With the in-depth development of AI big model, quantum computing, brain-like intelligence, cloud native, digital engine, audio and video and other technologies, new technologies, new models and new formats continue to emerge. For the above fields, we have planned a series of popular science columns, T-chat Frontier Hot Questions and Answers, to exchange the latest and hottest frontier scientific and technological topics with you in the form of ten questions and ten answers in popular language. We also welcome your guidance and put forward questions of interest to study together, continue to iterate and make progress together.
As a phenomenal hot spot, ChatGPT has received a lot of inquiries from all parties recently. Everyone is very concerned about its technology, impact and other contents. So we have the idea of writing this article. Please make comments from experts and industry colleagues.
How hot is ChatGPT now?
On November 30, 2022, ChatGPT came into being, forming a heated discussion around the world. According to the data of Similarweb, in January this year, about 13 million independent visitors used ChatGPT every day, more than twice as many as in December last year. The cumulative number of users exceeded 100 million, setting the record of the fastest Internet application breaking 100 million, surpassing the speed of TikTok breaking 100 million in the previous nine months.
In addition to the strange problems of the vast number of users, which triggered a large number of social media dissemination, ChatGPT's muscle in various fields has also further strengthened its popularity. The University of Pennsylvania found that ChatGPT was able to pass the final examination of its MBA program. The journal Nature announced on January 24th that AI tools were listed as the author's paper and could not be published in the journal. Nature said that the author's identity means that he is responsible for his work, while AI tools cannot bear such responsibility. It is said that ChatGPT also successfully passed the programming interview of Google and got the offer of L3 engineer with an annual salary of 183000 dollars.
The core technology behind ChatGPT?
ChatGPT is a form of generative AI, and Gartner takes it as the first in "Important Strategic Technology Trends in 2022". Gartner predicts that by 2025, generative AI will account for 10% of all generated data, but at present this proportion is less than 1%.
The support behind ChatGPT is the big model of AI. This technology point was described in detail in the report of Tencent Research Institute's "2022 Top Ten Frontier Application Trends of Digital Technology" last year: most of the current AI is trained for specific scene applications, and the generated models are difficult to migrate to other applications, belonging to the category of "small models". The whole process requires not only a lot of manual parameter adjustment, but also a large amount of annotation data for the machine, which reduces the efficiency of AI research and development, and the cost is high. The large model is usually trained on the unlabeled large data set by the method of self-supervised learning. Later, in the application of other scenarios, developers only need to fine-tune the model or use a small amount of data for secondary training to meet the needs of new application scenarios. This means that the improvement of large models can benefit all downstream small models and greatly improve the application scenarios and R&D efficiency of AI. Therefore, large models have become the focus of investment in the industry. Open AI, Google, Facebook, Microsoft, Baidu, Alibaba, Tencent, Huawei and Zhiyuan Research Institute have launched large models. In particular, the outstanding performance of the three models of OpenAI GPT in the fields of translation, question and answer, and content generation has shown the industry the hope of achieving universal AI. The current version of ChatGPT is GPT 3.5, which is based on GPT3 and is further enhanced.
One of the core technologies used by ChatGPT is Transformer. This can also be seen from its full name, Chat Generative Pre-trained Transformer. Transformer technology is one of the biggest highlights of AI technology in recent years. It is a deep learning model using attention mechanism proposed by Google in 2017. It can assign different weights according to the importance of each part of input data. The accuracy and performance of Transformer are due to the previously popular CNN (convolutional neural network), RNN (cyclic neural network) and other models, which greatly improve the effect of model training, and enable AI to further enhance its ability on the basis of larger models, more data and more computing power. In addition, he also has a strong cross-modal ability, not only in the field of NLP (natural language understanding), but also in voice and image.
Why can ChatGPT succeed?
ChatGPT is a very successful classic case from technology research and development to engineering, and then to kindling public enthusiasm.
technological innovation
Open AI continues to explore and introduce new technology routes in its continuous research. The reinforcement learning method introduced this time has greatly improved the effect of the model. In the research, Open AI introduced human experts in model training. On the one hand, experts help ChatGPT to write answers that are more in line with human habits, on the other hand, they also rank the generated results, and realize the fine-tuning and optimization of the model through such a reward mechanism.
organizational culture
Since its inception, Open AI has been committed to building the ability of general AI and has made steady and continuous investment, which is a key factor for the success of ChatGPT. Under such a vision, it attracted a large number of high-level talents and concentrated on R&D without any business KPI, and finally made a major breakthrough.
feedback mechanism
Through the use of users, Open AI can obtain users' feedback on the advantages and disadvantages of the model and the use experience, thus forming a closed loop from the use of the model to the experience feedback, so as to further optimize the model. In addition, Open AI also organizes a special feedback contest to encourage users to give feedback on the risks and hazards in the response, novel response suggestions, and other aspects. Participants have the opportunity to win $500 API points and can exchange corresponding prizes.
Project planning
An influential technology is inseparable from its project planning. ChatGPT's Q&A and multiple rounds of dialogue have well stimulated the enthusiasm and creativity of the public. People can freely play on topics of interest, either humorous or serious, and often swipe the screen in the circle of friends and the media. Including the previous launch of GPT3, his ability to write news, write poetry, translate, and code has also generated heated discussion and discussion in the society. This project design strategy of good interaction with the public is also worth learning. Including the previous AlphaGo Go game and Deepmind's solution to the problem of protein folding structure, these are good topics and tasks.
Is ChatGPT open source?
It is not open source yet. It is currently served by API (Application Programming Interface) calls, and there is no open source plan at present.
Open AI is not open, which is criticized by many people in the industry. Open source has been one of the core driving forces behind the vigorous development of the software and internet industries over the years. The open source approach can mobilize the enthusiasm of developers around the world. Everyone can download the source code and use it, optimize it and share it with the community. This mechanism of innovation with the strength of the whole society has greatly accelerated the process of technological research and industrial application.
Some experts in the industry also agree that ChatGPT is not open source, because AI technology is still a black box, and its internal mechanism is unknown. If the code is open source, it is difficult to avoid using the technology in some aspects that are not conducive to society and human beings. In addition, since Open AI abandoned its position as a non-profit organization and accepted investments from Microsoft and other companies, it will also adopt the approach of closed source and open application interface to promote the overall model from the perspective of commercialization. At the same time, it will also open up a small number of models to enrich the developer ecosystem.
Is ChatGPT free? Can we make money in the future?
Currently, ChatGPT is free for ordinary users.
According to Fortune, in 2022, OpenAI's revenue is expected to be less than US $30 million, with a net loss of US $545 million. With the popularity of ChatGPT, it may further increase its losses, because each call by users will make Open AI pay more computing resources and bandwidth costs.
Of course, Open AI also began to commercialize. In February 2023, OpenAI announced the launch of the paid pilot subscription program ChatGPT Plus, priced at $20 per month. The features of the paid version include queue free during peak hours, rapid response, and priority access to new features and improvements.
Open AI predicts that as ChatGPT becomes an important tool to attract customers, its revenue will grow rapidly. It is estimated that the revenue will be $200 million in 2023 and more than $1 billion in 2024.
In addition, it is reported that Microsoft is negotiating to invest about US $10 billion in OpenAI, which also brings its valuation to US $29 billion. Previously, Microsoft had invested $1 billion in OpenAI in 2019, providing it with a large amount of financial support and Azure cloud computing power support.
Will ChatGPT replace the search engine?
In the short term, no. In the long run, it is possible to form a new search engine through integration.
At present, the answers given by ChatGPT are uneven. It can give brief answers to many questions, and even write an industrial planning report that meets the level of a junior planner. However, ChatGPT also inevitably often appears "serious nonsense", which makes its credibility need to be further improved. Moreover, ChatGPT's current knowledge base is mainly based on the data before September 2021. The lack of new data greatly limits its ability to provide new information. At the same time, Open AI also launched WebGPT, which can find information on the network and provide information sources, which will largely supplement ChatGPT's real-time information sources, so as to further optimize the answer.
Under the logic of search engines, the right of choice lies with users. Search engines will generally give more relevant results and continuously optimize search results based on user click feedback. In recent years, search engines have also been trying to make the results unique, such as directly providing a unique answer choice in the search box, but at present, the number of unique results available is relatively limited, and it is unable to carry out multiple rounds of dialogue.
The launch of ChatGPT has brought great shock to search engine manufacturers and also inspired them to create new search experience. The two founders of Google have held several seminars to deal with the potential of ChatGPT to subvert the search engine business. Pichai, the CEO of Google, said that it would launch a large language model based on artificial intelligence similar to ChatGPT in the near future, a dialogue AI service Bard supported by LaMDA, which allows users to use it in the form of "search partner". Microsoft has added the ChatGPT function to the Bing search engine and carried out small traffic tests. Baidu also plans to launch a new function in March, adding a conversational robot like ChatGPT, "Wen Xin Yi Yan", to the search engine.
We can also look at ChatGPT's own answer:
What other models and applications are similar to ChatGPT?
ChatGPT is a typical AIGC big model. At present, corresponding generation models have emerged in text, code, image, video, 3D and other fields at home and abroad.
At present, a strong competitor of ChatGPT is Claude, who is the generative Al model recently launched by Anthropic. Anthropic was founded by several former OpenAl researchers in 2021, including Dario Amodei, former vice president of OpenAl research, Tom Brown, a GPT-3 paper author, and others. According to the Financial Times, at the end of 2022, Google invested about $300 million in this start-up company. Compared with ChatGPT, Anthropic put more emphasis on its work of building "reliable, interpretable and operable AI system" on its website.
Tencent's hybrid model, which integrates CV (computer vision), NLP (natural language understanding) and multimodal understanding capabilities, has successively topped the list of five authoritative data sets, including MSR-VTT and MSVD, to achieve a grand slam in the cross-modal field. In May 2022, Tencent's "Hunyuan" AI big model reached the top of the CLUE (Chinese Language Comprehension Evaluation Collection) total ranking list, reading comprehension list and large-scale knowledge map list at the same time, breaking three records at one stroke. In December, Hunyuan launched the first low-cost and landing NLP trillions model in China, and again reached the top of the list of natural language understanding tasks CLUE. The hybrid element can be hot started with a hundred billion models, and the training of the trillion-parameter large model HunYuen-NLP 1T can be completed in one day with 256 cards at the fastest. The overall training cost is only 1/8 of the direct cold start training trillion-model.
Will big models such as ChatGPT enable us to achieve universal AI?
Since the launch of GPT3 with 175 billion parameters in June 2020, the industry has further heated up the discussion on achieving universal AI. As Microsoft CEO Nadela said in an interview with the Wall Street Journal recently, the development of GPT is not linear, but exponential. Therefore, compared with GPT3, the current GPT3.5 has shown stronger capabilities. It is widely predicted in the industry that GPT4 will be launched this year with more powerful general capabilities.
But at the same time, we also need to have a clear identification. AI is not everything. That is, no matter how powerful AI is, the problems it solves are only a small part of all the problems facing mankind. Because there are massive problems in the real world that are not mathematical problems, it is impossible to solve them by calculation. At present, the problem that AI has found solutions is only a small part of the computable problem.
What social or ethical problems may ChatGPT cause?
ChatGPT also faces problems that AI has faced for a long time, such as ethical issues such as discrimination, security issues such as fraud, fraud, political rumor, identity forgery, and intellectual property issues. Recently, OpenAI said that it was considering adding model watermark to ChatGPT, that is, monitoring the special marks that the model can recognize, to reduce the problem of model abuse.
In addition, in response to the general concern about unemployment, ChatGPT shared a post on LinkedIn that said: "AI will not replace you. A person who uses AI will replace you".
What changes and new opportunities will ChatGPT bring in the future?
ChatGPT's own answer has shown a typical path of its development, that is, as a companion robot for human beings.
