The so-called ChatGPT is to answer a large number of questions while imitating the human speaking style, generate some semantic answers through the new model of OpenAI, use the form of web application to interact with users, and realize simple chatting, information consultation, writing poems and compositions, modifying codes, etc.
Due to its high-quality answers, efficient ways to obtain information, and addictive interactive experience, ChatGPT has attracted widespread attention and attention from social media at home and abroad, and has created discussions and possibilities to replace Google and other search engines.
Even the world’s richest man and Tesla CEO Musk said: “Many people are caught in the damn crazy ChatGPT cycle.” Even domestic investors have begun to look for OpenAI, the company behind ChatGPT, to buy its shares.
Although there are a lot of praises for it, Stack Overflow, a question-and-answer community where programmers gather, has announced that it is forbidden to use machine replies generated by ChatGPT, and ChatGPT has unstable output results, no error correction system, difficulty in commercial implementation, and difficult protection of data security. Issues such as being unfriendly to real-name users in China, not open-sourcing key models, and entrepreneurship and investment bubbles caused by the AIGC (Artificial Intelligence Content Production) track have sparked extensive discussions in the industry.
In an exclusive interview with Titanium Media App, Dr. He Xiaodong, Vice President of JD Group and IEEE Fellow, said that China should develop its own ChatGPT model product.
“Frankly speaking, the direction (voice robot) is right, and there will be market demand for this in the future.” He Xiaodong emphasized that in addition to continuing to explore in research, ChatGPT has unique experience and value in text generation, and it is almost It’s close to being commercially available.
How exactly did ChatGPT come into being? Can it replace search engines?
“AI is about to become good!” Some investors expressed emotion in Moments after playing ChatGPT.
So, what is ChatGPT?
You can understand that ChatGPT is a task-oriented multi-round dialogue or question-answering system, which is positioned as a “general-purpose AI technical assistant”. Compared with Apple’s Siri and Xiao Ai’s “ordering meals, booking tickets, and booking hotels”, ChatGPT performs more open-domain tasks, such as question and answer, reading comprehension, reasoning, brainstorming, composition writing, and error correction. It can be said to be “AI version of Zhihu”.
On March 4 this year, the OpenAI team submitted a paper entitled “Instructions for Training Language Models to Follow Human Feedback” on the pre-printed website arxiv, detailing the InstructGPT large model with 1.3 billion parameters behind ChatGPT. technical capabilities.
OpenAI pointed out in the paper that the core of ChatGPT lies in the output of the InstructGPT model, and the deeper level is based on the GPT-3.5 base capability developed by OpenAI. Although the parameters of the InstructGPT model are 100 times less than the 175 billion parameters of the NLP pre-training model – GPT-3, the data model of the former has been selected, calibrated and fine-tuned, and the training results are more accurate, and the language model can be optimized more easily .
According to the team, ChatGPT mostly followed the training method of InstructGPT, and recruited human AI trainers during the data collection phase, while playing the roles of users and AI assistants to collect data. In the process, these AI trainers can modify the responses generated by the model based on the results of the initial model, which can be used to supervise the fine-tuning of the trained model. Subsequently, the engineer will compare the multiple generated results of the model, and select the output information in the data that is more in line with human preferences, and even has continuous learning ability. OpenAI statistics found that its API users prefer InstructGPT more than 70% of the time.
At present, ChatGPT is still in the free beta stage and is open to the public in unlimited quantities. And based on its millions of users, the cost of the platform is not small. Musk once asked “what is the average cost per chat,” and Altman replied: “The average user spends probably single-digit cents per chat.”
We found that some information on ChatGPT is very accurate, a bit like answers from Zhihu or Baike, and sometimes like search engine results from Google and Bing. But the accuracy of its answer is still open to question.
ChatGPT demonstrates the capabilities of a new generation of AI technology, perhaps in the not-too-distant future, a certain degree of AGI (General Artificial Intelligence) can do it. At least seamless human-computer interaction is optimistic. He pointed out that InstructGPT’s data is very good, and its advantages lie in high data quality, diversity, and RLHF algorithm. Among them, the gain brought by reinforcement learning is very large.
At present, the upgrade of ChatGPT lies more in the technical achievements after the adjustment of computing power and corpus. Both fluency and logic are very strong, and logic is stronger than content. However, the quantity of “dry goods” is relatively low, and some knowledge accuracy is lacking. “Actually, in a sense, I think its practical value is still inferior to that of Google (search).”
A former scientist at Google headquarters, said that compared to voice assistants and search engines, ChatGPT still has many things that it can’t do or can’t do well. ChatGPT’s “cognition” is based on virtual training texts, and it is not connected to real-time databases or information. It still has a taste of “castle in the air”, so it is especially suitable for gags.
Some users concluded that currently the topics that ChatGPT cannot answer include: cannot predict events that may happen but never happened and their consequences; cannot OpenMind or break the routine to establish new rules and new models; cannot go deep into the details of the question to answer but stay in the summary level; it is impossible to distinguish whether it is a factual question or not, and some non-factual questions cannot be answered; very partial questions (probably questions that have never been thought of by anyone) will not be answered; non-mainstream views cannot be listed; the dialogue context cannot jump too much Large, the domain span is too large, etc.
Author Parmy Olson argues, however, that Google Search crawls billions of web pages to index and then rank them by the most relevant answers, including a list of links for you to click. ChatGPT is directly based on its own search and a single answer of information synthesis, which makes the reply process easier.
In response to the outside world’s discussion on whether ChatGPT can replace search engines, ChatGPT itself gave a “negative” answer: ChatGPT is not a search engine. It is not intended to provide informational searches. Compared with search engines that provide information by indexing web pages and matching search terms, ChatGPT helps users solve problems by answering natural language questions. Therefore, there is no direct competition between them and they cannot subvert each other.
Earlier, OpenAI also released DALL·E 2 – an AI program that can generate images from text descriptions. Guosheng Securities pointed out that compared with the previous AI painting, ChatGPT has the potential to combine AI capabilities with workflow. If ChatGPT is integrated with creative tools such as Midjourney, it is possible to directly output design drawings and other content. Both ChatGPT and DALL E are new changes brought about by AIGC technology.
“I think the potential of AI can be further enlarged and tapped. Looking back at the past ten years, the progress of AI technology is indeed beyond imagination. 10 years ago, you could see breakthroughs in image and speech recognition, but you certainly didn’t think of paintings, articles, etc. AI can also be used to generate logical products, which are different. In the past ten years, AI has continuously surpassed and broken the expectations and imaginations of many people, which is quite remarkable.” He Xiaodong told Titanium Media App that this time The biggest innovation of ChatGPT is the new semi-supervised learning algorithm, which makes it possible to understand the user’s intentions and reach the level of a large-scale commercial search engine.
the
At the end of 2021, OpenAI has completed a US$250 million Series A financing. Investors include Microsoft, Musk, Google Ventures, Tiger Fund, A16z, and Altman’s former owner Y-Combinator. According to The Information, the seven-year-old OpenAI company is valued at nearly $20 billion, 500-800 times its estimated revenue in 2022.
After OpenAI released GPT-3 in 2020, Microsoft obtained an exclusive license on September 22 of that year. According to reports, OpenAI spends about $70 million a year on Microsoft cloud computing.
GIPHY App Key not set. Please check settings