The craze of AI painting is not over, and AI Q & A has become popular again recently.

OpenAI recently released a new chat robot model, ChatGPT, but once it was launched, it was quickly out of the circle because of its high-quality answers, efficient access to information, and addictive interactive experience.

Many people who have experienced it exclaimed.ChatGPT could wipe out journalists, programmers and search engines.

Even Musk could not help tweeting that ChatGPT is terrifying and that we are not far from the dangerous and powerful AI. Recently, Musk tweeted and "ridiculed" Musk this weekend for not following ChatGPT news in the New York Times.

In terms of user data, less than a week after its launch, OpenAI CEO announced on Twitter on December 5th that ChatGPT had reached 1 million users.

What is ChatGPT?

ChatGPT is a chat robot model released by OpenAI. Its interface is simple and there is only one input box. AI will reply to the input and allow continuous chat in one context.

Since its launch, ChatGPT has been in the limelight, whether it's writing a rhyming poem, checking the bug of the code, answering scientific questions, and making suggestions for the future development of Twitter. It's all out of the question, and the content it replies to is random every time, but it generally maintains a certain degree of professionalism and information, which is of strong reference significance.

For example, knowledge explanation.

For example, writing a novel.

At present, some netizens have tried to let ChatGPT take the US college entrance examination; write novels; trick ChatGPT into planning how to destroy the world; and even let ChatGPT play the role of OpenAI and build ChatGPT dolls in the system.

Why is it so powerful?

According to the research report of Zhejiang Merchants and Guosheng Securities, the main improvement of ChatGPT compared with the past is the memory ability. ChatGPT can store dialogue information and extend the context, so as to achieve continuous dialogue, which is very important in the dialogue scene and greatly improves the user experience in the dialogue interaction mode.

Specifically, the new training method is called "reinforcement learning from human feedback" (ReinforcementLearning from Human Feedback,RLHF). This training method increases the demonstration of the output results of the model, and sorts the results. In terms of specific operation, the artificial intelligence trainer acts as both sides of the dialogue, that is, the user and the artificial intelligence assistant, to provide dialogue samples. When a human acts as a chatbot, the model will generate some suggestions to assist the trainer in writing responses, and the trainer will rank the response options and input better results back into the model. Fine-tune the model through the above reward strategy and continue to iterate.

Compared with previous models, ChatGPT also has the following characteristics:

1) the error can be admitted, and if the user points out the error, the model will listen to the advice and optimize the answer.

2) incorrect premises can be questioned and false descriptions can be reduced. for example, when asked about "Columbus came to the United States in 2015", the robot will explain that Columbus does not belong to this era and adjust the output.

3) due to the moral training adopted by ChatGPT, ChatGPT has significantly improved in reducing harmful and untrue responses, such as refusing to answer questions seeking bullying programs and pointing out their unfairness.

In addition, the ChatGPT can not be separated from the big model, big data, big math.

Behind ChatGPT becoming the milestone of AIGC is the large model training supported by big data in the development of computing power and the formation of the digital age in order to achieve the current results. ChatGPT, developed by OpenAI, is a fine-tuned GPT-3.5 series model with as many as 175 billion model parameters, and training was completed at the beginning of this year. The support of big data is indispensable behind the model training. The public crawler data set mainly used by OpenAI has more than a trillion words of human language data set. In terms of computing power, GPT-3.5 trains on the Azure AI supercomputing infrastructure (a high-bandwidth cluster of V100GPU), which consumes about 3640PF-days (that is, 1 trillion calculations per second, running for 3640 whole days).

AIGC mileage monument application

In the previous article, we introduced AIGC-- artificial intelligence to automatically generate content, which is currently mainly used in text, painting, images, video, audio, games and virtual human. Obviously, the AI question and answer ChatGPT belongs to a branch of AIGC.

Guosheng Securities said that AIGC will not only change creation, but will also change the main way of obtaining information. On top of this, ChatGPT shows that as the existence of productivity, ChatGPT has partially surpassed today's search engines in finding answers and solving problems. ChatGPT may change the way we get information and output content in the future. AIGC is expected to become a killer application that drives the outbreak of demand in the digital economy era.

In terms of market space, Gartner expects generative artificial intelligence to account for 10 per cent of all generated data by 2025. According to Generative AI:A Creative New World, AIGC has the potential to generate trillions of dollars in economic value. Guotai Junan said that in the next five years, 10% of the picture content may be generated by AI, corresponding to more than 60 billion of the market size.

Related industries

Compared to previous AI paintings, ChatGPT has the potential to combine AI capabilities with workflow. At present, giants such as OpenAI have begun to build their own computing clusters to meet the requirements of AIGC. This centralized training has a huge demand for infrastructure such as GPU, storage, computer room, power supply and so on. Some domestic entrepreneurial teams are currently based on public cloud training, and are expected to turn to self-built clusters as the demand increases.

Guosheng Securities said that since the beginning of this year, China's strategy of "counting east and counting west" has been advancing rapidly from top to bottom. As the foundation of the digital economy, the importance of "computing power is power" is self-evident. Before, more of our Internet applications focused on "saving" rather than "calculating". With the rise of AIGC technology in the world, the demand for arithmetic power of "counting east and west" will become more and more significant.

As for the company, Zheshang Securities said that from the perspective of downstream related beneficial applications, including but not limited to code robots, novel derivatives, dialogue search engines, language partners, voice work assistants, dialogue virtual human, and so on. From the upstream increasing demand, including computing power, data tagging, natural language processing (NLP) and so on.