Author|Liu Yuqi, Hao Xin, Editor|Wang Yisu
The emergence of ChatGPT has made the technology bigwigs who have been "silent" once again enthusiastic.
Bill Gates firmly believes that "GPT is the most revolutionary opportunity in 40 years"; Huang Jenxun sang "ChatGPT is the iPhone moment of artificial intelligence" all the way, arching NVIDIA's market value to trillions of dollars; Li Yanhong can't hide his excitement, believing that "big models have changed artificial intelligence and will also become the fourth industrial revolution"; Lu Qi repeatedly reiterated that "the inflection point of the new paradigm has arrived, and the big model will be everywhere".
Zhou Hongyi is also one of the biggest guys who have reacted the fastest to this scientific and technological revolution. He likened the big model to a technological revolution in the general technology of "power generation", predicting that "every industry, enterprise and individual will have customized GPT", "the combination of multimodal large model and the Internet of Things is the next outlet", "GPT still has to develop, and not developing is the biggest insecurity".
Compared with laying the foundation from scratch when laying out AI in 2016, this time, technology companies have accelerated the practice of AI because they have a business foundation.
Implementing it as soon as possible is the common mood of practitioners in this scientific and technological revolution.
In just two months, 360's movements are also quite frequent. On March 29, Zhou Hongyi held a press conference and decided to show the 360 large model before the conference, "Take the newborn baby out and take a look." In April, Zhou Hongyi sent an internal letter asking 360 to fully embrace AI, and also left homework for all product managers to think about how their products can be reshaped with large models.
In May, with the blessing of the 360 model "360 Wisdom Brain", Internet search, browser and security guards handed over their homework one after another. On May 31, 360 smart life also handed over the homework, Zhou Hongyi proposed that "big model + Internet of Things is the next new outlet", and believes that the visual big model will comprehensively reshape 360 intelligent hardware.
"The scene is mature and one is open," Zhou Hongyi said.
The big model came, allowing the entire scientific and technological community to find the next direction, and also brought Zhou Hongyi and 360 Second Spring.
01 Practicing the big model, Zhou Hongyi fancy the Internet of Things
"In the past, we talked about Internet +, and in the future it may be artificial intelligence +." Zhou Hongyi said that smart hardware is also a hardware-based APP, and I often think, how can it be combined with a large language model? What role will smart hardware play in this industrial revolution?
When the industry discussed the transition from large model parameters and emergence phenomena to industry large models, Zhou Hongyi set his sights on the application scenarios of large models in the Internet of Things, "The Big Boss Moss in "The Wandering Earth 2" may be the ultimate version of smart cameras."
Zhou Hongyi believes that the premise of connecting large models and the Internet of Things is to evolve from a language large model to a multimodal large model, and find a scene to combine the multimodal large model with intelligent hardware.
For the multimodal large model, Zhou Hongyi has his own understanding, he believes that the multimodal in the perception layer is not really multimodal, and the multimodal of the cognitive layer is the real artificial intelligence.
To take a simple example, the multimodal at the perceptual level can recognize "there is a cat and a dog in the picture", while the multimodality at the cognitive level can recognize "there is a cat and a dog fighting in the picture". The change between the two lies in the ability of the universal big language model to recognize and understand the world.
360's visual big model is based on the hundreds of billions of parameters of 360 intelligent brain, creating a general text-based natural language processing cognitive model; At the same time, the advance layout of 360 on the intelligent hardware side has accumulated billions of security data in the vast consumer market, creating a 360 visual model.
"On the road of big model development, it is not to simply put the large model on intelligent hardware, but to combine the ability of multimodal large model with intelligent hardware - from "seeing" to "understanding", understanding the relationship between characters and logic in the diagram, being able to read some terriers, and having human cognition. Zhou Hongyi explained.
The AI camera with the "brain" can transform the function of post-recording into real-time judgment. For example, the old man fell in the living room, the child climbed high risk, the original camera can recognize the old man and the child, but does not understand the semantics of the picture. But when the multimodal capabilities of the big language model support a cognitive brain, the camera will naturally associate the old man who may be in danger, and the child who may fall out the window, so that it can give early warning.
The combination of multimodal large models and the Internet of Things, Zhou Hongyi believes that this is the next outlet. "Big model + Internet of Things, from the perspective of empowering the Internet of Things, realizes from perception AI to cognitive AI, so as to better understand human instructions and execute human instructions."
For the Internet of Things, the addition of large models will also bring new increments in two aspects.
First, data increment: the Internet of Things collects more dimensions of data, different from Internet data, is a new data increment in the era of large models, especially after combining voice and visual solutions with hardware, the massive data generated by the terminal will usher in the second data explosion.
The second is scene increment: large models are finding landing scenes, and the combination of massive hardware is undoubtedly a huge scene increment.
"I have a prophecy because the valuable information produced by humans is limited, and many people worry that human texts are not enough. But I think as long as there is a multimodal capability, connect many cameras around the world to it, and collect all kinds of data in real time, even if Lao Zhou is talking nonsense here, the audience is whispering, and the big model can identify and obtain information through the camera. ”
Today, 360 has found the first scene combined with AI hardware - AI smart camera, which is the most basic and versatile scene. In the future, it may migrate to all smart terminals.
At the same time, if the data collected by the smart terminal wants to be better used by the large model, it will also force the manufacturers on the application side to solve the problem of data dispersion in the Internet of Things 1.0 era.
Just as in 2016, when Google beat Lee Sedol with AlphaGo, Amazon immediately entered the Internet of Things with the smart speaker Echo. The advent of ChatGPT will definitely drive the physical world further towards the direction of "intelligent connection of everything".
02 Batch landing, "one open one when the scene is mature"
ChatGPT has allowed the entire tech world to find its next direction.
When he first met ChatGPT, Zhou Hongyi was pleasantly surprised, and the mysterious attraction exuded by the large model was like the computer calling him when he was young.
Zhou Hongyi first showed his strong interest in ChatGPT to the outside world, starting with a conversation with his old friend Zhang Chaoyang in early February this year. At that time, ChatGPT was bursting out of the circle, sweeping the world, and the two bigwigs exchanged entrepreneurial experiences, and ChatGPT became an inescapable topic.
Zhou Hongyi, who has been paying attention to large model technology since 2020, showed his judgment: "ChatGPT may represent 'the beginning of a real revolution in the history of artificial intelligence', although it is not perfect and has many shortcomings, but there is unlimited potential and unlimited application scenarios in the future." ”
Although most companies at home and abroad are still waiting, in the face of the upcoming wave of large-model entrepreneurship, Zhou Hongyi has a preliminary understanding of the market, "It depends on whether it has invested in server computing power in the past ten years, whether it has a team that does artificial intelligence, and whether it has big data storage." It's not always a bad thing to have a bubble, there will be more good people, there will be more money pouring in. ”
After deciding on the direction of the large model, Zhou Hongyi thought about how to combine the general capabilities of the large model with the pendant industry. "GPT will become a 'power plant', data is coal, GPT will turn 'coal' into 'electricity', delivered to thousands of industries, all industries are worth reinventing with GPT." In the future, every industry, enterprise, and individual will have customized GPT. ”
On April 16, in his speech on "Artificial Intelligence Leads the New Industrial Revolution", Zhou Hongyi put forward six views, four major challenges, two predictions and one strategy.
Focusing on ChatGPT, Zhou Hongyi analyzed the reasons for the success of ChatGPT, and came to the conclusion that "GPT-4 is already the smartest person in the world" and "GPT already has real artificial intelligence", pointing out that GPT still has four unexplainable phenomena: emergence, hallucinations, language ability transfer, and logic enhancement. At the same time, under the big model fever, Zhou Hongyi keenly found that GPT will bring four major challenges, covering security, society, energy, artificial intelligence evolution and other issues.
Based on the insight into large models, Zhou Hongyi believes that China has the ability to develop its own artificial intelligence big language models. As one of them, 360 has formulated a strategic attack model of "two wings flying together and three concurrent ways", so as to grasp the technology with one hand and the scene landing with the other.
Subsequently, Zhou Hongyi implemented the strategy within the company. On May 31, Zhou Hongyi said on social media platforms: "In April, an internal letter has been released, asking 360 to fully embrace AI, and also leaving homework for all product managers to think about how their products can be reshaped with large models." "The strategy of combining multimodal large models with the Internet of Things has also planted seeds in April.
The results are obvious, and it is based on Zhou Hongyi's deep thinking that 360 can find its way when the big ship turns. In addition to this action in smart hardware, it is reported that 360 will be fully integrated in search, browser and other businesses, and further disclosure is expected in the future.
"The scene is ripe and open, and the 360 big model is ready to fight a protracted battle." Zhou Hongyi said.
03 Safety is a must-answer question in the AI 2.0 era
Zhou Hongyi's thinking about the big model did not stop because he determined the present, he was thinking about the next "to-do".
"Two days ago, I didn't attend the industry conference because I was impotent, so I used my voiceprint to clone a digital human of my own and gave a 10-minute speech. The presentation was good, but I also broke out in a cold sweat. I wondered, what if someone used my voice to say something I shouldn't? This is a security problem that multimodality inevitably entails. "Practicing the experience left over from the last Internet era, the development of technology is always accompanied by a symbiosis of security issues.
The problem that Zhou Hongyi is worried about is also the issue that pays the most attention to AI at the current social level.
AI scams stayed on hot searches for days. The owner of a domestic company chatted with a fake "friend" of a fraud gang for 10 minutes and was defrauded of 4.3 million. According to police investigations, the success rate of using DeepFake to imitate faces and voices is close to 100%; A mother in the United States received a call from a strange man, and her daughter's cry and cry for help were heard on the phone, and fraudsters used AI bionics to extort $1 million from her mother. AI scams are frequent in society, and for inexperienced people, it is almost impossible to judge the truth or fake.
"In the past, hearing was false and seeing was believing, but now seeing may not be believing." Zhou Hongyi sighed.
The technical confrontation and legal regulation of AI security set off large-scale discussions for a while. On May 16, Sam Altman, CEO of OpenAI, attended a hearing in the US Congress to discuss whether AI will generate dangerous content, whether AI will infringe on citizens' privacy, and whether data rights and AI will affect employment. Sam Altman said that in the face of increasingly powerful AI models, the intervention of government regulators is essential to control the associated risks, and agreed that AI should be treated with reference to the regulatory approach of nuclear weapons.
From network security and data security to AI security, Zhou Hongyi believes that it is necessary to upgrade through AI rights confirmation. "I think AI security cannot rely solely on technical confrontation, but must have regulations for manufacturers, such as AI practitioners must add fingerprints to AI-generated products, whether it is pictures, videos or sounds, to increase the cost of crime."
The cost of technology will also increase, but it is also the way to go.
Such security issues may also permeate other industries. "GPT, as the highest level of digitalization, will definitely bring a lot of security issues. Enterprise cloud, big data, Internet of Things, and computers from disconnected to networked all bring many security challenges. ”
In this regard, Zhou Hongyi also began to plan ahead, stemming from 360's security accumulation in the Internet era, and the exploration of AI security was half a step earlier than others.
"We have the world's largest cybersecurity big data, and we have recorded a lot of attack sample data, a lot of cybersecurity attack and defense examples. Use these to train a secure brain GPT to make it more effective in blocking and detecting malicious attacks. ”
04 Conclusion
Facing the problem squarely and solving the problem, whether it is Zhou Hongyi or 360, is changing and developing.
Regarding what is real security, Zhou Hongyi emphasized in his speech: "Not developing is the biggest insecurity. AI is an industrial revolution, and it can't be choked because it has safety concerns. ”
Zhou Hongyi's strong curiosity and vitality in front of the large model made this middle-aged man in his 50s look even younger. Curiosity is always the best driver, especially in the face of technological change, some people rest on their laurels and others rub their hands.
Only such people can not be left behind by the times and get tickets for the next era.