GPT-4 integrated into Microsoft Office, "the next step may be ChatGPT operating system"

·“ Human-computer interaction is a technology field that strongly affects the global business landscape, and each generation of interaction has produced global giants... The next step may be the ChatGPT operating system, which has too much impact. ”

·“ Unfortunately, the recently released technical report on GPT-4 does not reveal too many technical details, such as model size, parameters, training data, or training methods, all of which are difficult to review or replicate. At the same time, there may be copyright issues, there are questions about data quality, which sets a negative precedent for the secrecy of the model. ”

This week is the week that AI is transforming the world. On March 14, local time, OpenAI, the developer of ChatGPT, officially released its landmark multimodal large model GPT-4. On the same day, Google announced that it would pack generative AI into its Workspace "Family Bucket." On the 16th, Microsoft announced the integration of GPT-4 into a series of office software tools, saying that "the way humans interact with computers has entered a new stage."

For the announcement of GPT-4 and Microsoft's big move, The Paper (www.thepaper.cn) interviewed a number of experts in the industry. Gong Yeming, dean of the School of Artificial Intelligence Management (AIM) and director of the Global Business Intelligence Center (BIC) at the French Business School of Lyon, said that it is a pity that the technical report of GPT-4 did not disclose too many technical details, which set a negative precedent for the confidentiality of the model. The integration of Microsoft's office software into GPT-4 will trigger a substantial revolution in the enterprise office field. "The next step could be the ChatGPT operating system." "Asian companies that didn't capture Windows, the operating system for graphical user interfaces, may have missed the opportunity this time," he said. And Microsoft is also the company most closely associated with OpenAI. ”

Duan Zhiyun, founder and chairman of the AIGC (artificial intelligence-generated content) field enterprise "Proxima Planet", believes that from a security point of view, if there is an AI that can collect all the information on the Internet and then process it, there will actually be a big information security risk, so China will definitely promote independent and controllable AI capabilities.

The Paper: What is the GPT-4 update that makes sense and what makes sense?

Chen Yijun (researcher at Shanghai Institute of Artificial Intelligence): I read the 98-page technical report released by OpenAI. I didn't expect GPT-4 to be released quickly, but its performance upgrade direction made sense.

The most important upgrade of GPT-4 is, of course, the support for multimodality, which greatly expands the application scenarios and possibilities of AI. People can tell the AI some simple ideas and ideas in the form of sketches or simple pictures, and let it create music, lyrics and paintings. The combination of multi-modal is completely the effect of 1 plus 1 greater than 2, and with the assistance of GPT-4, people's work efficiency will be greatly improved.

In addition, compared to ChatGPT, GPT-4 has improved the correctness of answering questions and grasp the details, which has also greatly improved the user experience. In its report, OpenAI attributes this improvement to "more data and computation," which is a predictable improvement from the perspective of GPT's iterative technical roadmap.

Yeming Gong (Dean of the School of Artificial Intelligence Management and Director of the Center for Global Business Intelligence at the Lyon Business School, France): It should be said that the GPT-4 update is very much in line with my expectations. The market response was also relatively good. GPT-4 has limitations similar to earlier GPT models: it is not highly reliable (the so-called "hallucination" problem) and has a limited context window. But this is also exactly what was expected, and I didn't expect OpenAI to solve the problem anytime soon.

Unfortunately, the recently released technical report on GPT-4 does not reveal too many technical details, such as model size, parameters, training data, or training methods, all of which are difficult to review or replicate. At the same time, there may be copyright issues, there are questions about data quality, which sets a negative precedent for the secrecy of the model. For early GPT models, relevant technical information has been released.

The Paper: Andreas Braun, CTO of Microsoft Germany, revealed, "We will launch the GPT-4 next week, which will be a multimodal model that will offer completely different possibilities - such as video." "Why doesn't there seem to be any mention of video at the time of GPT-4's release?

Duan Zhiyun (founder and chairman of Proxima Planet): Video will still be a little more complicated. Under the current technical route, video is actually a lot of pictures, and there is no obstacle to the generation or reading of a single picture, but the amount of video information is much larger, closer to the real environment than the picture, which is a problem of calculation.

The Paper: What do you think about OpenAI?

Kong Rong (Co-President of TF Global Prospective Industry Research Institute): Its organizational culture has a vision for the future world, not just short-term profit as the goal, but a strong goal on how to change the world. Under such a big vision, a group of people who want to change the world are gathered before creating this product. I think this needs attention, and the original environment in China may be relatively hopeful of quick results. But in fact, some companies are more concerned about the future opportunities and development of human beings, and think from the perspective of human well-being to make a product.

The Paper: So what is the situation in China?

Duan Zhiyun: First of all, I don't think we can expect too high, after all, we try to do the effect as well as possible in a relatively limited time. Second, don't be too pessimistic, because the race has just begun, which is equivalent to OpenAI just exploring the right direction and taking that step, but the starting gun has just started running. It does have advantages, especially in the European and American markets. But after running for a while, after the application and business model are becoming more and more abundant, it is still full of possibilities.

The Paper: When GPT-4 performs so well Chinese and the quality of the Chinese corpus may not be satisfactory, is it necessary to make Chinese large models?

Duan Zhiyun: I think it's necessary. First of all, from a security point of view, if there is an AI that can collect all the information on the Internet, and then process it, there will actually be a great information security risk, so China will definitely push such an autonomous and controllable AI capability. And GPT-3 has not been open source since then, which means that it is a complete black box for all users, even the countries that use it, so you don't really know what's behind it. From this level, we must have such an autonomous and controllable model in China. Even if you may be a little behind, you have to catch up.

Second, I don't think there's any point in blindly benchmarking against OpenAI, after all, they are pioneers. The advantage of Chinese is that the scale of use in the world is not small, and it is good if it can be differentiated in the Chinese. The problem with the corpus is that the information that is easy to be queried on the Internet may not be of high quality Chinese, but it does not mean that there is no such data. This is actually a problem of information silos, and as everyone slowly realizes the importance of this data, many fragmented databases may open up. With this matter, I think it will encourage the whole industry to do more data exchange.

The Paper: How do you see the possible impact of GPT-4 on this field?

Gong Yeming: The most direct impact is the impact on human-computer interaction. From the first generation of interactive command-line interface, to the second generation of interactive graphical user interface (Graphic User Interface), to the third generation of interactive natural user interface represented by ChatGPT. Human-computer interaction is a technology field that strongly influences the global business landscape, and each generation of interaction has produced global giants.

Microsoft is integrating OpenAI's big language model into its Microsoft 365 offerings, which is a particularly big impact on the industry, triggering a substantial revolution in the enterprise office and having a huge impact on the service industry. The next step may be the ChatGPT operating system, which has too much impact. Asian companies that failed to grasp Windows, the operating system for graphical user interfaces, may have missed another opportunity. And Microsoft is also the closest company to OpenAI.

The gap between other countries and the United States in the information industry may widen further. ChatGPT has a high moat, similar to chip manufacturing, you know that the principle cannot be made without conditions.