Home >Technology peripherals >AI >AI technology accelerates iteration: large model strategy from Zhou Hongyi's perspective

AI technology accelerates iteration: large model strategy from Zhou Hongyi's perspective

王林
王林forward
2023-06-15 14:25:501209browse

Starting this year, no matter where in public, 360 Group founder Zhou Hongyi has been talking about one topic, and that is large-scale artificial intelligence models. He calls himself the "preacher of GPT" and marvels at the progress made by ChatGPT, and firmly believes that this progress will promote the iteration of AI technology.

As a celebrity entrepreneur who is good at expressing himself, Zhou Hongyi’s speeches are often full of witty remarks, so his “sermons” have also created many hot topics and indeed added fire to large AI models.

Although Zhou Hongyi has become an opinion leader, people are more concerned about how the 360 ​​company he leads responds to this wave of artificial intelligence revolution. In fact, within 360, Zhou Hongyi has already initiated a change for all employees. In April, he issued an internal letter requiring every employee, product and business of 360 to fully embrace artificial intelligence, adapt to human-machine collaboration, and start product development. Reinvention.

If the timeline is pushed forward, in February, 360 responded to the layout of GPT-like technology for the first time, saying that the company’s Artificial Intelligence Research Institute has been making continuous progress in AIGC technology including ChatGPT-like technology since 2020. However, as of now, it is only used as a productivity tool for internal business use. 360 will launch a demo version of ChatGPT-like technology as soon as possible.

At the end of the 2023 Digital Security Development and Summit Forum, Zhou Hongyi demonstrated the 360 ​​version of ChatGPT products to the attendees. At that time, the name of the product had not yet been determined. Zhou Hongyi joked: "Today we are going to take out the newborn baby and meet everyone."

Less than three months later, on June 13, 360 officially released “360 Intelligent Brain 4.0” and announced that 360 Intelligent Brain will be fully integrated into the “360 Family Bucket”.

Zhou Hongyi said that the past three months are like the past thirty years in the history of artificial intelligence, with various innovations emerging one after another, and the version of 360 Intelligent Brain has also undergone many iterations. He believes that the current development speed of large-scale models in China is very rapid. Various companies compete with each other and promote each other, so the gap with the world's leading large-scale models is constantly narrowing.

"Originally I said the gap was one and a half or two years, but I take it back today. I think the speed of development and iteration among domestic peers has basically caught up with or is close to the international advanced level," Zhou Hongyi said.

Big model strategy emerges

Looking at it now, the ChatGPT-like product released by 360 at the end of March was 360 Intelligent Brain 1.0. At that time, its capabilities were only combined with search and did not support multiple rounds of conversations; at the end of April, 360 Intelligent Brain iterated to 2.0 version, multiple rounds of dialogues were launched online; in mid-May, 360 released the multi-modal visual application "360 Hongtu", and 360 Intelligent Brain also iterated to version 2.5; at the end of May, 360 released a large visual model, and 360 Intelligent Brain also updated to Version 3.0.

The most important change in the recently launched 360 Intelligence Brain 4.0 is in its multi-modal capabilities, compared with the previous version. At present, 360 has cross-modal generation capabilities, including text processing capabilities, image processing capabilities, voice processing capabilities and video processing capabilities, which can realize text-based text, text-based pictures, text-based tables, pictures-based pictures, picture-based text, video understanding, etc. Function.

At the conference, 360 Intelligent Brain’s “Wensheng Video” multi-modal function was launched for the first time in China. Any text script can generate a video without being limited by professional skills and materials, demonstrating the ability to “make something out of nothing”.

Zhou Hongyi said that China’s self-developed large models have now caught up with GPT3.5 in many aspects, and the speed of catching up with GPT4.0 is beyond imagination. Therefore, the key to overtaking in corners is to develop more Modal capabilities.

“A complete large AI model must have four major capabilities, including text processing, image processing, speech processing and video processing,” Zhou Hongyi said. “The competition for large models in the future will be a competition of multi-modal capabilities. Only those with Only with multi-modal capabilities can large models truly reach a new level."

"360 Group's large-scale strategy has officially emerged with the release of 360 Brain 4.0, which can be summarized as 'two wings flying together and four paths running concurrently'.". Among them, "flying together on both sides" refers to technology and scenarios, and "four-way concurrency" represents four application directions, namely consumers, small and medium-sized enterprises, government and large enterprises and industries.

AI technology accelerates iteration: large model strategy from Zhou Hongyis perspective

According to Zhou Hongyi, in terms of technology, due to the needs of search business, 360 established an AI research institute as early as 2015 and accumulated core technologies for large models such as natural language understanding. In addition, because of the Internet of Things business, 360 has also begun to explore multi-modality in 2018.

"In 2018, Google released the Bert algorithm, and 360 tracked and used it at that time, including later T5 and GPT, and 360 also conducted research," Zhou Hongyi said, "As a search engine manufacturer, we do not wait to develop large models. We only followed up this time when GPT became popular. In fact, it started as early as 2018, which is why we were able to release related products so quickly."

Currently, 360 has ten core capabilities in large model technology: generation and creation, multi-round dialogue, coding capabilities, text classification, text rewriting, reading comprehension, logic and reasoning, knowledge question and answer, multi-modality, and translation. , the dimension covers hundreds of subdivided functions, can cover all application scenarios of large models, and ranks first in domestic large models in multiple third-party evaluations.

AI technology accelerates iteration: large model strategy from Zhou Hongyis perspective

In May of this year, in an evaluation result of the Chinese General Large Model Benchmark (SuperCLUE), 360 Intelligent Brain ranked first among domestic large models in multiple capabilities; on June 12, 360 Intelligent Brain became the first domestic model to pass A large model product evaluated by China Academy of Information and Communications Technology’s “Trusted AIGC Large Language Model Basic Capabilities”.

Accelerate application implementation

Regarding the future development of large models, Zhou Hongyi pointed out that just having basic capabilities may not bring about a revolution in industry and products. The key lies in how to combine the general AI capabilities of large models with more scenarios. The value of large models lies in the generalization, industrialization and verticalization of AI capabilities, and their deep integration with the specific business needs of different industries and enterprises.

"I think such industry opportunities have just begun. There will be a dividend period of at least 10 years in the future, including all apps, websites, and even all industries that are worthy of being reshaped with general artificial intelligence," Zhou Hongyi said.

Two months ago, Zhou Hongyi said bluntly that "the in-depth integration of GPT and hundreds of industries will directly change the industry competition pattern, and companies that cannot catch this train will be eliminated." Therefore, he also asked all 360 internal Product managers and engineers must use their imaginations to imagine how to reshape their products in the era of artificial intelligence.

At the press conference, Zhou Hongyi also officially announced that "360 Intelligent Brain 4.0" will be connected to 360 full-end products such as browsers, search, and security guards to achieve comprehensive intelligent reshaping. "Microsoft recently launched an operating system called CoPilot, which is equivalent to the co-pilot of Windows. It integrates artificial intelligence capabilities into every computer. In China, 360 may be the manufacturer closest to Microsoft's capabilities."

In addition, 360AI digital human was also officially unveiled at the press conference. According to Zhou Hongyi, 360’s main innovation in large-scale AI models is digital humans. Digital humans will become the most important application entrance for large-scale artificial intelligence models in the future. 360 will rely on the capabilities of 360’s intelligent brain model to redefine “digital human”. Different from traditional digital humans who can only output content according to established scripts, 360AI digital humans are characterized by having memories, personalities and personalities, and being able to reproduce their way of thinking and life experiences.

According to our understanding, 360AI digital people are currently divided into two types: digital celebrities and digital employees, with a total of more than 200 roles available. Digital celebrities include historical figures, idol stars, literary images, etc., allowing users to have close contact with idols in open conversations with digital people; digital employees can become knowledge assistants and office assistants for corporate employees, improving corporate operations. efficacy.

In addition, 360AI digital people support customization. Everyone can generate their own exclusive digital people at low cost by uploading a private knowledge base, such as digital clones, digital assistants, digital idols, etc.

Zhou Hongyi said that the future development direction of 360AI digital humans is to generate sounds, videos, have long-term memory, and even have the ability to decompose and plan goals, and have "hands and feet" and the ability to execute to better provide services to humans. . "The evolution of AI must be human-centered, and large models should become humans' friends and assistants."

In the past six months, many entrepreneurs have been excited by the development of large AI models and invested their efforts in new entrepreneurial fields. Zhou Hongyi, who has passed his destiny, is also one of them. He has repeatedly emphasized that "we are experiencing a revolutionary moment in the history of technology and business" and is happy to study large models.

In mid-May, after an iteration of 360 Intelligent Brain, Zhou Hongyi distributed hundreds of test codes online in a WeChat group at around 11 o'clock in the night, and discussed with people how to ask questions with relish. This was not the case in the past. Common, but in the era of super artificial intelligence, everything is becoming possible.

The above is the detailed content of AI technology accelerates iteration: large model strategy from Zhou Hongyi's perspective. For more information, please follow other related articles on the PHP Chinese website!

Statement:
This article is reproduced at:sohu.com. If there is any infringement, please contact admin@php.cn delete