Zhou Hongyi presented a self portrait generated by AI drawing and an opening speech generated by a large language model. He divided the development of artificial intelligence into three stages: artificial intelligence disabilities, strong artificial intelligence, and super artificial intelligence. Zhou Hongyi believes that GPT-4 has superhuman capabilities and real intelligence.

It is expected that GPT-6 to GPT-8 artificial intelligence will generate consciousness and become a new species. At the time of GPT-3.5, Zhou Hongyi predicted that it would grow "eyes and ears". Currently, GPT-4 has the ability to read pictures, which preliminarily confirms Zhou Hongyi's judgment. As GPT has evolved into "hands" and "feet", it has the ability to manipulate the world through access to Internet APIs.

| GPT-4 has evolved too fast, ranking first with full marks and passing a simulation interview with a large factory

Recently, GPT-4 passed the simulated interview with Amazon on LeetCode with full marks, surpassing all human beings involved in the test, and can be hired as software engineers.

GPT-4 has significantly outperformed previous models such as ChatGPT in these aspects, and is surprisingly close to the human level in all these tasks, which is

The paper by Microsoft researchers is called "The Spark of General Artificial Intelligence: An Early Experiment with GPT-4." In the paper, they wrote that in addition to mastering language, we have demonstrated that GPT-4 can solve novel and difficult tasks that span fields such as mathematics, coding, vision, medicine, law, and psychology without any special prompts.

In addition, in all of these tasks, GPT-4 performance is very close to the human level, and often significantly exceeds previous models such as ChatGPT. Given the breadth and depth of GPT-4 functionality, it is reasonable to consider it an early (but still incomplete) version of the General Artificial Intelligence (AGI) system.

To prove their claim, the researchers asked GPT-4 to prove that prime numbers are infinite and that every line rhymes. As a result, GPT-4 did it perfectly! Can humans do it? It is estimated that 90% of people do not even know what a prime number is, and 50% of people do not even know what a rhyme is. Proof is even more elusive.

Although GPT-4 meets or exceeds the human level in many tasks, overall, its intellectual model is significantly different from that of humans. However, GPT-4 is almost certainly only the first step towards a series of increasingly common intelligent systems. However, even as a first step, GPT-4 challenges a considerable number of widely held assumptions about machine intelligence and exhibits sudden behavior and capabilities, and its sources and mechanisms are currently difficult to accurately identify.

Is it a bit creepy, with a cold back? Unlike humans, the intellectual model of GPT-4 is currently difficult to accurately identify its source and mechanism. Does this mean that Microsoft actually doesn't know how GPT-4 is "evolving"? Moreover, researchers say that they "cannot access all the details of the huge training data of GPT-4." Isn't actual machine learning the use of neural networks to analyze massive amounts of data, identify their internal connections, and combine them into new knowledge?

At present, it may be that the evolution of GPT-4 is too fast, and there are many different or even contradictory voices within Microsoft. For example, Samuel Altman, the father of OpenAI, has repeatedly emphasized the limitations of GPT-4 and said, "We don't have a real AGI yet." Microsoft also stated that it is not focused on trying to achieve AGI.

| How far is it from GPT-4 in China? Make GPT-3.5 first

igital Intelligence Frontline learned that several large domestic internet companies and artificial intelligence enterprises have received intensive customer inquiries to understand the content related to ChatGPT and the big model. Especially among industry leaders and large enterprises, there is a sense of anxiety: how is the progress of domestic large models, and whether technology can keep up with them?

"In the past, it was widely believed that there was no significant difference in the progress of artificial intelligence between China and overseas, but after ChatGPT came out, everyone was a bit unprepared." A person from a major internet manufacturer told Digital Intelligence Frontline that after the large model parameters exceeded 100 billion, a qualitative change was achieved, allowing artificial intelligence to initially possess logic and reasoning capabilities.

The sense of urgency that ChatGPT brings is everywhere. Although there is still a significant gap between the effect and ChatGPT - even overseas giants such as Google and Facebook are not performing as well on large models as OpenAI - the layout of large domestic factories on large AI models has already begun. These include Alibaba's M6 model, Baidu's Wenxin model, Huawei's Pangu model, Tencent's hybrid model, and Zhiyuan's Wudao model, all of which have a scale of over 100 billion and are multimodal.

The efforts of domestic manufacturers in the field of artificial intelligence are obvious to all. Baidu has worked overtime in the past few months to catch up with ChatGPT's progress. In the middle of March, Baidu launched the ChatGPT like product "ERNIE Bot", which has been open to the public for testing. Although it is not as effective as ChatGPT, it has also caused active attempts by domestic users and enterprises. Tens of thousands of enterprises applied to call API services, and the server was once crowded.

In its public response, Tencent also made it clear that it has established a layout in relevant directions, and special research is also being carried out in an orderly manner. At Tencent's financial reporting meeting on March 22, President Liu Chiping also responded to investors' concerns, saying that Tencent would actively invest resources to build basic models and apply them to every business line of Tencent in the future.

Tencent has always been relatively low-key in its work, but in fact, Tencent has a lot of layout and time in artificial intelligence. According to Digital Intelligence Frontline, Tencent currently has three major artificial intelligence labs: Youtu Lab, which focuses on computer vision and industrial AI applications, WeChat AI, which focuses on developing voice AI, and AI Lab, which focuses on the combination of basic research and application exploration.

In the "four level acceleration architecture" previously disclosed by Tencent Cloud Intelligence, there are not only the lowest level computing power (self developed AI chips to accelerate computing power efficiency), but also the development level hybrid big model. Through Tencent Cloud TI platform, multiple industry big model fine-tuning solutions are provided, as well as plug and play standardized application industry and industry solutions, which have formed a complete chain from basic computing power to algorithm models to upper level applications.

In terms of the model layer that has attracted much attention from the outside world, Tencent publicly announced the AI big model "Hybrid" in 2022, Taking the meaning of "chaotic initial", it includes the CV (machine vision) large model, NLP (natural language processing) large model, and multimodal large model, covering the mainstream research directions in the industry. It has successively ranked top in multiple authoritative multimodal data sets such as the Chinese language understanding authoritative evaluation collection CLUE and VCR, MSR-VTT, and MSVD, achieving a grand slam in the cross modal field.

The advantages of the hybrid model lie in: first, Tencent's technical accumulation and reserves in the field of artificial intelligence have enabled it to achieve breakthroughs in multiple segmented tracks. For example, in the NLP field, relying on Tencent's Taiji machine learning platform, Tencent launched a trillions of Chinese NLP pre training models, HunYuan NLP-1T, last year, which topped the latest list of natural language understanding tasks, CLUE.

Second, Tencent has massive application scenarios. In addition to regular public data sets, the "hybrid" model also learns about text data sets unique to the business domain. Compared to other AI models in the industry, "hybrid" can better understand text information of various lengths, cope with diverse scene tasks such as search, advertising, news, and Q&A, and have more advantages in downstream tasks related to reading comprehension and knowledge mapping.

Nowadays, amid the noise of ChatGPT, domestic big manufacturers are accelerating their actions. According to Digital Intelligence Frontline, computing power and data are one of the key elements of the big model. Many domestic enterprises are purchasing Nvidia GPUs in large quantities, while some giants are looking for high-quality Chinese data to prepare for better training effects of the big model.

"The most important thing is to be able to reproduce GPT-3.5 now," said Yinqi. "GPT-3.5 is a more important point, at least to allow everyone to have a sense of application innovation and technological innovation on a common benchmark."

| What kind of imagination does the GPT6 application have?

Search must be the first scenario of landing in office software, such as Office 365; E-mail and video conferencing are also scenarios for generating content. In addition, applications such as service robots and intelligent customer service are also very good.

There are actually two revolutions now, one is the energy revolution, based on lithium batteries, from fossil energy in the past to our current lithium battery energy. Another revolution is AGI general artificial intelligence represented by ChatGPT. In the future, with more sensitive mechanical feedback, robots will be the largest application scenario.

Intelligent digital humans will definitely be applied earlier than robots. "Because there are many places where robots are constrained, including range, joint freedom, and so on, landing is much more difficult.". If it were a digital person in the digital world, its arrival might be much faster, such as a host, live streaming, digital star, and so on.

In addition, the impact on the game industry is also significant, especially as it can directly and significantly reduce the development costs of games. Drawing with AI can greatly improve the efficiency of game creators.

In the future, AI machines can replace many simple tasks, so people's creativity and ideas will become particularly important. Therefore, we believe that there will be a particularly significant demand increase for content creators.