I've always wanted a customized LLM knowledge base that can integrate various information sources to improve answer accuracy. WeChat's IMA Knowledge Base provides some customization capabilities, but since most of my work files are not stored on WeChat, manually importing data and interacting with a mobile-based model does not fully meet my needs. Similarly, ChatGPT's Build Your Own GPT follows the same principle but lacks convenient collaboration features for internal team members. After browsing GitHub for potential open-source tools, I found no suitable solutions readily available. Since the functionality itself is not particularly complex, I decided to let Devin develop an internal knowledge management system for our company.

My requirements included:

• Account login system

• Collaborative knowledge base maintenance (contributions and management)

• Ability to switch between different large models

In less than 30 minutes, I had a fully functional internal knowledge management tool deployed on Devin's virtual machine, complete with a test link for verification. While the UI was basic and plenty of optimizations could be made (e.g., supporting more models via API key imports, handling additional file formats), the first MVP already met all my functional requirements. Future modifications could be accomplished largely through prompts. Building your own app—was it really this seamless? (BYOA platforms are plentiful, and Devin is just one example.)

This is Worrying—in a Good Way

In the past, rapid product development was a competitive edge for teams. A somewhat amusing example is that every student who took Stanford’s first iOS app development course reportedly earned at least $1 million because, at the time, the App Store had just launched, app development had a certain entry barrier, and any semi-useful app (remember those utility tools?) could attract massive traffic.

However, we are now entering a foreseeable future where code is no longer a scarce resource, and neither is the ability to build products. Ordinary users can generate the tools and products they need simply through natural language commands. How will this reshape competition among companies and user behavior? Here are some bold predictions.

Categorizing the Impact: Consumer vs. Business Users

I first divide users into consumer (C-end) and business (B-end), then further classify C-end users into two primary use cases: productivity and killing time.

C-end: Productivity Tools—Limited Disruption

For productivity-related use cases, the impact on user behavior may remain relatively small. While AI coding tools enable high customization, the scenario where individual users build personal productivity tools remains niche. We can draw a parallel to website creation in the Web2 era—even though many no-code site builders exist, the proportion of users actually running personal websites remains minuscule compared to the overall internet user base.

Users prefer existing platforms and products—even if they lack full customization, convenience is far more important.

ChatGPT has already educated the market, proving that there are prosumers willing to pay $20/month for enhanced productivity. Senior professionals may pay for customizable efficiency tools, but the format of such tools—whether they remain covered by underlying large models or evolve into a new product category—remains uncertain.

A simple assumption: even if customization is available, the frequency of such needs remains low. For example, the KMS system I built with Devin likely won't require frequent updates.

C-end: Killing Time—The Rise of Reverse Recommendation Algorithms

This is where I anticipate the emergence of a reverse recommendation algorithm-driven content network. If code abundance leads to content abundance, the most valuable elements become user preferences and social relationships.

In an ideal scenario, I could carry my preference data (protected through zero-knowledge proofs for privacy) and use it to initialize a completely new content network.

Consider an extreme but not impossible case: AI generates an entire content platform (akin to RedNote or TikTok) based solely on user preferences.

This diminishes the network effects derived from content barriers while amplifying social-driven network effects. Social platforms could deliver faster, better-matched content—long Tencent, short Bytedance?

The Problem with Recommendation Algorithms—A New Solution?

Over the past decade, recommendation algorithms have been the biggest variable in consumer-facing internet applications. The smartest minds in the world have spent years figuring out how to keep users engaged for longer.

However, as Jonathan Haidt highlights in The Anxious Generation, research increasingly links recommendation algorithms to mental health issues in teenagers.

From a historical perspective, it's hard to say whether recommendation algorithms represent a productivity enhancement or a slow erosion of human well-being.

Fighting AI with AI may be the best strategy—using AI as a curator of high-quality content.

In the Web3 world, there is an emphasis on users owning their data and preferences. Through decentralized economic incentives, Web3 aims to establish social networks driven by users rather than opaque algorithms.

While this vision is ideal, it is unlikely to happen overnight. Before we overthrow TikTok, Instagram, and Twitter, we will likely experience intermediate phases of evolution.

The Shifting Role of Tools vs. Networks

For companies built on the “Come for the tool, stay for the network” model, the significance of tools may diminish, while the importance of networks grows. Furthermore, the definition of a network itself may change.

Compared to social relationships, e-commerce supply chains, and payment networks, content-based network effects are inherently weaker.

B-end: How SaaS Companies Can Adapt

On the business side, existing SaaS companies can offer more customization to users. While core processes and data structures (e.g., SAP) won’t change easily, the UI/UX layer provides room for significant improvements.

For example, I may not migrate my entire company from Slack to another platform, but Slack could introduce extensive no-code customization:

• Auto-replying to messages from the boss (“Got it 🫡”)

• Summarizing group chat discussions

• Generating weekly reports from chat history

Thus, AI won’t kill SaaS, but it will (hopefully) make SaaS more customizable.

Code Abundance & Crypto: Net Positive Impact

The crypto industry suffers from several well-known issues:

• Poor user experience

• Lack of PMF (Product-Market Fit)

• High learning curve for new users

• High development barriers

Unlike traditional internet products, crypto lacks clear “winner-takes-all” players.

AI-driven code abundance accelerates crypto’s search for PMF, improves UX, and makes on-chain applications more accessible to non-crypto users.

⚠️ Note: Code abundance does not guarantee that crypto will find PMF, but it significantly accelerates the process—just as it will for other emerging industries.