2025 was a dramatic year for me.

Many things changed. I began devoting more time to Fumadocs and other side projects, and the adoption of my libraries grew rapidly as they matured.

I created my ORM layer FumaDB, launched Fuma Content, and continued shipping new features to Fumadocs, which surpassed 10k stars on GitHub.

Yet the pace of frontier innovation and development accelerated even further. The AI hype showed no signs of slowing, and I, too, felt increasingly excited about the future while integrating AI agents into my daily workflow.

Undeniably, the rapid advancement of AI models came as a surprise. At times, coding with AI felt like a real game-changer, delivering a huge productivity boost.

I used to treat AST manipulation as a daily brain exercise, but much of the appeal of programming tarnished once AI could mass produce efficient code, including remark/rehype plugins.

Near the end of 2025, I began shifting my attention back to my studies. The period felt tumultuous as I struggled to find direction within the constraints of reality: uncertainty about the future, my ambition to build a startup, and the reality that my current position and schedule prevented me from fully committing to software development or joining the startups I admired.

The viral growth of Fumadocs brought many opportunities. It was a genuine “wow, my dream came true” moment when Guillermo Rauch DM’d me on Twitter about joining Vercel.

At the same time, I felt deep internal conflict. While I took the difficulty of US companies hiring teenagers overseas as an excuse, it was also true that I was not fully prepared to balance a career with high school studies. As a perfectionist, I found it treacherous to devote insufficient attention to either responsibility.

So, I continued working as an open-source maintainer throughout the year.

Before 2025, I treated open source purely as a personal hobby, pure and carefree, to mess around and building something cool.

2025 became the year of Fumadocs. It was adopted by many excellent projects and teams, including Zod, BetterAuth, Vercel, Sim Studio, Better-T-Stack, and many others.

2026 has a good start as well: Prisma announced their new documentation powered by Fumadocs, and I am already planning more AI-native features for the project.

Fumadocs has truly emerged as the successor to Nextra and a closer alternative to Mintlify. Suddenly, Fumadocs have the responsibility to remain reliable for many consumers.

But that responsibility is thrilling! Being useful is actually a remedy against inferiority complex, and personally, I am really curious about what Fumadocs will be.

For Fumadocs, the top priority is to make first-class LLM integration smoother, ideally seamless, and beginner-friendly.

At the same time, I am working on a stronger foundation for content processing in Fuma Content, which Fumadocs will soon build upon.

Beyond that, I occasionally allow myself time to reflect on broader questions.

Noise fills Twitter as people speculate about the future of AI, yet that very cacophony has sparked several thoughts worth considering.

Many pointed out the possibility of reading source code instead of docs, treating code & comments as the only reliable source.

While it is plausible, this model is not efficient because interpreting code still wastes a lot of computational power. Docs captured the high-level vision of a project, which is exactly the information that is missing from source code.

User-facing docs are still common, for instance, the help center for apps, and docs of UI libraries. Their intended audience remained to be human rather than LLMs/crawlers.

Fumadocs invested in the tooling to keep docs up-to-date effortlessly without spending tokens, like docs generated from TypeScript & OpenAPI spec, and my favourite Story integration that brings beautiful playground for component libraries.

There are many cool ideas I have yet to explore, but I am certain the future is more about writing docs under the assistance of AI, than deprecating docs entirely.

My own prediction is that AI will become integrated at the operating-system level. Software will expose interfaces for AI agents to interact with, just like how applications today provide IOS widgets.

This approach would be far more efficient and intuitive than current tools such as MCP, which remain familiar only within technical circles.

The story of OpenClaw reinforces a neutral but plausible view: AI will ultimately become an OS-level capability.

I plan to invest more heavily in the infrastructure and ecosystem that AI will depend on (Fumadocs is already powering some vibecoded documentation sites!).

AI models and hardware remain the most irreplaceable components. Competing in the broader ecosystem surrounding AI will likely be the primary objective of many startups in the coming years.

Although the future remains full of uncertainty, I am optimistic that we will witness the promises of AI come to fruition.

What a great time to be alive!