In just a few days, OpenAI's chatbot ChatGPT swept the Internet, and the number of users easily exceeded one million.
Answering questions, writing codes, writing papers, creating poems and piano music, when human beings rack their brains to set up "nine-nine-eighty-one difficulties", ChatGPT basically asks and answers, even if it doesn't know how to make it decent.
ChatGPT is by far the best text generation AI out there for the masses, not to mention free to use.
When ChatGPT is proud of the horseshoe, some people see its frustration under its halo. The jobs of programmers and text workers may not be guaranteed, and even traditional search engines such as Google may be revolutionized by it.
With search engines, we still need to spend a lot of time flipping through the web to find answers. Wouldn’t it be better if AI could deliver the answers directly to your eyes and ensure the correct rate?
But the question is "what if".
ChatGPT: I can't compare to Google
On December 1, developer Josh Kelly posted the different results of the same code question on Google and ChatGPT. ChatGPT's answer seemed to be of higher quality, which made him sigh "Google is done" (Google is finished).
Has the fledgling ChatGPT really put a knife on the neck of Google search?
First look at the difference in definition between the two.
The core of the search engine is the collection of massive information, rather than the creation of information. You enter keywords in the search box, and the search engine crawls, indexes, and sorts the results that match your query according to the algorithm, and then you see a large number of links, and then find the information you need from them.
ChatGPT belongs to AIGC (Artificial Intelligence Generated Content), which is a new way of content creation. It has been trained by the data set. Through one-on-one dialogue and human-like tone, it can give a single and instant answer. It can also combine context to achieve multiple rounds of dialogue to help you solve more complex and continuous problems. .
You can guide the rules step by step, let it design games and other products, or give it a program, let it check for bugs, or give it a demonstration case, so that it can draw inferences about other cases from one instance. The more complex the interaction, the richer the ability of ChatGPT will be. It is only used as a search engine for one round, but it is a bit "talented".
Smarter than traditional chatbots and faster than human Q&A sites, ChatGPT makes the process of finding information more intuitive and simple.
One is generative search, and the other is large-scale search. The current ChatGPT is far from being able to replace Google.
In the basic setting of ChatGPT, the timeliness of information has fallen behind. ChatGPT is trained on billions of text examples based on the Internet, and its learning career is stuck in 2021. It is not connected to the Internet, does not call external network resources in real time, cannot recommend restaurants for you based on location like Google Maps, and cannot help you open Twitter, Facebook.
In contrast, Google not only has a huge database of indexed websites, but also provides functions and tools such as images and maps, allowing you to quickly access other websites is its "basic literacy". In fact, the most popular searches on Google are for keywords like “Facebook” and “Amazon.”
However, ChatGPT is the most complained about its accuracy. It listed the chicken and rabbit in the same cage equation but solved it wrongly. It insisted that the founder of Douban was not Abei. The ending of Journey to the West was made up and nonsense in a serious manner. That's all, the tone is very confident, and the source of information is not provided. When you use a search engine, you can at least cross-validate with multiple sources.
On December 4th, Musk revealed that ChatGPT had access to the Twitter database, which meant that ChatGPT must have learned a lot of data that lacked fact-checking. Because of the flood of spam answers, the programmer question-and-answer website Stack Overflow has temporarily banned users from sharing content generated by ChatGPT starting from December 5.
In addition, ChatGPT's answer was mixed with a lot of soup, which was ridiculed as a "hydrological artifact". OpenAI explained that this is mainly due to training data bias (trainers prefer longer answers that appear to be more comprehensive) and over-optimization.
With various deficiencies on the table, ChatGPT will not threaten search engines for the time being.
▲The AI killing incident in "2001 Space Odyssey".
Although ChatGPT is more human-like than other AIs, it is essentially a "random parrot" that cannot be separated from large models, big data, and large computing power. Its knowledge comes from the statistical laws in the training data, and it does not understand what it says. On December 5th, "Time" magazine "exclusively interviewed" ChatGPT. When asked "what are the limitations of large language models", ChatGPT replied:
One limitation of large language models is that we cannot understand the context or meaning of the words we generate. We can only generate text based on the probability that certain words or sequences of words occur together given the training data.
This means that we cannot provide explanations or reasoning for our responses, and we may not always generate responses that are fully coherent or meaningful in the context of the conversation.
But ChatGPT begs a question: what is understanding? "Understanding" is a human concept, can it be used in ChatGPT? If it can be used, is its understanding equivalent to the result of "generating a meaningful answer", regardless of the track?
At least for a period of time in the future, we may need to worry more about whether the quality of the results given by search engines will improve or decrease when more and more web content is generated by similar models.
I also asked the ChatGPT deity, its attitude is very humble, and I don’t think it should be compared with Google:
I'm a big language model and can't compare to Google or any other search engine. My purpose is to help users get information by answering their questions, not to provide a search function. My advantage over search engines is my ability to answer complex questions and provide more personalized and detailed information. In contrast, my disadvantage is that I can't provide a search function, and I can't provide the latest information.
AI may be the future of search engines
While ChatGPT is still in its infancy and hasn't done a better job than Google in search, it could change the way we search for information online and take search technology to the next level.
There is a simple example right now, there are already a large number of plug-ins that embed ChatGPT into Google, and some of them display the results of ChatGPT on the right side of the webpage, and one search takes into account two experiences.
Referring to the linkage between ChatGPT and image generation model Stable Diffusion (probably because AI understands AI better, the description of ChatGPT is easier to be extracted by Stable Diffusion, and the final picture quality is higher), ChatGPT may also be used to explain and guide keywords, Help us better find information with search engines.
In addition, the lack of timeliness and accuracy of ChatGPT is not unsolvable.
On the one hand, Zhihu answerer and natural language processing expert @张俊林 pointed out that it is very challenging to integrate new knowledge into large-scale language models in near real time. One solution is to store it in the index of traditional search engines. If ChatGPT cannot answer time-sensitive questions, you can turn to search engines to extract corresponding answers.
On the other hand, Bloomberg reports that OpenAI is developing an AI system called WebGPT that will be able to answer questions more accurately and even illustrate the source of a citation.
These are the combination of AI and search engine. If we imagine more boldly, regardless of technical limitations, and throwing away search engines, there is an omniscient AI that provides relevant and accurate information in an easy-to-understand form of questions and answers. This is the ideal appearance of future search ?
Many AI experts believe that the vision itself is problematic. Benno Stein, a researcher at the Bauhaus University in Weimar, Germany, points out that it can hide real-world complexities:
The problem is not the limitations of existing technology. Even with perfect technology, we can't have perfect answers. We don't know what good answers are because the world is complicated, but when we see these straight answers, we stop thinking.
So how to make the answer appear more "complicated"? Some people think that simply providing a list of documents will be more useful than giving answers directly; others suggest that the answers can be explained and the pros and cons of different views can be given, so that people can know what they are and why.
▲ Picture from: Getty Images
But most of the time, there is no truly perfect answer. These accurate and detailed measurement standards are more aimed at factual and knowledge-based questions, rather than those open-ended propositions that are wild and unconstrained.
Frame AI by the accuracy or detail of the answer, but it is a bit "picture". Let us go back to the positioning problem mentioned above. ChatGPT is a generative search, Google is a large-scale search, the former is chat, and the latter is search. They are essentially different.
ChatGPT has been popular for a while, and we have a general consensus on it: it has a lot of wrong answers, especially in knowledge and factual questions, but if you put it in a link of creation, it can be used to inspire ,increase productivity.
It is not a search engine, nor is it like a chat robot, but more like a "super brain" that you can consult at any time. In other words, ChatGPT will not necessarily subvert Google, but it has fundamentally changed the way we get along with knowledge. You can talk to it about the stars and the moon, from poetry to philosophy of life.
ChatGPT's stimulation of creativity and broad thinking may be more important than the accuracy of factual information. It can completely complement each other with search engines and human labor, and it is not necessary to fight each other to complete a piece of the puzzle leading to the unknown. This is what we respect for " The basic need of search".
A search engine is more than just a question answering machine
Since the birth of ChatGPT, there is no shortage of voices that Google search will be replaced.
In fact, Google has not fallen behind. It trained the AI chat robot Sparrow on DeepMind's large-scale language model Chinchilla, and also developed the dialogue neural language model LaMDA.
In May last year, Google researchers published a paper entitled "Rethinking Search" , describing a new type of search engine. Large language models provide concise professional answers with the help of algorithms, and users do not need to search for information in a large list of web pages. It sounds like ChatGPT.
Why didn't Google launch something like ChatGPT directly to the public like OpenAI did, or integrate it into its own search? Alphabet engineer @hncel believes that the problem is mainly cost and delay:
Large-scale language models like GPT are one of Google's main areas of research. Google has a lot of budget and people to deal with these models, but the economics of actually using these language models in the largest Google products (such as search, Gmail) are not yet. fully present.
It's one thing to release an interesting beta, but it's quite another to integrate it deeply into a system that serves billions of requests per day, taking into account the cost of serving, the added latency. It would take at least a 10x cost reduction to integrate such a model into products like search.
At the same time, large language models will also affect Google Search's current business model-about 81% of Google's parent company Alphabet's 2021 revenue of $257.6 billion comes from advertising, most of which are Google's pay-per-click ads.
AI like ChatGPT greatly reduces the number of pages, preventing people from browsing and clicking more ads, and the advertising revenue will drop accordingly.
Having said that, the explosion of ChatGPT has also made us more or less aware that the inherent model of search engines "indexing, retrieval and sorting" has ruled for more than 20 years, and Google makes thousands of changes to search engines every year. Most are minor and haven't changed radically.
In 1998, a pair of Stanford graduate students published a paper on a new type of search engine:
In this paper, we introduce Google, a prototype of a large-scale search engine that makes heavy use of structures in hypertext. Google efficiently crawls and indexes the web and produces more satisfying search results than existing systems.
The innovation of the past has become the tradition of the present, and traditional search engines such as Google face more than future AI.
For example, some people have called TikTok the "new Google". Foreign netizens use TikTok to search, which is a bit like we search for strategies in Xiaohongshu. It is really useful in the fields of food and movie lists. There is a trend hidden behind this: In the world where TikTok and Douyin "dominate", the Internet is more intuitive, more visual, and more interactive than before, and search is no exception.
But TikTok won't really shake Google. To find more information, to visit more websites, you still have to go back to Google.
Now that the changes have taken place, Google also needs to bring a better search experience in a more natural and intuitive way.
In recent years, Google has been shifting in this direction thanks to advances in artificial intelligence, machine learning, and computer vision, including the introduction of camera and microphone search, multiple searches for images and text , immersive views in maps, and more.
To put it simply, the input and output of Google search have become more "multi-sensory", more active, and more able to guess the user's mind .
▲ The machine learning model MUM makes the Google search engine "smarter".
Many of Google's search-related projects are still in the exploration and testing phase. At the annual Search On event in September this year, Liz Reid, vice president of Google's search products, gave a possible future example:
If Google knows you're interested in woodworking, in addition to answering one of your search questions, it'll also show you new tools you didn't know, YouTubers you've never heard of, and where you can learn more New skills and more.
Liz Reid believes that Google Search is not just a quick question answering machine, but a system for exploring, discovering, and learning about things you don't have a clear answer to yet.
To some extent, whether iterative search engines or aggressive general-purpose AI models, one is the fine-tuning of the inherent framework, and the other is the reform of starting from scratch. They are all making knowledge easier to obtain, making information screening more intelligent, and reducing your Learning threshold, shorten your learning process.
Prabhakar Raghavan, senior vice president of Google, made an interesting point that search is still a problem that is far from being solved . "If you give me all the machines, I will still be bound by human curiosity and cognition."
Before searching for better answers, we need to know how to ask questions. In the future, the ability to organize data may no longer be scarce, and the ability to ask questions and original opinions based on individual experience and emotions will be more precious. When you are introduced to the gate of knowledge, the speculative and creative nature of human beings is highlighted in an unprecedented position.
#Welcome to pay attention to Aifaner's official WeChat public account: Aifaner (WeChat ID: ifanr), more exciting content will be presented to you as soon as possible.