Author| Utada
| Tiger Sniff Technology Group
Cover | Visual China
Just when Yuan Xingyuan happily said that his new software was more popular with the post-00s in Station B, my mood gradually sank to the bottom.
This is an AI story continuation program called "Caiyun Little Dream". Just type a few crosses and it will give you a balding story orientation.
On the B station, many young up owners who like online texts, although they have not obtained any writing inspiration from it, they seem to have found the traffic password. Because, algorithms are obviously much cuter than humans:
In the world of AI, "Trump can fall in love with Macron the next day after falling in love with Biden...";
Cao Cao in the Romance of the Three Kingdoms could borrow money and grain to Liu Bei;
Beginning with "Chinese football won the World Cup", in a few paragraphs, there was a dream-back prediction that "after the World Cup, the national football team suffered a heavy blow and was ridiculed by the network group"...
The traffic from station B made young people hit the results of the continuation of chicken blood
Hundreds of thousands of plays are the best evidence that Xiaomeng is sought after by young people. It's just that after laughing, any profit organization has to trace the commercial imagination of this program.
Frankly, over the past 6 years, this kind of entertaining algorithmic use has not been uncommon.
Even the public outside the artificial intelligence circle has witnessed the huge success of AlphaGo in the game arena to the revenue dilemma of its production company Deepmind; from the explosion of the face-changing software Zao to the rapid passing of the air in just one year; and the "integrated technology marketing package" of Google, Baidu and other technology companies such as AI playing the piano and painting and writing poetry, and the commercial significance is only at the level of technical verification.
It can be said that what I have seen in the 6 years of the industry is that most of the "hot time" in the Chinese market that does not rely on large platforms and entertainment-oriented AI algorithms and systems has only one Spring Festival holiday time on average.
"So, if the novelty is over, will they still come to play?" I'm more curious about young people's loyalty than the app's cumulative 1 million downloads.
And the developer of this software, Yuan Xingyuan, is very honest. Because this is the pain of all app developers since the beginning of the Internet era. The app, launched in the second half of 2021, although the user growth is still good, but the retention rate is not high.
"7-day retention is 5%, while 30-day retention may be lower."
After all, young people's curiosity comes and goes quickly. Once they master and get used to the AI's routine of continuing stories, it is difficult to guarantee that they will stay in a world composed of only words and simple logic for a long time.
As the founder of the startup Caiyun Technology, Yuan Xingyuan has launched two very well-known AI software products in the past 6 years - Caiyun Weather and Caiyun Xiaoyi (most people should be more familiar with the former).
The former is said to be the first weather forecasting software in China to accurately determine the rain time by a few minutes, thanks to their algorithmic craftsmanship in the field of natural language processing, a branch of artificial intelligence technology;
Screenshot from the mobile APP
The latter is a foreign language translation plug-in artifact that a journalist friend strongly recommended that I "must download" on Google Chrome.
However, in the global huge mobile phone application market, these two products can not be regarded as the ultimate success, but at least "small and beautiful". This also ensures that even in the past 5 years when the business potential of artificial intelligence has fluctuated, the company can have a sustained and relatively stable revenue revenue.
However, it is clear that both software are relatively favored "productivity tool-based applications" in the market. That is, you can have a longer and persistent retention because you "have to use it";
And this time," "Caiyun Little Dream", for the time being, is more like an algorithm program that "can only make young people laugh" and "the value of literature and art is much higher than the commercial value", and this is one of my biggest doubts.
"There are reporters who sniff tigers to visit today" is a sentence I output
The hardest fortress cannot be conquered
From another point of view, it is precisely because of the weak association between this set of algorithms and "commercialization" that Yuan Xingyuan should be one of the most sincere technology entrepreneurs I have ever met who have the most sincere feelings about artificial intelligence technology.
Compared with image recognition and face recognition, which have been "rolled out of the flowers" in the industrial application industry at this stage, another important branch of artificial intelligence, "natural language processing", the fundamental reason why it cannot be applied on a large scale today is because it is unable to make a "code-level description" of a large number of abstract concepts, semantics and logical reasoning in the real world.
For example, two illiterate people can't read, but the two of them can chat completely without obstacles.
In this conversation, "200 times" is more like an adjective or an exclamation to express the mood, but from a semantic point of view, the machine is difficult to understand.
In recent years, the rise of deep learning has suddenly solved the problem of computer vision (CV) "using manual extraction of features", so it has brought explosive progress in the field of CV.
However, in essence, even computer vision has only made a breakthrough at the perceptual level, and has not made much breakthrough in the next layer, semantic and logical reasoning. That's why "planning decisions" for autonomous driving are so difficult.
Compared with intuitive images, it is necessary to go beyond "perception", and require extremely high natural language processing for "high-level semantics, memory, knowledge abstraction and logical reasoning" (image recognition is also needed, but it is better to commercialize in the perception stage), which is obviously more difficult and has a stronger sense of achievement.
Conversations from online Siri
You might be wondering, how can you say that "natural language processing" has no large-scale application? Apple's Siri, smart speakers, machine translation, and algorithmic recommendations hidden behind various e-commerce and social platforms are all forms of this technology.
But in reality, these are the first fruits that are best harvested at the foot of the "natural language processing" mountain.
Just like smart speakers, whether they sell 50 million units or 100 million units, they are still titled "mentally retarded". Because it still can't "understand the logic behind the world" – this is the most valuable medicinal material on the top of the mountain.
"Except for writing, nothing else is worth studying by artificial intelligence!" Yuan Xingyuan was resolute. He made the project more like an algorithm engineer or scientist, wrestling with the market and technology.
"It's like 'writing essay' is the last question on the language test paper. It's harder than playing Go, than autopilot and face recognition! Because it is difficult, it has the value of doing!
Because image data is high-density, low-information data, while text is high-density and at the same time highly abstract data, the data space behind it is much larger than we think. ”
Of course, Yuan Xingyuan is not the only one in the industry who holds the same "ideal". Foreign countries not only started earlier, but also showed their strength to the outside world.
Microsoft CEO Nadella (left) and Open AI CEO Atman (right) at the 2021 Microsoft Build Conference
In May 2020, OpenAI launched GPT-3, known as the "World's Most Advanced Language Model." The AI research organization, which was backed by Elon Musk and other Silicon Valley moguls in 2015, officially embarked on the road to commercialization the following year after partnering with Musk Say Goodbye in 2018 and immediately accepted $1 billion from Microsoft.
Today, the new generation of language generator GPT-3 is sold as a cloud service to institutions that want to use it to automatically renew text. Needless to say, it's built on natural language processing technology—
Nearly 1 trillion words of text were collected on the Web and trained on a supercomputer embedded with hundreds of thousands of processors. By the way, this computer was made by Microsoft, which also paid back some of the $1 billion invested.
Interestingly, the Guardian once wrote an op-ed with GPT-3 titled "Are You Scared?" Humanity". The general central idea is "Although I am a thinking robot, don't be afraid, I will not destroy you, I am a community of human destiny."
Excerpted from the British Guardian
At the same time, however, I think a small portion of Microsoft's $1 billion investment will probably be lost.
Because two years after its release, this language generator model, although it has made a major breakthrough in technology, has even been called "one of the most important and interesting AI systems in history" by Australian philosopher and cognitive scientist David Chalmers; however, it has not won the favor of many users in the corporate world.
In addition to enterprises for research and training purposes, there are only some educational auxiliary jobs that "play a role in text fine-tuning".
Although I have not been exposed to GPT-3, but also as another form of "text generator", since the key basic technology has not yet made a breakthrough, then Caiyun's story continuation algorithm, in fact, is not essentially out of the category of "will be stupid".
But it did grow a few years.
First of all, although in the continued paragraphs, it can not only reproduce the pattern of the text, but also perform very normally logically, even better than some online writers who write patchwork hydrographics (the picture below, the performance is indeed good). But because of the lack of common sense understanding of the world, it will say "Trump is in love with Macron" without thinking. Perhaps this is a big reason why engineers define their attributes as "writing novels."
Second, it is still limited by the intellectual properties of the original training text. The algorithm is trained by 30,000 online novels, so when I enter news reports and non-fiction stories in the technology industry, the content I continue to write will be wrong.
Third, a lot of times, I find that it keeps saying the right crap (pictured below). That is to say, ai is constantly imitating and copying the text you output, and replacing the original sentence several times with synonymous sentence meaning.
AI is better at crap literature
Fourth, the first few points determine that it is not yet possible to tell a story that is super long, complete, and logically structured. So the program can only limit the number of words you can input and the number of words it can output.
Interestingly, when Yuan Xingyuan talked about the reason for this project, he repeatedly mentioned "experiment", "research" and the more ambitious theme of "pushing human civilization to the next stage", and his reverence for "The Three-Body Problem" was overflowing.
Although as an ordinary science and technology editor, I am far from being able to share the imagination of engineers, and can only listen to the other party's obsession with "the future of light-speed spaceships described in science fiction "Three-Body" is the goal of contemporary science", from echoing to gradually falling silent.
But I am well aware that solving the problems at hand is equally important.
Obviously, in addition to the bottleneck encountered by basic science, Caiyun will definitely not avoid the commercialization problems encountered by GPT-3.
Let young people pay, can't rely on fun
Although Yuan Xingyuan optimistically told me that the current project is indeed "doing public welfare with a happy heart", completely free, only hope that young people can enjoy the temporary happiness brought by the algorithm. But as an entrepreneur and business manager, he must consider designing a paid solution for China's C-end youth market.
At present, their ideas about this, in my opinion, there are still large gaps - they do not have enough product power.
First of all, the simplest and most brutal commercialization model is to let users pay for "continuation".
However, according to the company's team's preset 100 yuan / month pricing, hundreds of thousands of daily active users, and the premise of "taking online literary authors as target users", whether the revenue space of this product can exceed his first two applications may cause some doubts.
We found the "2019 China Internet Literature Development Report" released by the China Audiovisual and Digital Publishing Association in 2020. It is clearly mentioned that although the number of online literature authors in 2019 reached 19.36 million, the number of signed authors was 770,000, and the average monthly income of online authors was 5133.7 yuan.
Of course, we can't use TAM (the total amount of a product in the market) as the market size that this continuation software can cover, but should circle its SOM (the market that has the ability to get it).
According to the current entertainment attributes of the program, public popularity, limited length of continuation paragraphs, poor ability to build the core of the story, and "original works are more easily recognized in the literary field" (although the plagiarism problem is serious), it will be used as a productivity tool product market coverage, which needs to be further verified.
In November 2021, the data of online writers released by Reading Articles, of which the last smoke is after 00
Second, be an intermediary docking platform for content producers and content consumers. Each user creates their own "story small world" inside. Yuan Xingyuan feels that content consumers should be willing to pay for their favorite story ideas.
He opens the backstage, and each user ID corresponds to the "small world" he created— the name of the protagonist of the story, the main task performed, and the main story line. Here, you can feel the imagination of young people: Qin Shi Huang's dialogue with Trump, Harry Potter vs. Contra, vampire love stories, interstellar city Xiu Xian...
"In 3 to 5 years, there will be a large number of user-contributed 'worlds' on our platform, and these people who set the world can share it, so he is the content producer on our platform;
And people who are interested in using these world settings, that is, our content consumers, pay the producers. I'm like making a game platform and extracting a certain rate. ”
At first glance, this may sound fine and is very similar to the business model of an open source platform. But on the one hand, as a content practitioner, I know that the domestic content payment industry is not good enough to be compared with code; on the other hand, I quickly found that the level of these user-built stories is uneven:
Someone constructed a complete storyline, but the storyline was not appealing;
Someone wrote halfway and left;
Some people can't even write the main line of the story clearly, and deliberately type it randomly. For example, someone only recorded a name that "the rice dumplings are delicious" and there was no follow-up...
What cannot be ignored is that data quality will not only affect the quality of algorithms, but also affect the sustainability of commercialization. And even if there are 1 million such "small worlds", consumers who have a demand for creative content will not pay for it.
Yuan Xingyuan agrees with my question that only the "boutique world" has the possibility of payment, and now it is more of a disorder. But at the very least, they insist they've seen a possibility that "young people are willing to give their time on this software."
From Dwarven Fortress, it is hailed as one of the most difficult and complex games in the world
The third is to make a very complete game.
If you think of Dream as a creative piece of software with a playful nature, it has some of the qualities of a game or metaverse product. And now moving closer to the game and the metaverse, indeed... Can attract the attention of investors.
Yuan Xingyuan feels that this algorithm, which has not yet been evolved, can be turned into a set of text versions of "Minecraft" in the future. The latter is a very famous Microsoft sandbox game, a bit like a virtual version of Lego. Its biggest feature is that it allows players to collect various resources in a 3D world to freely create and customize their own world.
However, whether you want to make the next Minecraft or do Dwarf Fortress, no matter how many randomly generated "worlds" players encounter, in the game design and development process, it will not all be a disordered state that allows the AI to give random results.
For example, caves, temples, monster lairs, abandoned towers, these elements are already present in the game design script, but ultimately depends on how the player activates and obtains it, and uses these elements to create their own world.
Of course, the more important point is that the application of a purely text-based algorithm evolves into a sandbox game of 2D or 3D graphics + text for gamers around the world, and there is a huge experience and cost gap in between.
The game market is fiercely competitive, even laymen like me know a little bit:
There are tens of thousands of games in development every year, dozens of games are being tested every day, and those that can survive may only be hundreds of them every year, and the only ones that have caught fire are the ones we have heard;
In addition, in addition to forming a game development team of at least 5 people, companies also have to bear the huge capital required for operation and distribution. These are all things that an application software company that made its fortune as a "natural language processing technology" is difficult to afford. At least during the conversation, I didn't hear the specific planning details about the team.
Of course, from another point of view, since natural language processing technology relies on "text", caiyun will inevitably not be able to jump out of the business model set by "text" - so, from "caiyun weather", "caiyun small translation", and then to "caiyun small dream", in essence, there is no jump out of the "text world" of the five finger mountain.
The existing consumer-side megatrend is a business model that is comprehensively constructed by natural language, images, videos, and metaverse-related technologies; is the pure text world a kind of reverse?
Image from Visual China
Creating a simple text world, this makes me have a vague sense of déjà vu: in the past 6 years, many artificial intelligence companies have tried to use single-dimensional technology to open the door to competitions that require comprehensive quality to win steadily.
But one thing, I must admit that Yuan Xingyuan's inspiration for using AI to make games is well documented – he sees a game industry that has existed for many years:
Use words and images as mere shells of the game. Most games seek to draw everyone's hair very finely, but the words they say are very mentally retarded.
"Probably 90 percent of the development cost is on top of that. Then it's all about me punching you and you don't move; you punch me, and I don't move. The world revolves around an inherent rule, and the game's graphics are far more intelligent. This needs to be changed by the improvement of natural language processing technology. ”
Therefore, in the ideas, questions, rebuttals and acquiescences raised by each other, we are constantly looking for a more feasible business intersection that is not all supported by imagination.
Perhaps, providing api technology output for more comprehensive game studios and joint operations is more suitable for "small and beautiful" companies.
More importantly, the business model of exporting API technology to other enterprises has been effectively verified on the first successful application they launched, Caiyun Weather: 90% of the application software with "minute-level precipitation forecasting" on the market is the relevant underlying technology provided by Caiyun Weather.
For example, the takeaway software commonly used by our workers is "Hungry", every time you see the page where the takeaway brother delivers in rain mode on a rainy day, you use caiyun technology.
As for such a single-dimensional product and technology, whether it has encountered the encirclement of large factories and the imitation of competitors. Of course, a weather app that ranks first on the market today "mimics" Caiyun and follows with the release of the same product. This made Yuan Xingyuan feel both angry and helpless.
During Yuan Xingyuan's roadshow in 2017, an investor stood up and demolished the stage with the same question:
"No matter how good your algorithm is, I can't spend tens of millions to do better than you, won't I just kill you?"
And Yuan Xingyuan's answer is that there is a technical man's temper and dignity:
"Then you will compare yourself with me!"
In fact, according to my observations in recent years, although the comprehensive strength of the big factory is extremely strong, it seems that there are too many "children who need to be favored", and they cannot always support it in many subdivided vertical fields.
For example, a game manufacturer, its AI, automatic driving or medical, semiconductor projects, said no, said no, said no, do not want.
Therefore, for the segmentation of technology "take-ism" applies to large factories; for heavier product models, small and beautiful software technology startups unless they have the courage to make a desperate bet, the support of capital and strong execution, I have only seen two endings:
The mouth cannon related to "technology changes human beings and lives" is jingling, but it is unwilling to do dirty work and tired work, and has evolved into a "content creation company" or real estate company that eats old books and patrons;
Shut down the company, go back to school or go home to inherit the family business.
Written at the end: the possibility of absurdity
Although in the course of the exchange, I have repeatedly raised doubts and concerns about the commercialization of this program, my confidence in the technology held by Yuan Xingyuan has not changed.
This comes first of all from the good feelings for Caiyun Xiaoyi: as a text worker who balds every day, this 12 yuan per month translation plug-in is convenient enough to give me some courage to open difficult and obscure foreign language technical articles.
Second, it also comes from my confidence in the nature of a group of Internet software technology people in China: whether in the free market or not, engineers naturally seem to kneel to absolute strength - although the expression always does not change, but there are always a few dry him grand ambitions popping out of the mouth.
What does it feel like?
Just like when we watch the Winter Olympics, why do we feel admiration for the "absolute strength" of Famous Players such as Wang Mao, and why do we feel heartache for the garbage strength of Chinese football that "cannot be shaken like a mountain"?
In the world of technology, everyone wants to challenge the hardest thing, pick the fruit at the highest point, and finally use this "absolute strength" with the technical dimension as the only criterion to break out of the commercial siege.
However, we do need to make good technology have a good shell and a reliable application method.
If one rots, the others will rot very quickly
In fact, what worries me the most is not the bottleneck of natural language processing technology and the "no money" of AI story continuation.
Instead, often in the process of human beings looking for business models, in the end, the most easily commercialized successful models excavated are precisely some "crooked evil ways" derived from the development of normal industries.
For example, face recognition technology from the paper into the access control and gate, the whole process has never been mentioned, it can be used by real estate agents to kill, identify old customers to raise hundreds of thousands of house prices.
But this kind of application not only happens, but if you think about it, this is precisely the business path that 2B technology is most easily paid for.
And what happens when the AI story continues?
It may be used in some side-door businesses in the education market (which I didn't want to write about). For example, "essay writing", "homework ghostwriting", "exam cheating", and even exacerbating the probability of "plagiarism of works".
Among them, there has been some heated discussion in foreign academic circles about the negative behavior of students triggered by GPT-3.
But from the perspective of pure copper stink, aren't these more rigid than "entertainment" and "game development"?
I hope that my worries and worries will not become a reality because of the outbreak of natural language processing technology after the next breakthrough bottleneck.