SuperApp's Future Outlook: How Andrew Ng predicts how communities can coexist with AI

Andrew Ng: We need community. People who have friends and allies will perform better than those who go it alone. Even though the field of AI brings breakthroughs every week, having friends to help you tell the truth from the hype, test your ideas, support each other, and co-create with them will put you in a better position.

At the end of the previous article (Non-consensus Interpretation: Perplexity.AI vs Tiangong AI), Xiaobian threw out the knowledge search ability based on large models, and built a SuperApp product idea from "tool value (low premium)→ process value (low premium)→ knowledge value (high premium), → network value (high premium)" in the domestic C-end market. Due to the length of the previous article, I did not share this product idea. In the past two days, I have seen the importance of the "AI community" mentioned in Mr. Ng's "2024 Prediction: About AI, These Things Will Not Change in the Next Ten Years". The editor also continued the last topic of the previous part, "SuperApp: Knowledge Sharing & Trading".

This article will focus on the following questions to think about and share this product idea.

• In what track, knowledge quizzes are just needed?
• In this track, what are the pain points of users?
• What can large model capabilities do in these pain points?
• If you want to solve these pain points, what is needed, and where are the stuck points?

Before starting, the editor reiterates the two non-consensus cognitions shared in the previous article:

• In the domestic market, SuperApp based on the search capability of large models must be free for C-end consumer users.
• In the domestic C-end market, the product value of SuperApp based on the search capability of large models is not only the value of tools, but also the high premium realization brought by resource matching by meeting the needs of multiple parties.

1. In what track, knowledge quiz is just needed?

1. In China, what track is the demand for knowledge quizzes?

Whether you're starting a business or building a new app, you must first take advantage of it. Today, many applications take advantage of the momentum of AI 2.0 to break the circle or even stand still. For example, Miao Ya Camera, Tongyi Tingwu, Midjourney. In addition to AI2.0 being popular this year, what other tracks have exploded in the past two years?

1. Knowledge payment: The scale of China's knowledge payment market will reach 112.65 billion yuan in 2022 and 280.88 billion yuan in 2025.
2. Super individuals: In 2022, there will be about 200 million flexible employees in China (750 million total employed population), and 18.6% of college graduates in 2022 chose freelancing, an increase of 3 percentage points from last year.

In the post-epidemic era and the downturn in the job market, the mentality of migrant workers has undergone a huge change. Many post-85s, post-90s and even post-95s workers have begun to realize that they only have one main business, and it may be difficult to guarantee a stable tomorrow in the current highly changing market. A new term has been mentioned and followed by many people, "super individual"!

The editor is more concerned about the knowledge super individual with the highest degree of relevance to the Internet among the super individuals, commonly known as: knowledge IP.

Knowledge Super-Individual:

It usually refers to individuals who have deep expertise and skills in a certain field and are able to monetize business through knowledge payment. Not only are they proficient in one or more professional skills, but they are able to apply those skills to real-life and work in order to monetize their business.

You may wonder, "Q&A service scenario of knowledge IP" and "Q&A scenario of AI search engine" don't seem to be the same kind of scenario, right?

That's right, the Q&A scenario of the head knowledge IP, the current mainstream scenario is the Q&A service on the private domain platform through live broadcast, or 1-to-1 service.

However, Xiaobian does not focus on the head knowledge IP, because what Xiaobian is thinking about: it is a C-side-oriented AI native application idea, so we must pay attention to the coverage breadth of its target users.

Borrowing from Midjourney's product idea (using AI capabilities to assist UGC users to transform them into PUGC and elevate a small number of users to the level of PGC).

So what I really focus on is:

• Users who are already intellectual IPs, but are at waist level or below;
• Users who want to open the second growth curve for their personal workplace by becoming a knowledge IP.

Both types of users have two core and common pain points: self-improvement & precise linking

Why do you say that?

Supply-side user demand:

• Precise link: Reach consumers in demand through major traffic platforms, empower knowledge, and gradually establish their own influence in the field.
• Self-improvement: Through constantly breaking the circle and socializing upwards, improve your vision and cognition, repeatedly polish your skills, and improve your ability to realize knowledge.

Consumer-side user needs:

• Self-improvement: Based on major traffic platforms, we can screen out knowledge IPs that meet our own knowledge needs, and obtain knowledge through payment and other methods to improve our competitiveness in the workplace.
• Precise linking: Through the strategy of upward socialization and horizontal circle breaking, we can obtain information gaps and knowledge and improve personal ability.

If you want to give an example, the recent "how to cross over to AI 2.0 to be a product manager" can be said to be one of the hottest topics in the workplace.

As a lucky cross-border success, Xiaobian has been thinking about two core points every day from the beginning of cross-border to today:

• How can I ensure that I am in a correct Xi posture and obtain information that is valuable to me?
• How can I accurately reach the network resources that are valuable to me, instead of people who want to cut leeks?

2. Why is there a chance for SuperApp to appear in this track?

Wang Xiaochuan: People need to have three things: one is to be creative, the second is to be healthy, and the third is to be happy.

Wanchen Moonshot, public account: Geek Park "Dialogue with Wang Xiaochuan: The core of large-scale model entrepreneurship is to think about how technology matches products"

2.1. Emotional level:

After conceiving this product, Xiaobian has been asking himself "What fundamental needs does this product solve?" After seeing this sentence shared by Mr. Wang Xiaochuan, Xiaobian has some answers. The improvement of users' self-ability after acquiring knowledge will bring happiness and a sense of accomplishment to themselves. And when users empower other people or things, it is also a sign of creativity.

When an app can meet two of the three basic needs of users, the app has its value.

The last generation of products called super apps were basically "link" type products, whether they were social with acquaintances or strangers, and the first task for users to use these products was to get more links.

We later discovered that the new generation of super apps may be more about providing users with a playground in which each user can create and generate their own things.

Plan Z, Official Account: Rustic Speech "Infinite Imagination of SuperApp & Limited Ability of Large Models|Z Salon Issue 4"

2.2. Rational level:

According to the speech of a VC partner in the "Z Salon of Mass Spectrometry Speech", the core value of super applications in the previous generation is "link", and the core value of this generation is "creativity". Whether it's "link" or "creativity", super individuals meet their criteria.

Knowledge super individuals have a "link" with consumers through "knowledge", and in the process, through the continuous understanding and application of "knowledge" (creativity), value empowerment is generated.

The technical boundaries of the current large model (hallucination, lack of explainability, catastrophic forgetting, etc.) can be cleverly avoided through scene segmentation or optimized to a certain extent based on the assistance of user data in the construction of Xiaobian's products.

2. What are the pain points of users in this track?

When it comes to user pain points, it is necessary to mention user scenarios. In the direction of personal self-improvement, which is the user's demand. What does the user scenario look like?

After seeing what the editor mentioned in the previous article: "Data Acquisition → Information Differentiation → Knowledge Classification → Skills Summary → Knowledge Monetization", you may have doubts? "Knowledge Classification → Knowledge Monetization It is very easy to understand, so what is the difference between data acquisition and information differentiation? Isn't it usually all about learning knowledge Xi and improving skills through practice?"

Here, the editor is going to talk about the Plus version of the smart model mentioned by the editor in a sharing a long time ago~

In the past ten years, many friends who have entered the Internet industry have quickly obtained the mature methodology (knowledge & skills) summarized by industry seniors through reading books, literature, and accelerated training when they first entered the workplace, and directly used them in the work scene. Over time, the Xi of "teaching people to fish is better than teaching them to fish" has been cultivated.

In the past two years, the job market has been sluggish, AI 2.0 has exploded, and the rise of super individuals has caused some readers to start working hard to improve their personal skills. At this time, I found that the methods of my predecessors were not so easy to use in my own hands. Therefore, I began to pay frantic attention to all kinds of knowledge IPs and join various skill sharing communities, hoping to improve my competitiveness in the workplace.

Scenario 1: Data acquisition → information differentiation

Therefore, the first segmented user scenario has "data acquisition → information differentiation":

Scenario examples:

As a friend who is trying to improve personal cognition and workplace competitiveness, do you see a lot of articles on different channels (WeChat official account, 36 Krypton, Tiger Sniff, Knowledge Planet, etc.) every day, but due to problems such as time, inertia, and lack of tools, there will be:

• Sometimes I can only take a brief look at → Collections → I have time to look at it again → I can't remember this thing at all / There are too many collections, which one is coming;
• Sometimes, after reading it carefully, → bookmarked→ and I had time to write a note→ I didn't have time to write at all;
• Sometimes I look at it and suddenly think of an inspiration → collection→ and write down this idea for a while→ what is the idea after I finish it;

Are these pain points rigid enough?Are they common enough?Before AI 2.0, these problems could not be achieved through product capabilities due to technical capabilities.

Friends who may be passionate enough about AI tools will think of Evernote's Elephant AI, Zhiliao Reading, Tiangong AI Assistant AI Reading and other products currently provide application apps for this kind of needs.

But after the experience, the editor summarized 3 problems to be solved:

1. The scene is not accurate: for example, the AI reading of Tiangong AI assistant, the main scene of the user is WeChat official account, 36Kr and other APPs, you actually asked me to manually copy → jump out of the current APP → open Tiangong AI → open AI reading → paste the link;
2. Paid to reduce retention: The product is limited by the impact of marginal cost (computing power), and users cannot use a large number of tools for free, such as knowing and reading, and the free quota is only enough for two or three content summaries, which is for C-end users, and it is difficult to grow without free;
3. Technology leads to poor experience: The product is limited by computing resources and technical capabilities, and the AI summary experience is poor, such as Elephant AI, when using AI to read, it either collapses, or waits for 1 minute to appear before the content with poor summary quality appears;

User pain points:

• Valuable data (articles/videos) manually judged by users need to be quickly "stored" to the "personal information database";
• Interested data (articles/videos) that are manually judged by users need to be able to be quickly "identified" to produce valuable "information";

Product Value:

• User traffic entrance: the link origin of the user's knowledge Xi & self-ability improvement;
• Private domain data acquisition: Currently, large model search can only be obtained through public domain data to solve user problems. It is difficult to obtain private domain data such as public accounts.
• The premise of better user experience: it is easier to get the user's preference information, which provides value for subsequent user stickiness (hot recommendation) and transaction matching (creative preference on the content supply side).

Scenario 2: Information Classification → Knowledge Classification

If the user pain points in the first scenario are solved, what will you, as a reader, do next?

User Scenario:

This "personal database" may be used when you have time to study Xi or when there is an emergency that requires you to consult your knowledge base (or material library).

• I'm going to start writing and need some information to refer to.
• I was communicating with friends in the community, and I needed to add an idea that I remembered, and I remember seeing a paragraph about it.
• I'm going to do an AI comic video now, and I'll share tips on AI output of continuous related footage, where I remember watching it.

This involves the transformation of "personal information database → personal knowledge base", that is, "information distinction → knowledge classification".

User pain points:

• Users can quickly find the "information" that is targeted in the "personal information database" through natural language or business tags.
• Based on the "information" that has been found, users can quickly find relevant information in the public domain and the private domain of the product for knowledge completion.
• Users can quickly summarize the "information" used this time and supplement it to the corresponding knowledge points of the "personal knowledge base";

Product Value:

• Reduced user churn rate: In the previous scenario, users retained a large amount of personal data, and in this scenario, AI capabilities + product strategies can further increase the cost of silence and reduce the probability of churn.
• Improvement of product recognition: because the product can help users sort out their knowledge, improve their cognition, and make users feel the improvement of personal creativity, the product can also let users find "confidants" and make users feel happy through the "Xueyou" mechanism at this stage;

Scenario 3: Knowledge classification → skill summary

The biggest difference between this scenario and the first two scenarios is that this scenario is an offline scenario for users. Because "knowledge classification→ skill summary" and then detailed dismantling is actually "classification→ trial and error, → practice, → summary, → execution→ review→ review, → review".

Conduct trial and error on knowledge points, eliminate irrelevant knowledge, practice relevant knowledge, summarize cognitive biases through practice, and continuously polish and iterate skills through PDCA strategies.

There are two main pain points for users:

1. Online tools: handy sorting tools → text creation, version management, logic combing;
2. Offline action: find practical application scenarios → practice to complete experience accumulation or learn from the practical experience of others;

In this link, it is necessary to adopt the strategy of building on strengths and avoiding weaknesses. Because at the tool level, there are applications with high user stickiness in various vertical directions: such as Feishu Docs, Xmind, Graphite, and Evernote. Migrating users from these apps to my product for content creation is a very cost-effective (user migration probability / app development cost) event.

Therefore, in this scenario, product tonality is not the creation of native content by users in the product. Instead, let the user re-enter this content data into the product after the content creation is completed by the third-party tool.

At this point, you ask: why do users do so much?

Because before AI 2.0, although various text authoring tools provided the ability of global search in the product, the problem that was not essentially solved was "based on an information point or knowledge point, the content involved is associated to form a knowledge graph." ”

The product I conceived can quickly learn Xi the large model, plus the knowledge graph and other in-depth learning Xi technologies, to complete the construction of the user's personal "skill tree".

Just ask: If there is a product that can connect your usual scattered thoughts, ideas, and summarized content through mind maps, etc., and your operation cost is only to quickly import these contents, and simply proofread the labels after AI marking, will you use it?

What do you mean by the "skill tree" mentioned here? In layman's terms, it is a mind map. For example, the skill of "long text to video", which is currently popular. Although in the AI 2.0 era, AI video generation has greatly improved the efficiency of creators. But the skill of "script → video" has been around for a long time. So this "skill tree" itself is not a new product. The knowledge points on this skill tree are simplified only through AI capabilities. But if you're a business master of "AI video generation", don't you know how traditional processes can work, and how can you empower others with knowledge?

Product Value:

• Reduction of the hallucination probability of large models: Through the "skill tree" sorted by users, the accuracy of large models' answers to a certain type of directional questions can be further improved and hallucination bias can be reduced.
• The value of the product in the second scenario is further enhanced.

PS: In this scenario, a friend of the editor once put forward a point of view: "The more standardized the content, the less valuable it is for AI integration." ”

After I talked to this friend, he wanted to express the idea that "the more standardized the knowledge, the less valuable it is for AI integration." The editor agrees with this point of view, but what the editor wants to express here is that "knowledge has standards, but skills are not." It's like saying, "The registration and login module of an APP has a function definition, but there is no fixed answer to how to design it to make the user experience better." ”

Therefore, Xiaobian's "skill tree" strategy is to make up for the cognitive gap of users' "knowledge", and gradually improve the ability ceiling of a certain "skill" for users.

Scenario 4: Skill summary → Knowledge monetization

First of all, I would like to emphasize that the editor here describes "knowledge monetization", not "knowledge monetization", and knowledge monetization focuses on the economic aspect. Profit can be understood as: self-satisfaction of self-spirit, effective expansion of network resources, and significant improvement of economic income;

User Scenario:

• Waist knowledge IP: It has a skill, but due to its lack of industry influence and reputation, it cannot reach accurate consumer users.
• Consumer users: There is a clear need for knowledge acquisition, but due to the difficulty in identifying the service capabilities of knowledge IP, they are often cut leeks and cannot obtain valuable knowledge information relatively fairly.

User pain points:

In the current market, only a small number of leading knowledge IPs or institutions can achieve a significant increase in economic income through knowledge payment.

What is the cause of this problem?

• Head knowledge IP: The "authority effect" brought to ordinary users leads to the majority of users relying too much on the authority of head knowledge IP due to "security psychology", "herd mentality" and "cognitive dissonance".
• Knowledge IP of the waist and below: Due to the lack of personal influence, consumers cannot understand whether the "waist IP" can meet the knowledge needs of consumers through word-of-mouth.

In other words: "As a seeker of knowledge, before I need to pay to obtain the knowledge service of a certain knowledge IP, the biggest consideration is whether the money is worth it? The traditional way to judge this matter can only be to inquire about his reputation, or look at his past sharing and students' feedback.

In this product, Xi all the knowledge demanders can become knowledge suppliers (three-person lines must have my teacher), because from the process of "data acquisition → skill summary", although the large model is providing services to users, in turn, the large model can also judge the user's ability and degree of expertise through these data.

This is something that was almost impossible for internet applications to do before AI 2.0.

When the large model can calculate the competitive advantages and disadvantages of users relatively clearly through data and business tags:

• For the part of the skill shortcoming, you can introduce third-party education and training institutions and other users who are higher than the user's ability to conduct transaction matching.
• For the skill longboard part, it can be channeled to the field of enterprise recruitment.

Product Value:

Monetization method: Bring actual economic value to the product by completing the resource matching service at both ends of the supply and demand ends of knowledge payment.

3. What can be done in these pain points?

In the previous section, in fact, each scenario mentions the large model capability, and it can also be said that the entire product is built based on the large model capability. In other words: without the ability of large models, this product cannot be implemented.

Here is a brief description of several core functions that rely on the capabilities of large models in user scenarios:

3.1.AI Article Summary (Scenario 1 & 3)

I believe that friends who have used Evernote's Elephant AI, Zhiliao Reading, Tiangong AI Assistant AI Reading and other products know that the core function of its products is the AI summary (AI summary) capability. In the scenarios of information differentiation, knowledge classification, and skill summary, the product capability based on the large model is very important.

At present, in the design of this function of these companies, Xiaobian believes that this function may not be done well due to computing power cost, technical capabilities, scenario analysis and other reasons. Regarding the detailed analysis of this function, the editor has elaborated on the "AI speed reading" and "AI intensive reading" scenarios in the "Tiangong AI" chapter in the previous sharing, so I will not repeat it here.

3.2.AI Knowledge Search (Scenario 1-3)

Whether it is overseas ChatGPT and PerplexityAI, or domestic Wenxin Yiyan and Baichuan, etc., they all provide AI knowledge search engine services. In Xiaobian's product conception, AI intensive reading in information differentiation and knowledge completion in knowledge classification & skill summary all require AI knowledge search capabilities.

In this function, the user data (hot information, knowledge classification, and skill tree nodes) precipitated by the product itself can be reversely empowered by knowledge search based on large models, reducing the probability of problems such as hallucination, lack of explainability, and catastrophic forgetting. This is also a problem encountered by current AI knowledge search engines.

3.3.AI Marking of Information (Scenario 1-4)

Data marking through AI technology has been around since the AI 1.0 era. However, even the most authoritative medical knowledge graph (SNOMED-CT) cannot cover 100% of all medical Q&A scenarios.

Therefore, in the products built by Xiaobian, AI information marking is not a substitute for users to complete the data classification of "information→ knowledge → skills". Instead, AI marking is used to screen the huge and messy data first, and the user manually corrects it based on the results after the initial screening.

Here I would like to put forward two non-consensus cognitions:

• AI can replace users to complete some extremely standardized workflows, but it cannot replace users to complete non-standardized work tasks.
• When AI can replace users with 100% of complex tasks in a certain field, such users will not be competitive in this field.

3.4.AI Matching Alumni (Scene 2-4)

For the vast majority of ordinary people, due to workplace pressure, they need to learn and Xi out of their comfort zone and complete self-improvement. It is not a joyful process and can also be seen as a pain.

Therefore, in this product built by Xiaobian, in addition to providing users with a full-link service from "information→ knowledge → skills" through tools, another core capability is the service of matching "Xueyou" through AI. Let users link to the right mentors and friends in the process of self-improvement, reduce user churn, and build knowledge value & network value for the product.

"Xueyou" strategy: Xiaobian positions it as a precise link between people through knowledge needs. Non-traditional meanings are vaguely linked through topic communities/communities.

Here you may have two questions:

1. Why not link people in the same way as topic communities?
2. If the precise link between people is carried out through knowledge needs, how to solve the problem of low matching caused by demand personalization?

For the first question, the editor explained through a phenomenon: when the editor was Xi with AIGC 2.0, he joined 10+ relevant private domain communities, but the difference in obtaining valuable information was not as much as what I obtained by watching in-depth literature/videos.

For the second question, I believe that we should first pull together a cognition: what is demand personalization, and what is its granularity?

In the field of knowledge empowerment, the needs of users are not ever-changing and untraceable.

The user's needs come from the industry, the position, the skills required, etc. The individualization of the user's need for "knowledge" is caused by the cognitive bias of the results summarized after the synthesis of the above dimensions.

The product can use AI technology to calculate the user's current knowledge needs based on the user's personal information, preferences, technology tree and other data, and match them with "Xueyou" at the appropriate node of the user's path. Of course, there must be a margin of error for this match. However, the cost brought by this function to users is similar or better than that of users finding "Xueyou" through traffic platforms and private domain communities, but the cost is lower. Users will use this feature.

4. What does the team need to build this product?

From the above analysis, it is not difficult to see that the entire product is built based on large model capabilities, and each scenario involves a large number of large model calls and fine-tuning of model capabilities for directional scenarios. Therefore, the first important premise of building this product is:

1) The team itself needs to have the ability to develop large models.

Since the product is oriented to the domestic C-end market, it is difficult to bring growth to the strategy of charging users. However, if the API service of a large model provider is used, ordinary application factories cannot afford the cost of computing power in the short and medium term. Therefore, the second important premise of building this product is:

2) The team itself needs to have computing resources.

Therefore, if you want to achieve such a product construction, in addition to the Internet manufacturers with the ability to self-develop large models, they are model factories.

For model factories, the value that such a product can bring in the short to medium term:

• In scenario 1, users can actively input "directional domain & manual judgment valuable" data for the large model by quickly cutting into the user's pain points, and manually annotate the output data based on the summary of the large model, so that the large model can obtain a large number of valuable and well-labeled data through the way of "white prostitution", so as to improve its own large model service capabilities.
• When the full-link construction is completed, the business model oriented to the C-end will change from a tool service subscription system to a matching transaction. (Low premium tools → high premium knowledge & connections paid)

Whether it is Wenxin Yiyan, Tongyi Qianwen, or Baichuan model, the dark side of the moon and other model factories, they all provide information search services for domestic C-end users, but in addition to releasing the signal that the team is paying attention to the "domestic C-end large model application ability", what is the value brought to the enterprise itself? Is it expected that users will take the initiative to mark the wrong answers? Or do you expect to analyze valuable user scenarios through extensive use by users? Xiaobian has a shallow understanding of this issue, so it will no longer be a class axe.

That's all for this sharing, the full text does not mention the business model and monetization methods, which is not something that the editor did not consider. It's just that due to space issues, I won't expand it this time. To put it simply, when users become sticky to the product, they can start to "sell" through subscriptions, matching seats, and consumer traffic monetization. As for how to sell, the editor has also conceived the whole case, interested friends, you can chat with the editor privately.

Refer to the article

Dialogue with Wang Xiaochuan: The core of large-scale model entrepreneurship is to think about how to match technology with products

Infinite imagination of SuperApp & limited ability of large models|Z Salon Issue 4

iiMedia Consulting|2022-2023 China Knowledge Payment Industry Research and Consumer Behavior Analysis Report

Thankfully, this is the age of super-individuals

Columnist

Yang Sanji, WeChat public account: Yang Sanji, everyone is a product manager columnist. Senior product officer with 8 years of Internet experience, deeply engaged in the content field, ex Ali AIGC.PM, and now AI2.0 PM, a leading enterprise in a vertical field.

This article was originally published on Everyone is a Product Manager. Reproduction without the permission of the author is prohibited.

The title image is from Unsplash and is licensed under CC0.

The views in this article only represent the author's own, everyone is a product manager, and the platform only provides information storage space services.