The realism of Tencent's large model: solving the "AI anxiety" of enterprises in the scene

Original source: IT Times

Author: Hao Junhui

Image source: Generated by Unbounded AI‌

This is an interview in which the question can only be "grabbed" by speeding up the speech rate and increasing the decibel.

On the afternoon of July 7, before the 2023 World Artificial Intelligence Conference Tencent Forum, in a small and noisy conference room, Wu Yunsheng, vice president of Tencent Cloud, head of Tencent Cloud Intelligence, and head of Youtu Lab, accepted a group interview from the media. Nearly 20 days ago, Tencent officially announced the MaaS panorama, cutting into the hot "large-scale model track" with large-scale industry models. A path that looks more "realistic".

"What enterprises need is to truly solve a certain problem in an actual scenario, rather than solving 70%-80% of the problem in 100 scenarios." Wu Yunsheng said that from the perspective of the company's strategy, Tencent is more focused on solving the actual problem of landing problems, and the general large model cannot completely solve all the problems of users.

Tencent, which has the largest number of individual users in China, is taking the first step in changing the industrial Internet to AI when a new wave of artificial intelligence comes.

Tencent Cloud MaaS upgrade

On June 19th, Tencent Cloud announced for the first time the research and development progress of Tencent Cloud's large-scale industry models. Provided more than 50 large-scale industry solutions for more than 10 industries such as media, cultural tourism, government affairs, and finance.

At the World Artificial Intelligence Conference, Tencent Cloud once again announced a number of upgrades.

Among them, the newly upgraded Tencent Cloud self-developed Xingmai high-performance computing network can increase GPU utilization by 40%, save 30% to 60% of model training costs, and bring a 10-fold improvement in communication performance for large AI models. Based on Tencent Cloud's new-generation computing power cluster HCC, it can support a super-large computing scale of 100,000 cards. Tencent Cloud's AI native vector database supports up to 1 billion-level vector retrieval scale, and the delay is controlled at the millisecond level, which is 10 times higher than the traditional stand-alone plug-in database retrieval scale, and has a peak capability of millions of queries per second (QPS).

In terms of application innovation, Tencent Cloud's large-scale industry model capabilities have been applied to scenarios such as financial risk control, interactive translation, and digital smart customer service, which has greatly improved the efficiency of intelligent applications.

The financial risk control solution supported by the industry's large-scale model has 10 times the efficiency compared with the previous one. Through Tencent's accumulation of more than 20 years of black and gray production confrontation experience and thousands of real business scenarios, the overall anti-fraud effect is 20 times higher than the traditional model. An increase of about %. In the field of digital humans, Tencent Cloud launched a small-sample digital human factory this year, which can reproduce 2D digital clones within 24 hours with only a small amount of data, greatly reducing the cost of enterprise application digital human services.

"In fact, for more than half a year, we have been thinking and exploring what is the most essential logic behind the combination of large models and various industries? There are actually only two points: one is that the fundamental starting point of technology is to solve practical problems, and the other is If you can’t go deep into the industry, you can’t really solve the problems facing the industry.” The “test” brought by the real scene to the big model made Wu Yunsheng feel a lot.

Intelligent customer service is recognized as the most applicable industry for LLM (Large-Scale Language Model). At this conference, Tencent created a large-scale industry model for an online travel OTA company. The fine-tuned customer-specific model can solve business problems end-to-end without configuring dialogue processes. Improve the task completion rate and reduce the cost of dialogue construction. But in fact, it is not as simple as imagined for the big model to truly understand the customer's problems.

"During the communication process, the customer's thinking is jumping and changing. For example, he just proposed to book the hotel on the 10th, but before the machine answered, he suddenly said, help me find the hotel and flight on the 11th , when the AI is still giving feedback on the second requirement, he may say, show me the twin room.” Wu Yunsheng pointed out that it is still quite difficult for the large model to realize multi-intent recognition, and the general large model cannot It is a simple solution, but needs to be combined with the scene, especially the interaction with the customer's system to reconstruct some very complex models.

The era of "group models dancing together" is here

After the initial hustle and bustle, how to commercialize AI large models, how enterprise customers can enjoy this round of AI dividends, and solve "AI anxiety" have become hot topics at this World Artificial Intelligence Conference.

Zheng Qingsheng, partner of Sequoia Capital China, has entered the investment field since the mid-term of the PC Internet. In his opinion, the winners of each era are derived from the original technology of that era. For example, in the PC Internet era, people value e-commerce and social networking Software has become the biggest winner; since the era of mobile Internet, people have paid attention to social software and long videos, but short videos occupy the most time. "Now we don't know which original scenes generated by AI itself will change our life. basic behavior."

Although it is still unknown when AI's native "killer" will appear, "entering the game" must be the first step. Among the more than 30 large-scale models unveiled at WAIC this time, except for the first round of general-purpose large-scale models such as Baidu Wenxin Yiyan, Ali Tongyi, Xunfei Xinghuo, and Shangtang Ririxin, the latecomers basically focus on the industry large model.

"For customers, enterprise-specific large models with few parameters, low investment, and quick results are more likely to be accepted, and their willingness to pay is relatively clear." An exhibitor of a start-up company told the "IT Times" reporter that some are already using large models. Bank customers who transform the customer service system usually choose a private domain deployment method that integrates software and hardware, and use their existing knowledge graphs and data to train and implement reasoning, which not only ensures data security, but also reduces the cost of computing power, " If only one scene needs to be inferred and output, the computing power board can even be done in single digits.”

"Industrial scenarios have become the best training ground," Tang Daosheng, Senior Executive Vice President of Tencent Group and CEO of the Cloud and Smart Industry Business Group, said at the WAIC plenary meeting-Industrial Development Forum, choosing a one-stop industry model The cooperation of cloud vendors with service capabilities to build their own exclusive models based on large industry models may be a feasible path for enterprises to explore the application practice of large models.

This means that the future will be an era of "coexistence of all models", and every enterprise will have its own large model, and Tencent has decided to be the enabler of the new era.

In the MaaS service panorama released by Tencent Cloud last month, it was pointed out that based on the Tencent Cloud TI platform, a large-scale industry model selected store can be built. Tencent Cloud can provide 10 major industries such as finance, cultural tourism, government affairs, medical care, media, and education. solution. At the same time, Tencent Cloud launched an industry large model fine-tuning solution to help model developers and algorithm engineers solve tasks such as model invocation, data and label management, model fine-tuning, evaluation testing and deployment in one stop, and reduce the pressure of creating large models.

On the basis of these models and tool platforms, enterprises can quickly generate their own "exclusive models" only by adding their own scene data.

"It is still in the early stage of the development of large-scale models. I personally hope that a hundred flowers will bloom and everyone will try different possibilities in different fields." Wu Yunsheng believes that the development of artificial intelligence is a huge data project, which requires common knowledge and It also requires a professional, profound and authoritative knowledge organization, and the joint efforts of all parties are needed to truly enable technology to serve the industry.

AI for Science Captures Cosmic "Flicker"

Of course, in addition to exerting effects in the digital transformation of industries, Tencent Cloud's large-scale industry model also accelerates the application of AI technologies such as large models in the field of scientific computing.

Beginning in 2021, Tencent, the National Astronomical Observatory, and the School of Computer Science and Technology of Fudan University jointly launched the "Star Exploration Project", using cloud + AI to help China Tianyan FAST process the huge amount of data received every day, and find fast radio bursts and pulses through visual AI analysis. According to star clues, 30 pulsars have been discovered so far.

At this year's WAIC, Tencent announced that the star exploration program has made further progress, and for the first time discovered 2 fast radio bursts through AI technology.

Fast radio bursts are a mysterious astronomical phenomenon. Every 1 millisecond, the energy released by the sun throughout the year will be emitted, "flickering" the universe. However, its "flashing" frequency is extremely low and the time is extremely short. It is easy to ignore in the massive data and extremely difficult to capture. It was not until 2007 that humans discovered the first pulsar, 40 years later than the discovery of pulsars.

Compared with pulsar exploration, in order to discover fast radio bursts that occur at a lower frequency in massive data, AI models are required to have higher accuracy and faster calculation speed. In order to improve the calculation speed, Tencent specially designed a set of brand-new, end-to-end AI algorithms for the exploration of fast radio bursts. Under the same computing power, this brand-new astronomical data processing paradigm promotes signal processing efficiency to be 1800 times faster than conventional processing.

Previously, before AI could understand the map, it was necessary to complete complicated astrophysical preprocessing on the signal map, such as Fourier transform, chromatic dispersion... These tasks are professional and complicated. Now Tencent Youtu has created an "end-to-end AI algorithm" for astronomical data processing, which can skip the preprocessing steps and directly enter AI recognition, greatly improving efficiency.

FAST generates hundreds of terabytes of data every day and tens of millions of signal maps every week. In the face of massive data, Tencent Cloud can quickly locate and identify useful information in the data through the "multi-instance learning method + attention mechanism", and provide powerful underlying computing power support.

Today, Tencent Cloud and FAST are continuing to detect the radio signals of M31 Andromeda 2.5 million light-years away, and it is expected that more "cosmic flashes" will be captured in the near future.