Dismantling AI products from the perspective of users: Character & Tool

**Source: **3D Walkthrough

Author: Liang Jiasheng

Image source: Generated by Unbounded AI

What are the barriers to AI products? -- Talking about Jasper.ai layoffs

I woke up on Wednesday and saw the news that Jasper.ai had laid off staff.

"It's too fast, I thought I could hold on for a while" -- My friend and I both expressed emotion.

My attention to Jasper.ai originally stemmed from an article (Jasper.ai's $1.5 billion GPT shell? Is there a moat?). Jasper is a particularly interesting company, standing out among the crowd of AI upstarts: A group of young men, none of whom have a Ph.D, no dream of stars and oceans, and no determination to make a big brick, just want to start a company to make money -- Pragmatism. In the early days, I established a "marketing service distribution company". As the name suggests, it provides marketing services, but the method is to distribute to outsourcers--on the surface it is marketing services, but actually it is an agent, and its core competence is sales. I have to say, sometimes a company's DNA is written at birth (laughs)

The later development route was very Internet: doing AI marketing writing, relying on early community operations to win valuable test users, and then iterating step by step according to user preferences--Ren Xin said, this is stepping on the watermelon rind, where it counts where. All the attempts that were full of subjective will, they all failed; all the small step iterations that followed the trend were unexpectedly done well.

Among AI enthusiasts, I'm the type who pays too much attention to Jasper.ai. The reason is that I think it is an excellent example for me to establish a cognitive system for AI applications:

I think the core of the killing feature of ToC products in the AI era is: "high demand + barriers"; among them, "barriers" is the most difficult part to analyze at present. What kind of AI applications have barriers? What are the short-term/long-term barriers:

1. Is the interaction and interface design based on user process dismantling a barrier?
2. Is engineering for vertical scenarios and intent understanding a barrier?
3. Is the embedding solution for vertical scenarios a barrier?
4. Is only the ability of LLM a barrier?
5. Still, the ability of LLM is not a barrier, only privatized data + data flywheel + LLM fine tuning is a barrier?
6. Or, in the end, only powerful LLM + plug-in ecology + developer user community are the barriers?

After thinking about it layer by layer, the more I thought about it, the more chaotic it became, but there was no result.

Going back to the example of Jasper.ai, what did it do?

1. The ability to select the appropriate model for the task
2. Network real-time information, certain retrieval function
3. Overall: record customer information, preferences, styles and settings with sufficient timeliness
4. Templates and parts: marketing copywriting based on different platforms
5. Easier-to-use interface and interaction scheme based on workflow disassembly design
6. Ultimate user education and product guidance
7. Overwhelming talent marketing

I have to admit that what Jasper.ai has achieved so far is only a finished product of a technology company's 1-3 months of work: ChatGPT casing + workflow dismantling UI + engineering; Further increase in the window period (customer development and marketing). In the age of AI, this is really not a solid barrier. But I used to think that its workflow disassembly ability is valuable-this means that the team has a deep enough understanding of the "marketing copywriting" industry, which may be enough to support them to accumulate a batch of users over a period of time, and then build a data flywheel , and then build a stronger data barrier in the next step:

This set of business models: high gross profit eats industry dividends —> focus on marketing and promotion —> occupy users’ minds (budget) —> crack down on competing products/monopoly; in traditional industries, when the flywheel turns, it becomes invincible.

But it's different now, Jasper laid off staff this week - because the popularity of ChatGPT is equivalent to D2C brand direct sales, middlemen like Jasper have no way out.

A way of dismantling AI applications: Character & Tool

Recently, two of the most popular AI Apps have been launched (except ChatGPT). The domestic one is Wenxin Yiyan, and the foreign one is character.ai; both products are essentially "shelf-type" products -- different debugging (mainly Or engineering)’s LLM is placed in front of users through shelf display, allowing users to choose and use:

If you classify the AI on the shelf, there are actually two types: one is Character (corresponding to Figure 2 & 4), and the other is Tool (corresponding to Figure 1 & 3).

The core of Character is character, which provides emotional value and personalization; while the core of Tool is function, which provides non-standardized efficiency value. Emotions and personalities are ever-changing, so there will be many Characters--As of March, there are already 2.7 million characters on Character.ai; but the functions are limited, there are only a few major core needs, and most of the subdivided scenes are differentiated Commonality, so Tool focuses on quality rather than weight.

This classification is very interesting and very natural: Everyone says that LLM changes the human-computer interaction mode, turns the original Database and Code into Model, and turns GUI into LUI... But from the user's perspective, the product is actually From Tool —> Character + Tool -- the tasks I originally wanted to complete with the help of tools can now be handed over to a person. **

It happened that I recently read some articles (links) related to Tool learning, and saw the potential of AI as a brain to use Tool. At the same time, I remembered that I had discussed the relationship between LLM, Agent and Plugin with my friends before, and suddenly I started brainstorming and asking myself questions and answers. The following are all violent comments, welcome to correct me:**

For any AI product, it can be disassembled into Character and Tool:

●Tool is capability, characterized by stable output and high efficiency. It includes traditional tools (such as calculators, alarm clocks, weather forecasts, etc.) and new AI capabilities (such as Wenshengwen, Wenshengtu, search for understanding intentions, etc.), and of course some popular products (such as Midjourney) and some Plugins.

●Character is the role, which is the subject of the ability to use, and is characterized by having different "characters" -- similar to different people. The implementation method includes training different LLMs (such as GPT, Claude, and Infection-1, etc.), building different Agents, and debugging LLM and Agents through different data/different methods.

**Q: Does LLM have to be Character? **

A: Not necessarily. If a large model achieves perfect compression and decompression, then the main source of "character" becomes the Agent and command engineering around LLM, so LLM is Tool. Just like from the current point of view, ChatGPT is a Character, because compared to Infection-1, it is knowledgeable, indifferent, comprehensive, unfocused, and extremely rational; but in the long run, ChatGPT does not necessarily have the emotional ability of Infection-1.

**Q: Where are the barriers to AI products? **

A: In the short term, both Character and Tool can become barriers, and even Tool is the main barrier. But in the long run, the barriers to AI products are mainly in the Character, and the Tool will be received by the Character and used by the Character.

**Q: Do humans not use Tools? **

A: No, humans will use Tool, but it will be used under GUI interaction, which is similar to now. The Tool under LUI interaction will be received in Character, because LUI meets non-standard service requirements, people often not only pay attention to functionality when choosing services, but also consider the fit with the other party (such as finding a childcare wife, hiring a waiter, purchasing Consulting Services), so Character + Tool > Tool.

**Q: What are the differentiating abilities of Character? **

A: Character's ultimate form is a personal butler. Therefore, the primary ability is personalization, and the core of personalization is model memory, which can learn the user's personality and preferences if the memory is long enough. The second is a distinctive personality/three views, which are used to accumulate early users, similar to the "community tone" of contemporary Internet communities. Noam Shazeer, founder of Character.ai, said: "If you try to present a public character that everyone likes, then this character will definitely be boring."

The further barrier here is privatized data, which can build differentiated "roles". However, the advantage of long-term data will also be obliterated. After all, the concept of privatized data does not exist once the product is launched, so it is more important to build a data flywheel based on personality.

**Q: What are the differentiating capabilities of Tool? **

A: In the short term, the differentiation ability of Tool is based on the underlying technology, and the competition is hard technical ability. At present, Midjourney, Runway, and Adobe Firefly are examples; but in the long run, the tool products brought by technology generally have two endings--Technology Those with high cost will achieve monopoly (such as Microsoft office suite, Adobe suite, etc.), and those with low technical cost will become the basic capabilities of the bad street (such as timers, calculators, etc.), and there is little chance for a hundred schools of thought to contend.

**Q: If you want to make a ToC product, do you want to be a Character or a Tool? **

A: For most people, it is better to be a Character. Because the characters are more diverse, the demand is longer tail, more products can survive, and it is easier to hold the barrier after survival. To make a tool requires absolute hard technical ability and the ability to monopolize the market.

**Q: Finally, to answer, what is the long-term route for AI products? What conditions are required? **

1. Discover high-quality Characters: you need to have good insights into user needs, or a quick screening mechanism (this wave of Character.ai is in the atmosphere)
2. Model construction based on the target Character: a data screening mechanism (marking) for the target Character is required
3. Seize the window period to accumulate users/corpus: clear product positioning and marketing capabilities
4. Build a data flywheel that deepens Character differentiation: exquisite products <—> data design
5. Repeat the above cycle