The End of the Information Arbitrage Era#

In February 2023, 28-year-old David Zhang excitedly typed away in his small apartment. As a former Google engineer, he had just discovered a “secret” about ChatGPT—through specific prompt techniques, he could make AI generate high-quality marketing copy.

“This is a money-printing machine,” David said to his roommate, his eyes gleaming with gold-rush excitement. “I’m going to create a prompt store, $49 each, conservatively estimating I can sell 100 a month!”

Indeed, in the first three months, David’s business was surprisingly successful. His Etsy shop received over 500 orders, with monthly income approaching $25,000. He even began considering quitting his job to run this “passive income” business full-time.

But the good times didn’t last long.

Four months later, David’s order volume began to plummet. By the end of 2023, his monthly income was less than $2,000. What frustrated him more was that his carefully designed “exclusive” prompts started appearing for free on GitHub, and they were better quality and updated more frequently.

The collapse of information gaps happened much faster than anyone imagined.

According to Stanford University’s 2024 AI Index report, in the past 18 months:

• Basic AI model usage costs have decreased by 87%
• Open-source model performance has improved by 3-5 times
• Average user acquisition costs for AI tools have risen by 220%
• Average time users spend on a single AI tool has dropped from 47 minutes to 12 minutes

Behind this data lies a brutal reality: when technology transforms from a scarce resource to a mass commodity, businesses built around information arbitrage will collapse rapidly.

David’s story repeated itself throughout 2023. From prompt stores to AI courses, from “ChatGPT usage secrets” to “AI art masterclasses,” these businesses relying on information gaps were like castles on the sand, rapidly collapsing before the waves of technological democratization.

More dangerously, this trend is accelerating. OpenAI’s GPT-4o mini model, launched in 2024, performs equivalently to 2023’s GPT-3.5 but costs only one-tenth as much. Meanwhile, open-source projects like Meta’s Llama series, Mistral’s mixture of experts models, and France’s BLOOM are continuously narrowing the gap with commercial models.

Just as the printing press made knowledge no longer the monopoly of a few, AI technology democratization is making “AI expertise” a thing of the past.

The Survival Crisis of “UI Wrappers”#

Walk into any Silicon Valley startup incubator, and you’ll hear similar entrepreneurial ideas: “We’re going to build a better AI writing tool,” “We’re going to create an AI image generator optimized specifically for designers.” The common feature of these ideas is that they all add a user interface layer on top of existing AI models.

The fundamental problem these “UI wrappers” face is: they don’t create any unique value.

Take the wildly popular AI writing tool Jasper from 2023 as an example. In the first few months, it did gain a large number of users due to its excellent user experience. But soon, users discovered they could use ChatGPT directly, or turn to integrated AI features in existing workflow tools like Notion AI and Canva Magic Write.

According to data from Second Measure, Jasper’s paid user count peaked in the first quarter of 2024, then dropped 35% in the following six months. The same story has been playing out repeatedly in AI image generation, AI code generation, and other fields.

The evolution of user behavior has exceeded most entrepreneurs’ expectations. The market is undergoing a fundamental shift from “AI is magical” to “AI is standard.” When users realize AI is just a tool, not an end in itself, they begin demanding tools that truly solve their specific problems, rather than providing generic “AI experiences.”

Just like the “portal era” of the early internet, when all websites offered similar news, email, and search services, those that ultimately succeeded were enterprises providing deep value in vertical fields. AI entrepreneurship today is experiencing a similar screening process.

In the era of technological democratization, superficial differentiation is like footprints on the beach—gone with the first wave.

The Death Spiral of Homogenized Competition#

Let’s look at some unsettling data:

In the AI writing assistant field, there were fewer than 20 major players in early 2023; by the end of 2024, this number exceeded 200. In AI code generation tools, competitors surged from 15 to over 180. In AI image generation, the number exploded from 8 to over 300.

This homogenized competition has led to a classic “death spiral”:

1. Customer acquisition costs skyrocketed: from $10-20 per user initially to$100-150 now
2. Price wars intensified: monthly fees dropped from $49 all the way to$9.99, or even free
3. User loyalty collapsed: average user lifetime value (LTV) dropped by 70%
4. Product differentiation disappeared: functional homogenization rates exceeded 85%

Technology is the entry ticket, not the reason to retain customers. This simple truth is being relearned by more and more entrepreneurs at painful costs.

When all products are based on the same underlying models, when prompt engineering best practices are widely disseminated, when user interface designs tend to homogenize, real competition must come from other dimensions.

Barrier One: The Data Flywheel—The Modern Embodiment of Network Effects#

In 2004, a young Amazon engineer asked Jeff Bezos a question in an internal meeting: “Why are we investing so much in the customer review system? This doesn’t seem to directly generate revenue.”

Bezos’s response later became a business classic: “When you have more user data, you can provide better service; better service attracts more users, which in turn generates more data. Once this positive cycle is established, it’s almost impossible to surpass.”

Twenty years later, this insight has demonstrated unprecedented power in the AI era.

But in the AI field, the logic of the data flywheel is more subtle and powerful than in Amazon’s time. It’s not just about the quantity of data, but more importantly, its quality and relevance.

Imagine: a general AI writing tool might have writing samples from millions of users, but this data is like finding a needle in a haystack—difficult to form targeted advantages. Meanwhile, an AI tool specializing in legal contract review, though it might only have a few thousand users, provides professionally trained, highly structured legal data with every user.

Just as a senior lawyer’s intuition comes from handling tens of thousands of cases, AI’s professional capabilities come from deeply digesting domain-specific data.

Let’s return to Harvey’s story. This AI assistant specially built for law firms didn’t try to be the best in all fields, but focused on the legal vertical. Their success secret is simple yet profoundly deep:

Every time a lawyer uses Harvey to review a contract, the system isn’t just completing a task—it’s learning. Every modification, every annotation, every feedback on results is making the system more “legal.”

The magic of this learning lies in its compound effect: the first user might improve the system by 1% in contract review, but the hundredth user might bring a 10% improvement because the system can apply previously learned knowledge to new situations.

Just like the “aha moments” humans experience when learning new skills, AI systems also suddenly “understand” the deep logic of a field after accumulating enough high-quality data. This understanding doesn’t come from more computing power, but from insights into the patterns behind the data.

The real power of the data flywheel isn’t that it makes your product better, but that it makes your product better in a way competitors cannot replicate.

When a general AI tool needs to process 1 million ordinary samples to reach a professional level in a certain field, Harvey might achieve better results with just 50,000 high-quality legal data points. This is like having a general practitioner and a specialist diagnose a rare disease simultaneously—the specialist’s intuition based on deep training is often more accurate than the generalist’s reasoning based on broad knowledge.

But building a data flywheel isn’t easy. It requires:

• Carefully designed data collection mechanisms: letting users contribute high-quality data seamlessly
• Effective data processing capabilities: transforming raw data into trainable formats
• Rapid application feedback loops: letting data improvements quickly reflect in user experience
• Strict privacy protection measures: maximizing data value while staying compliant

Barrier Two: Workflow Depth—From “Tool” to “Infrastructure”#

On a warm spring afternoon in 2024, in a café in San Francisco’s Mission District, 42-year-old veteran developer Michael Torres was having a heated discussion with his friend—a startup founder.

“I don’t understand,” the founder said confusedly, “my AI programming tool is very powerful, why do users still prefer GitHub Copilot?”

Michael put down his coffee, thought for a moment. “Let me ask you a question,” he said, “when you’re writing code, what do you hate most?”

“Context switching,” the founder immediately answered, “every time I have to switch from one tool to another, it breaks my train of thought.”

“That’s the answer,” Michael nodded, “GitHub Copilot succeeded not because its AI technology is much better than others, but because it understood a core pain point of developers: continuity of thinking is more important than functionality.”

This simple insight reveals a profound truth: the best tools aren’t those with the most features, but those that make you forget they’re tools at all.

Just like a pair of perfectly fitting shoes that you don’t notice but that support you through long journeys, GitHub Copilot’s success secret lies in it becoming a natural extension of the developer’s thought process, rather than a “tool” that requires additional attention.

Before Copilot appeared, the developer experience with AI code generation tools was like this: encounter a programming problem, switch to browser, open the AI tool’s website, describe the problem, wait for generation, copy code, switch back to editor, paste code, adjust formatting, continue working.

Every switch was a cognitive breakpoint, and every breakpoint consumed the developer’s most precious resource: attention.

Copilot simplified this process to: while writing code in the editor, AI suggestions appear naturally, accept or reject, continue working. No switching, no breakpoints, no cognitive load.

Just like water flowing naturally through pipes, this is the highest state of workflow integration.

But the value of workflow depth goes far beyond this. When AI tools are deeply integrated into users’ workflows, they begin to truly understand user context and intent. This is like a long-term partner who doesn’t need much explanation to understand your thoughts.

Notion AI’s success is an even more classic case. Notion spent years building a powerful knowledge management and collaboration platform where millions of users established complex workflows, knowledge bases, and project management systems. When Notion AI launched, it wasn’t just adding an AI feature, but adding thinking capabilities to this already established “digital brain.”

The cost for users to leave Notion AI isn’t just losing AI functionality, but having to rebuild their entire digital life. This deep integration creates extremely high user stickiness, not because of technical barriers, but because users have internalized this tool as part of their work and thinking patterns.

Barrier Three: Community and Ecosystem—From “Users” to “Co-builders”#

In the winter of 2023, in a small apartment in Brooklyn, New York, 29-year-old digital artist Elena Vasquez was experiencing a creative crisis. Her once-proud painting skills seemed so pale and powerless in the face of AI.

“I feel like a craftsman about to be eliminated,” Elena wrote on her art blog, “when machines can generate in one second what takes me a week to complete, where is my value?”

In desperation, Elena tried Midjourney. But to her surprise, the platform didn’t make her feel replaced, but instead found new creative motivation.

In Midjourney’s Discord server, Elena discovered over 15 million creators like herself. They weren’t just using AI to generate images, but were sharing techniques, inspiring each other, and building friendships. Some specialized in Renaissance-style prompts, others mastered cyberpunk aesthetics, and some created unique “AI + hand-drawing” hybrid techniques.

“I realized AI isn’t here to replace artists, but has given us a new creative language,” Elena said in a later interview, “just as photographers didn’t make painters unemployed, AI won’t make us lose creativity, it just changes how we express our creativity.”

This story reveals the real secret of Midjourney’s success: it’s not just an AI tool, but the birthplace of an artistic movement.

Midjourney’s success has always been a mystery. Technically, it’s no more advanced than Stable Diffusion or DALL-E; in terms of product, its user interface could even be called crude; in terms of business model, it uses the most traditional subscription model. Yet it has become one of the most successful companies in the AI image generation field, with monthly revenue exceeding $20 million.

The answer lies in it unintentionally triggering an ancient human phenomenon: the awakening of collective creativity.

In Midjourney’s community, users aren’t just consumers, but co-creators. Every generation, every share, every comment contributes to this collective intelligence. Like the builders of medieval cathedrals, everyone contributes their skills, jointly creating a great work that transcends individuals.

The value of community doesn’t lie in the number of users, but in what kind of connections form between them. Midjourney’s success lies in creating three seemingly simple yet extremely powerful elements:

1. The transformation of identity: Users are no longer “people who use AI tools” but “artists of the AI era.” This identity shift transforms users from passive recipients of technology to active creators.

2. The spontaneous formation of cultural norms: The community internally developed unique language, aesthetic standards, and creative methods. Terms like “v5 style,” “cinematic lighting,” and the rise of various style streams have all become cultural symbols of the community.

3. The positive cycle of value co-creation: Every user’s creation contributes value to the entire community. Excellent works inspire others, new techniques spread rapidly, and the community’s overall aesthetic level continuously improves.

But what’s most magical is that none of this was deliberately designed by Midjourney, but emerged naturally. Just as a city isn’t completely designed by urban planners but is the product of countless individual interactions, Midjourney’s community culture was also spontaneously created by users.

Once this spontaneously formed community culture is established, it’s almost impossible to replicate. Competitors can copy your technology, imitate your features, but cannot replicate the cultural DNA created by specific people at specific times and places.

This kind of community barrier, once established, becomes extremely difficult to replicate. Competitors can copy your technology, can copy your features, but can hardly copy your community culture.

Building community barriers requires:

• Clear value propositions: letting users identify with your mission and vision
• Effective participation mechanisms: letting users deeply participate in product development
• Cultural symbol systems: creating unique language, rituals, and identity markers
• Value distribution mechanisms: letting users’ contributions receive reasonable returns

Case 1: Cursor—From “Just Another Programming Tool” to “Developer’s Thought Companion”#

In early 2023, when Cursor first launched, there were at least 20 similar AI programming tools on the market. Most tools’ fate was to be fleeting, but Cursor stood out and became many developers’ first choice.

What did Cursor do right?

First, deeply understanding developers’ cognitive processes. The Cursor team realized that programming isn’t just writing code, but a complete cognitive process of thinking, designing, debugging, and refactoring. They didn’t simply create a “code generator” but created a “developer’s thought companion.”

Second, seamless workflow integration. Cursor is deeply integrated into VS Code, making AI assistance a natural extension of the development process rather than an additional step. Developers can get AI help without leaving their familiar environment.

Third, context-aware intelligence. Cursor can understand the overall structure and context of a project, providing more accurate suggestions than general code generation. It knows what type of application you’re building, understands your coding style, and can even predict what functionality you might need.

Fourth, continuous learning loops. Through developer usage feedback, Cursor continuously optimizes its ability to understand code logic and developer intent. Every developer using Cursor is contributing to improving the tool.

Cursor’s success teaches us: in the AI era, tool success doesn’t lie in the number of features, but in the depth of understanding user workflows.

Case 2: Perplexity AI—Surviving Under Google’s Shadow#

When Perplexity AI launched in 2022, many thought it was courting disaster—after all, who dares to challenge Google in the search field?

But Perplexity not only survived, but reached 50 million monthly active users in 2024, with a valuation exceeding $1 billion.

What did Perplexity do right?

First, redefining the value proposition of search. Perplexity didn’t try to compete with Google in “general search,” but focused on “academic and professional search.” It understood that for researchers and professionals, what they need isn’t as many results as possible, but as accurate, reliable, and well-cited answers as possible.

Second, building a credibility data flywheel. Every time users use Perplexity for professional searches, they’re providing high-quality professional data to the system. This data is used to improve the model’s performance in professional fields, thereby attracting more professional users.

Third, establishing authoritative reputation. By providing accurate citations, transparent information sources, and professional answer formats, Perplexity established strong brand credibility in academic and professional fields.

Fourth, creating a differentiated user experience. Perplexity didn’t imitate Google’s link list model but created a brand-new “conversational answers” experience that better matches modern users’ cognitive habits.

Perplexity’s case tells us: even in a market dominated by giants, as long as you find a sufficiently segmented and important user need, and provide truly differentiated value, there’s a chance to succeed.

Case 3: Notion AI—The Perfect Embodiment of Platform Advantage#

Notion AI’s success might be the least surprising, but also the most worth learning from.

What did Notion AI do right?

First, patiently waiting for the right moment. Notion didn’t rush to launch AI features during the most frenzied AI hype, but waited until technology matured and user needs were clear before making its move.

Second, deep integration rather than surface addition. Notion AI isn’t a standalone product but is deeply integrated into Notion’s entire ecosystem. AI functions can be used in any Notion page, seamlessly cooperating with existing database, project management, and knowledge management features.

Third, leveraging existing network effects. Notion has tens of millions of users and millions of established workspaces. When AI features launched, users didn’t need to learn entirely new tools, just add some new features to their already familiar environment.

Fourth, creating collaborative value. Notion AI isn’t just a tool for individual users, but an enhancer of team collaboration. It can help teams quickly summarize meeting notes, generate reports, analyze data, creating team-level value.

Notion AI’s insight is: if you already have a powerful platform, AI shouldn’t be a standalone product, but an enhancement of platform capabilities.

Self-Diagnosis: What Stage Is Your Product In?#

To evaluate your AI product’s competitive barriers, you can diagnose from the following dimensions:

Data Flywheel Maturity:

• Do we have systematic data collection mechanisms?
• Is the collected data highly relevant to our core business?
• Do we have the capability to quickly transform data into product improvements?
• Are users seamlessly contributing high-quality data during use?

Workflow Integration Depth:

• Is our product a natural extension of users’ workflows?
• Would users need to restructure their workflows to switch to competitors?
• Does our product solve industry-specific “last mile” problems?
• Do we deeply understand users’ complete work cycles?

Community Engagement Level:

• Is there meaningful interaction and connection between users?
• Have we created unique community culture and identity?
• Can users’ contributions receive visible returns?
• Have we established effective community governance mechanisms?

Gradual Building Strategy#

Optimal Strategies Under Resource Constraints#

For resource-constrained small teams, the following strategies are recommended:

Focus, focus, and focus again: choose a niche market that’s small enough but important enough, and achieve excellence. Don’t try to be “everything to everyone.”

Leverage existing platforms: fully utilize existing open-source models, cloud services, and developer platforms, don’t reinvent the wheel. Your value lies in the unique combination and application of these resources.

Validate quickly, iterate quickly: adopt lean startup methods, quickly validate hypotheses, quickly iterate products. Don’t pursue perfection, pursue progress.

Establish early user relationships: build deep cooperative relationships with early users, letting them become your product co-creators and promoters.