domeet webmaster

1

As a text worker, I deal with search engines every day.

For example, when writing Facebook’s earnings news, Google can tell me about its real-time stock price, market value, recent highs and lows, and other very useful information.

But actually, there is another tool that is better than Google, and that is Wolfram Alpha. It goes a step further than Google and can directly list the knowledge I might need in a structured way.

The simplest example: My home bandwidth is 75Mbps (9.375MB/s). How long does it take to download a 100GB file? I can ask directly in natural language. Wolfram Alpha will not only tell me the answer, but also write the formula:

It is not only a mathematical tool, but also a good knowledge aggregation tool. For example, the recent movie “The Clown” is very hot. If I want to write an article about it, I can find a lot of details on Wolfram Alpha, including not limited to video information, rankings and box office.Wait.

Exactly, Wolfram Alpha is not a search engine. Its official definition is called the “computational knowledge engine”, which can be used to answer questions that are not publicly available, but can be calculated by calculation. Moreover, it uses a structured approach to presenting answers, rather than listing links one by one like a search engine.

Next, introduce the protagonist of today’s article: Magi, a little-known tool among my friends in the technology circle in recent days.

Magi (https://magi.com) also looks like a search engine:

But once you play it, you will find that it is very different from all the search engines you have in mind.

When I used it to search for the term “易烊千玺”, I got the following result.

First of all, the answer provides several key descriptions of Yi Qian Qian, such as “Members of TFBOYS” and “Representation of National Idols after 00”. Then, it lists almost all of the important attributes of the owner of the entry, including the date of his birth, the film and television productions he participated in, and the published music albums.

The structured presentation of the answer is quite similar to Wolfram Alpha.

More interestingly, magi.com also replied to several pros and cons of Yi Qian, such as his nickname, his name and his pets.

Sometimes, Magi can give some results that I can’t help but…

Enter the “Evangelion”. The description of the answer includes “the famous model worker in the industry”…

After searching for Hideo Kojima, I was also drunk in the “Speciality” in the answer…

Next search for Fu Jian Yibo.

Maybe because of the majority of my career, I’m posting, magi.cOm tells me that Fu Jian’s hobby is “painting comics…”

Of course, most of the time the answer given by magi.com is quite reliable.

The answers to the search will be in green, yellow and red to indicate their credibility from high to low; on the right side of the answer will provide a few links, you can use the mouse to cross them. See, the answer is which one/some specific sources to learn from:

You will notice that in the result of magi.com, the answer is just below, and the link goes to the right, which is completely contrary to the user interface of the mainstream search engine.

This is the biggest difference between Magi and the mainstream search engine: The link is not the result, the answer is.

This is because Magi is not a search engine (although it has some search engine features). It is actually a machine learning-based knowledge engine that can retrieve and extract natural language texts from any field, extracting the knowledge and forming structured data.

It’s a bit simpler:

We all know that there is a lot of text-based information on the Internet, which contains a lot of knowledge. However, computers do not understand most of the information on the Internet, because such information is often not in the form of “sex: male”, “nationality: China”, but in the form of natural language.

For example, “the height of the Eiffel Tower” is an entry-level problem, because people have already sorted out the correct answers, written on Wikipedia and travel sites; but want to know “the second elevator line of the Eiffel Tower has How long is it, it is difficult to find accurate information on the search engine. This is because very few people will record these details in a structured way on the Internet.

This is what Magi wants to solve: Extracting knowledge from plain text in the open world and letting it be parsed, retrieved, and traceable.

2

Magi from the Chinese team Peak Labs, founder Ji Yichao is also well known in the developer community. In 2011, during his time at the Peking University High School, he completed the development of the mammoth browser iOS. The following year, he completed Rasgueado in just two days, the first iOS input method that supports swipe gestures to control cursor position.

In 2012, Ji Yichao founded his own company and continued to promote browser and input method projects. Currently, Peak Labs focuses on the Magi project, focusing on the technologies behind it and the development of related commercial products.

Peak Labs has no plans to compare Magi with mainstream search engines like Google and Baidu. Make Magi a “search engine” mainly to give the public a chance to experience the technology behind it and feel the value it can provide.

Even so, it looks like the search engine magi.com, the strength is not to be underestimated. In fact, for this exemplary product, Peak Labs did not choose a smart way to crawl results from other search engines, but instead developed an Internet search engine from scratch.

“The summary of our results is longer than the average search engine, yes, we deliberately. This is enough to prove that our results are unlikely to come from other search engines,” Ji Yichao wrote on the official website.

Depending on user input questions, keywords, and expressions, magi.com can present answers in different ways—the specific presentation also demonstrates the capabilities of the Magi system.

For example, if you enter “Taxis Software Company”, the Magi system can call all the mobile phones it knows to be listed in the answer by “collection”.

And on Baidu, the results are as follows. It can be seen that Baidu’s knowledge map also provides similar results, but it seems that it has not been updated in four or five years:

For example, if you enter “octagonal”, the Magi system will find that the two keywords are actually the same thing, and it will give the answer in the form of “assertion”.

As shown below, magi.com tells me that the octagonal and aniseed are “near-meaning items”, which are the relationship between “also known as” and “also called”.

The Magi system can learn 24 hours a day. Its timeliness is also pretty good. Peak Labs claims knowledge in real-time news, Magi takes only 5 minutes to master, and can also adopt new sources of information for cross-validation for automatic error correction.

If you stay on the home page of magi.com for a while, you can see the link it is currently learning:

In addition to the self-developed network-wide search engine, Peak Labs has developed a neural information extraction system based on attention mechanism, which does not rely on the distributed crawling system without the interface browser (the crawler program MagiBot), and supports mixing. Handle natural language pipelines in more than 170 languages.

These four are combined to be the full picture of the Magi system.

As an EVA powder, I have to interrupt it here: Magi and its four subsystems, all named fromEvangelion (EVA itself is based on other Western religious classics such as the Bible), and there are also eggs in the name:

• Magi (three sages, supercomputers with multiple systems)
• Search Engine Ramiel (Ray Angel)
• Neural information extraction system Ireul (horror angel, with learning and evolution ability)
• Natural Language Processing Pipeline Arael (Bird Angel)
• The reptile program Matarael (Angel of the Rain, looks like a spider)

Peak Labs pointed out on the official website that the current Magi technology is not yet fully mature.

It is true. At present, some problems can be observed through magi.com, such as many answers that can be easily found in mainstream search engines, which magi.com can’t give (usually because it has not yet learned).

For example, when searching for “the richest person in the world”, what I want is that Magi can tell me who is the richest at the moment, but it can only tell me the richest group of people:

“Maybe in the near future, along with the progress of the entire industry, the inclusive architecture of Magi will be the cornerstone for interpretable artificial intelligence,” writes the Peak Lab website.

——I hope this vision can be achieved. (It doesn’t matter if you can’t achieve it! Take magi.com and search for something weird, you can still get a lot of jokes…)

If you are interested in the technical details of Magi, you can go to the Peak Labs website Learn more about. Ji Yichao made a more detailed explanation in his answer.

[1] Ji Yichao knows how to evaluate Peak Labs’ 2019 Magi search engine? Answers https://www.zhihu.com/question/354059866/answer/881655371