success of the product has also made everyone in the team. For Gaoyan individuals, He likened his experience of participating in the Hunyuan Large Model Project to "life has opened an accelerator" - the technology in his hands can not only be fully applied and verified in a broader scene, but also expand new business cooperation, which makes people really feel that they are doing something that changes the world.
"After graduation, I teach AI to play games"
"Do your parents know that you play games at work?"
Hearing this question, Fu Zhiyuan (not his real name) smiled: "I only knew that I was doing research in the game department, but I didn't know that I also played games."
Compared with the rhythm of life in which most of the time during his doctoral period was immersed in papers and experiments, Fu Zhiyuan did "spend more time" on games after work.
Fu Zhiyuan's research direction is to strengthen learning and multi-agent systems. The environment of university laboratories is relatively pure. Most of the time, he focuses on breaking through a certain topic in the field. The game industry is a very ideal testing ground for the landing exploration of agent technology. Like the real world, games are also rich learning environments with responsive real-time settings and changing goals. Many researchers are committed to building more powerful AI agents, giving the whole system a stronger ability to plan or solve problems.
This is also what Fu Zhiyuan thought when he graduated: what new breakthroughs can the emerging AI technology bring to the game?
After receiving his doctorate from Nanyang Technological University, Fu Zhiyuan decided to return to China and join Tencent's IEG Photon Studio Group. Large factories like Tencent can provide the best conditions for computing resources and game environment resources used to train and optimize AI systems.
![](https://nimg.ws.126.net/?url=http%3A%2F%2Fdingyue.ws.126.net%2F2024%2F0621%2Fbbcb504fj00sff5jh0027d000u000ffm.jpg&thumbnail=660x2147483647&quality=80&type=jpg)
After coming to Tencent, Fu Zhiyuan From gamers to people who help develop games. "Roughly speaking, 60% of his time is spent dealing with the business of the game itself, and 40% is spent learning and applying new game technologies." Specifically, his task is to use AI to improve the game experience.
Generally speaking, AI will be set up a "reward and punishment mechanism" in advance, and then explore the playing mechanism independently and check whether there are bugs in the game function. Sometimes, it can also find some "highlights" that humans have not discovered: for example, which matching method is more conducive to winning, where to exploit loopholes, where to have unlimited resources. Ultimately, these findings will be used to improve the game experience.
Different from the laboratory, technology implementation needs more consideration of user experience. For example, the ability to join the AI model is beneficial, but it will increase the size of the game installation package. The team should consider how to make the model run efficiently under extreme conditions, or it will lead to high latency, high energy consumption, memory occupation, hot mobile phones and other problems.
"The 20 millisecond delay and the 1 millisecond delay are different in terms of user experience. The goal of AI is to make users more happy when playing games!"
In a place where excellent talents gather, there will be more or less "volume". In particular, research projects closely related to business will drive people to constantly think and create. They should not only be familiar with the game business, but also track the latest AI technology. Therefore, whenever an important new AI agent paper appears, the internal team will immediately sit down and seriously discuss it.
When the first model he participated in the training ran in the real business environment, Fu Zhiyuan's psychological activity was very strong: "There was a feeling of excitement and expectation when hearing the roar of the racing engine."
Do something of long-term value
Compared with Gao Yan and Fu Zhiyuan, Wang Aiwen (not his real name) has been "academic" in his research direction for many years, from doctor to Tencent AI Lab.
In recent years, the application of AI technology to traditional scientific fields, such as physics, chemistry, biology, and medicine, the so-called AI for Science, is a cross field with high hopes. Many complex problems that could not be calculated before can be well modeled, and effective predictions that can guide engineering practice in the real world can be obtained, It has promoted scientific discovery and technological innovation as never before.
Among them, proteomics is a very frontier subject. The scientific community once believed that the root cause of disease could be understood only by mapping the human genome sequence map, but this is not the case. The same gene often has different expression, which is the expression of different proteomes.
An interesting analogy is that the relationship between genome and proteome is like dictionary and article, element table and chemical plant. Therefore, to truly explain life, we must find the answer from the proteome.
This is what Wang Aiwen has been doing in recent years. This year, the three "proteome" studies she participated in have been published in the international top academic journals, respectively, and put forward highly forward-looking research programs to solve the problems of database, AI modeling and AI assisted clinical analysis in proteomics.
![](https://nimg.ws.126.net/?url=http%3A%2F%2Fdingyue.ws.126.net%2F2024%2F0621%2F6d1881adj00sff5jh001md000u000ntm.jpg&thumbnail=660x2147483647&quality=80&type=jpg)
For Wang Aiwen personally, The most difficult thing is that Tencent AI Lab provides a very pure research environment.
Compared with WYSIWYG business value, long-term value is often more difficult to see clearly. Today, it is difficult for us to determine whether several proteomic studies can leverage a large-scale market, but perhaps the next outbreak point of the biological industry will come soon.
Tencent 2025 Qingyun Plan Launches: 50% Expansion of Large Model Talents
If we take a long-term view, technical talents are not only the competition between technology companies, but also the competition of comprehensive strength at the national level.
The outbreak of the big model wave, to some extent, has increased the intensity of competition for top AI talents. This year, Tencent highlighted in the new Qingyun Plan: To strengthen the recruitment of talents in the field of big model The number of places will increase by 50% over last year. The scope of recruitment is for "top students in the world", which provides "highly competitive salary", as well as customized training programs, core business job opportunities, forward-looking technical topics and other attractive conditions.
In March this year, a "Global AI Talent Tracking Survey Report 2.0" tracked the global distribution and flow of many top AI researchers and scientists. The report found that China has cultivated a large proportion of the world's top AI researchers, which has increased from 29% in 2019 to 47% in 2022, and more and more AI talents have chosen to develop in China.
The growing industrial demand in China is the factor that attracts these AI talents. After a hundred model war last year, China has more than 200 large models and is still growing rapidly. In the era of big model, both technology giants and start-ups have broad development space.
It can be seen that China has taken the lead in the implementation of large model technology. In terms of making products, Chinese companies have more "scene" advantages than overseas companies, and the warming of the large model market will drive the upstream and downstream industrial chains, increase the demand for talents in data, computing power, algorithms and other fields, and provide a broad space for these talents to display their talents.
Tencent has always attached great importance to the recruitment and training of technical talents. Among the students who joined Tencent through the recruitment project of top technical talents in the past, there are dozens of students who have grown into core backbones and technical middle and senior executives of major business departments. They contributed their talents in various projects and found a clearer goal in life.
On June 19, Tencent announced the official launch of the new year's Qingyun plan 。
The recruitment scope of this year's Qingyun plan includes ten major technical fields, namely AI big model, infrastructure/hardware, financial technology, storage/database, robot, multimedia, game engine, security, quantum and big data. In particular, the talents who join Tencent's Qingyun Plan will be tutored by Tencent's chief scientist Zhang Zhengyou and outstanding scientists such as Yu Dong and Wu Shi.
The graduation time of the applicant must meet the following requirements:
- PhD students from January 2023 to December 2025
- Undergraduate students from January 2024 to December 2025
At the same time, candidates who apply for the Qingyun Plan must meet the following three requirements:
- Have a real technological ideal, technological enthusiasm, technological persistence, and are willing to use technological power to improve the quality of life of people around the world;
- During his student days, he made outstanding technical achievements, and had outstanding performance in academic, practical, competitive and other fields;
- With unique insight, it can penetrate the essence of technology and apply it to the ground, providing innovative and far-reaching answers to complex problems.
In addition, it is expected that from July, Tencent will organize a number of Green Cloud technology salons&technology open days to discuss the most cutting-edge topics in the technology field together with business leaders, academic celebrities in the scientific community, Tencent technology bulls, and so on, so as to help the development of technical talents.
For more information, please see: Tencent 2025 Green Cloud Plan Global Launch
Link:
https://mp.weixin.qq.com/s?__biz=MTkyNTM0MzA4MQ==&mid=2650942330&idx=1&sn=8d981ae72306a29dfe9ca7393e6b1350&scene=21#wechat_redirect