Author|Xu Jiecheng
On the eve of the release of ChatGPT, Zoubin Ghahramani, the head of Google Artificial Intelligence Laboratory-Google Brain, also publicly stated: "Chatbots are not reliable for people every day. Things that are used locally." However, Zoubin Ghahramani must not have expected that the "slap in the face" would come so quickly. Just weeks after his remarks, ChatGPT was being used and discussed nearly all over the world.
Since its release at the end of November, ChatGPT has quickly become popular all over the Internet, with over 1 million users in just one week. Today, ChatGPT has become the LLM (Large Language Model) chatbot with the largest number of users in the world in history. And due to ChatGPT's innovation in information search mode, many ChatGPT supporters currently call it a "Google killer", believing that ChatGPT will be able to subvert the traditional information search mode in a short period of time and replace Google in the information search mode. Search domain dominance.
So, is the fact true as the majority of ChatGPT fans say? Does ChatGPT currently have the ability to challenge search engines? Is Google, the technology giant that has dominated the search field for decades, really as vulnerable as ChatGPT fans think?
Currently, most people’s love for ChatGPT mainly comes from its strong communication skills. When we ask questions to ChatGPT, it can not only Providing us with conclusions, but also being able to establish communication with us, asking any question to ChatGPT feels like you are talking to a real human being. This is indeed a major advantage of chatbots compared to search engines. The output of answers in the form of chat is far more direct and friendly than the line-by-line URLs of search engines.
But on the other side of the coin, ChatGPT also has three major problems in information search: unreliable answers, vulnerability to external influence, and forged information.
First of all, ChatGPT’s answers are not always reliable and often make mistakes in reasoning and fact. Secondly, ChatGPT is easily affected by external information. Among the unique features of ChatGPT, the most important one is its learning ability. The model can remember the content of previous conversations with other users and retell them. This also results in users being able to very easily intervene in ChatGPT’s judgment and answers to questions. Recently, the Stack Overflow forum banned all answers created by ChatGPT. The reason is that ChatGPT provides error examples learned from users for a large number of questions.
In addition, when faced with certain vague or descriptive questions, ChatGPT seems to have chosen to falsify some of the content it generates in order to make its answers more convincing. For example, when ChatGPT quotes experts' responses to specific questions in various social media, it sometimes modifies the text content, or matches A's remarks or events to B's name. The most typical example of this is when a reporter asked ChatGPT to write an article on Microsoft's quarterly earnings. In order to increase the credibility of the article, ChatGPT falsified a quote from Microsoft CEO Satya Nadella.
It can be seen that although ChatGPT has achieved a certain influence and a lot of attention in information retrieval through innovative question and answer mode, in terms of the accuracy and reliability of information supply, ChatGPT still has There are some obvious flaws, which are obviously contrary to the positioning and goals of search engines. And because ChatGPT was originally designed for conversational question-and-answer and simulated human conversational behavior, although ChatGPT can give certain explanations when faced with certain keyword retrieval scenarios, it cannot provide users with enough helpful additions. amount of information. Therefore, in terms of the current positioning and capabilities of ChatGPT, it is obviously not realistic to challenge Google or even completely replace the search engine.
Looking back at Google, in fact, long before the launch of ChatGPT, this giant had already participated in the research and development of a large number of LLM projects, the most famous of which is The LaMDA chatbot was described a few months ago by Google engineers as "already capable of human emotions." So, compared with the currently popular ChatGPT, how does LaMDA, which Google has always intended to hide its edge, perform?
As Google’s most mature conversational language model at this stage, LaMDA consists of 137 billion parameters and is trained on 1.56TB of publicly available conversation data and web documents. ChatGPT, on the other hand, is based on the GPT-3.5 architecture with 175 billion parameters and is trained through supervised fine-tuning on demonstrations of human-written text and samples rated by human annotators on overall quality scores.
At first glance, the difference in parameters and training methods between LaMDA and ChatGPT is not obvious, but in terms of specific use experience, the two show different characteristics. Riley Goodside, a researcher at Scale AI, once said in a blog: When faced with the same problem, ChatGPT can often give a clearer response, but the response form given by LaMDA is more friendly and closer to human language patterns. . This may be directly related to LaMDA receiving dialogue training, while ChatGPT has received more training in network text.
Therefore, although LaMDA itself also has some loopholes, at the practical application level, if Google hopes to use LaMDA to compete with ChatGPT, the outcome of the two is still unclear. The reason why Google has not released LaMDA to the outside world so far is mainly due to considerations of the company's reputation. Jeff Dean, head of Google AI, also claimed in an internal meeting in mid-December: Google currently has the ability to make AI products and technologies, but compared to small and medium-sized companies, they must make some decisions more conservatively.
When talking about Google AI, Jeff Dean said: Chatbots are easily affected by bias and false information, and Google, a large enterprise with more than one billion users, is even more difficult to escape this influence. . Of course we very much hope to use these technologies into real products, especially using better language models, but more importantly, we have to do the right thing. In addition, it is worth noting that Google has also recently adjusted the work of its internal teams and strengthened its research and development support for AI technology and products. From the above actions, it is not difficult to see that Google is still very confident in facing some of the challenges posed by ChatGPT.
In the final analysis, comparing chatbots with search engines is a false proposition. The value of ChatGPT lies in its ability to simulate the "thinking" behavior of real people, and use the knowledge it has learned through conversation to answer some of our questions, or complete creative and simple writing tasks. For search engines, their main responsibility is to provide people with the information they need quickly and accurately. The barriers that exist between the two cannot be effectively bridged through technical means.
As for the future development trends of the two, many industry experts have given almost unanimous views, that is, ChatGPT conversational robots will form a 1 1 > 2 "symbiosis" with search engines in the future. "Relationship has become a new entrance to search engines. In this new mode, people may be able to communicate directly with search engines through voice or text, and search engines can also use the understanding and cognition of conversational robots to quickly summarize for people from massive data Produce more accurate and preferred search results, help people save steps in information search, and greatly optimize the efficiency of people's entire search process.
It seems that although ChatGPT and Google currently have a certain opposition on the surface, this trend is more based on interests. I believe that as time goes by and technology continues to evolve, the current contradictions between the two will gradually disappear.
Reference link:
//m.sbmmt.com/link/3142c42ca1d6675586ec636631a0c922
//m.sbmmt.com/link/86f3f50a6f945bbfa351e55faac043fb
The above is the detailed content of Can ChatGPT kill Google?. For more information, please follow other related articles on the PHP Chinese website!