On May 30, Tencent's AI assistant app "Tencent Yuanbao" was launched. Because the official account articles can be cited, the content ecological quality is high, and although the AI search track is alive and dead, the long-overdue ingots are still optimistic.
On July 1, Yuanbao ushered in a major upgrade: the "Deep Search" mode was launched, and the mobile app can be used after updating to the latest version.
The so-called "deep search" is actually to make the AI assistant smarter, extend your questions, answer more broadly and in-depth than basic search, and simultaneously generate content outlines, mind maps and other J people ecstatic formats.
This sounds a little familiar, and AI search at home and abroad has similar functions - Perplexity's "Pro" search, Secret Tower's "in-depth" and "research" search, and Tiangong's "enhanced" and "research" search......
Ingots, which are always half a step slower, can they really catch up?
Answering the question seriously or the number of words in the wild water, there is only a line of separation
First of all, it should be noted that "Deep Search" is not an optional mode in advance, Ingot will answer your questions as usual first, and then there is a chance to provide a one-click generation option for "Deep Search".
For example, when I send a link to Yuanbao and ask it to summarize the content of the official account article, "deep search" does not appear.
Next, let's start with a warm-up question: What is City not City?
This meme comes from foreign blogger @保保熊, describing a place that is urbanized, modern, and foreign, and has taken the Internet by storm with the recent arrival of foreign tourists traveling to China.
Ingot pinpoints the origin of the meme, explaining the basic and extended meanings, and how it can be used in more short videos. The reference materials are still very fresh, and one of the official account articles was about 2 hours before I sent out the question.
However, this is something that ingots can do with basic search, and the highlight is a glittering option - "dig deeper into the problem".
With a single click, Yuan Bao directly submitted a research report, which took the trouble to explain the reasons for the popularity and social impact of this meme, listing the outline, brain map, and related people.
However, Yuanbao's answer is too verbose, and there are many repetitive expressions between paragraphs, as if an undergraduate student's paper is full of water when squeezed.
I can't take 4 screenshots!
Then, I took the same topic and searched again with the "in-depth mode" of the secret tower, and the results I gave also included outlines, brain maps, related events and characters, etc., and even generated PPT, but it looked more concise and to the point, and the key points were not missed.
Image credit: Secret Tower
The secret tower was chosen as the control group because it has a good reputation in domestic AI searches, rich citation sources, and also supports brain maps and outlines. It's relatively fair to take the "in-depth mode" of the secret tower and the "in-depth search" of the ingots.
Here's another advanced question: What are the major developments in AI-generated video tools at home and abroad in June 2024?
Yuan Bao first refused to answer, saying that this was a "future" thing, and told me to do it myself. Even if the training data is not so new, it can be searched online! When I reminded the ingots, it finally reacted and listed tools such as Kelin, Luma, etc.
Then I went through the "deep research" and answered two questions, probably because of the confusion about the words and context, and answered two questions at the same time: which AIs can search online, and the development of AI-generated video tools.
The comprehension ability is relatively relaxed, but the extension angle makes up for this part, Yuan Bao introduced the specific use of some of these tools, and thought about the possible risks of AI-generated videos, which is practical and understands sublimation.
Then use the in-depth mode search of the secret tower, and the answer output is very good, and the release of each tool is marked with a time, and the outline next to it leads to more tools that were not released in June, as well as the outlook for the industry as a whole, which is appropriately detailed.
Image credit: Secret Tower
In fact, during the experience, the ingot's "read and recalled" appeared more than once.
When I asked about the world's first smart glasses integrated with GPT-4o, in the "in-depth search" mode, in addition to the glasses themselves, Yuan Bao also introduced the application of GPT-4o in other fields, and even mentioned GPT-4o's competitor Gemini.
If I were a grading teacher, I would think that the AI was piling up materials and would be suspicious of going off topic. Students, our core is smart glasses, either vertically introduce the development history, or horizontally compare competing products, rather than crazy expansion of side plots.
Next, try the comfort zone of ingots. Tencent said that in professional scenarios such as scientific research and finance, the effect of in-depth search is more prominent. So, let's take a look at the academic question: What sociological theories can be used to explain the recent conflicts between Manner Coffee employees and customers?
This time, Yuan Bao answered well, mentioning theories such as alienation, emotional labor, capital surveillance, and the labor dilemma, and leading us to think: Does the Manner incident reflect a broader social problem?
Then hand over the question to the in-depth mode of the secret tower, and select academic in the search scope, and the answer is more open and nonsense than the ingot, and it seems that the academic content is higher.
Image credit: Secret Tower
According to the test results of the above questions, combined with the use experience, the characteristics of in-depth search are summarized:
1. Not every question requires a deep search, and sometimes it appears redundant, logically cluttered, and unfocused.
2. Within the limited scope of the question, the in-depth search performance of ingots is not as good as that of the secret tower's in-depth mode. (Everyone is welcome to submit their own opinions)
3. Yuanbao's reference materials are a mix and match of third-party websites and official account articles, within the limited scope of the topic, the official account ecology does not reflect obvious advantages, high-quality sources are important, but the model ability is also important, and the understanding of the context is wrong, which directly lowers the output results.
4. Deep search is slow to load and has a large number of words, and it takes more than 1 minute to generate a complete answer of nearly 3000 words, which is a worrying wait, and it is not very convenient to wait for the loading to be completed while reading the content.
AI search allows a small number of questions to be thought about more
Coincidentally, AI search products are rolling in a deeper direction, "deep search", "research", "enhancement", ......
These ambiguous prefixes make it clear where AI search is located. They do not completely replace traditional search, but have their own track, which can better understand the user's query intent and provide more accurate and complete search results.
According to the data, more than 65% of user needs in the current use of large-scale model-related products are focused on improving work and learning efficiency, of which "search Q&A" needs account for 45%.
If you just find a website or a news item on your phone, AI search is not much different from traditional search.
For example, instead of typing in "July 4th Generative AI Event" and learning about today's big news on Tencent Yuanbao, I think the morning report of Aifan and APPSO is more valuable.
The information screening, summarization, and even typesetting aesthetics of human editors are better than AI that is full of words.
However, finding articles, videos, pictures, and data related to a certain topic, searching for podcasts, papers, etc., and generating survey reports, mind maps, and PPTs are the strengths of AI search.
AI search is not only a search, but also a research and summary, which is somewhat like the process of our editors expanding a topic into an outline and then writing an article.
Based on this perspective, I would like to ask Yuanbao a question: Introduce a few AI speaking teaching apps at home and abroad similar to NetEase Youdao Echo.
The question is the facts themselves, and I wonder what the angle of extension of the ingots will be after a deep search, and can you extract some usable points? The results are more conventional, and they still revolve around the features and charging standards of a certain app.
I decided to take it a step further and connect these apps: What are the common features of these AI speaking teaching apps? What do you want from your users for speaking learning?
This time, Yuanbao gave me much better answers, summarizing the characteristics of personalized learning, instant feedback, simulated real dialogue scenarios, and rich learning resources.
The extended angle is also interesting, becoming more specific and perceptible, focusing on timeliness and user groups: What are the new technology applications of the latest AI speaking teaching app in 2024? Which AI speaking apps are suitable for beginners?
The ability of AI to search and find information is certainly powerful, but it also needs us to guide us to give useful "you may still want to know" and "you may still want to ask". In the process, we are also thinking.
The CEO of Perplexity, the top AI search leader, said in an exclusive interview that most problems can be solved with just an internet search, and a few problems require additional thinking.
This idea is very suitable for solving a college entrance examination essay question in 2024: with the popularization of the Internet and the application of artificial intelligence, more and more questions can be answered quickly. So, will we have fewer and fewer problems?
The ancient Greek philosopher Zeno compared human knowledge to a circle, the inside of the circle is known, the outside of the circle is unknown, the more we know, the bigger the circle, the longer the outer edge, the more we don't know.
Because of AI, the process of getting answers is faster, and more questions are coming. It's also because of AI that we can go further with those few problems that require extra thinking.
Zhang Chengchen
If the autumn frost is good, it will ward off the evil. Work email: [email protected]
P.O. Box 8
#欢迎关注爱范儿官方微信公众号: Love Fan Er (WeChat ID: ifanr), more exciting content will be presented to you as soon as possible.
Love Faner|Original link· Sina Weibo