2025 Year in Review: Slow Down to Go Fast

Published:
  Lastmod: 2026-01-02
Year In Review
by Ernest Chiang

2025 Year In Review: Slow Down to Go Fast (Caption: Deliberately captured Tokyo Ginkgo with the Ginkgo Camera GR IV. Image source: Ernest Chiang.)

“Sometimes you got to slow down to go fast.”

This phrase kept echoing in my mind throughout 2025. While every industry chases the efficiency gains promised by AI, and anxiety about the future and artificial intelligence pervades, I kept returning to this note in my journal: “True progress often comes from deliberate pauses and deep thinking.”

The hard part isn’t stopping or slowing down—it’s being deliberate.

Deliberate practice, deliberate thinking, deliberate connection, deliberate verification.

✳️ tl;dr

  • Core mindset shift from “thinking code, writing code, reviewing code” to “building knowledge and workflow systems that enable AI to iterate autonomously.”
  • 2025 Vibe Coding (Agentic Coding) tool journey: Cline/RooCode → Cursor → Claude Code → Claude Code + Kiro.
  • Core insight: Coding is Easy, Context is Hard. Once the engineering bottleneck is removed, the bottlenecks of product thinking, human communication, and context management are just beginning (exhausting…).
  • Branded my knowledge management methodology as “Humbled Productivity (土炮工作法)”, emphasizing three pillars: People, Workflow, System.
  • Health and life balance: Weekly weight training habit since February, handwritten notes, Cognitive-Information-Physical (CIP) framework, family, travel, local experiences, good food, rituals.
  • 2026 direction: Building infrastructure, combined with Agents and Skills, to help teams and clients progress toward their goals through Humbled Productivity.

✳️ Foreword: Re-learning to Code This Year

Looking back at 2025, my biggest change wasn’t “getting better at using AI”—it was a transformation in my role and positioning.

By fortunate circumstance (privilege?!), I had the opportunity to touch a computer at age three and started learning BASICA programming. In elementary school, instead of reading the manga everyone else was reading, I was reading MS-DOS manuals. In junior high, while everyone was preparing for exams, I was translating PHP3 tutorials. In high school, while everyone continued stressing about college entrance exams, I received a rare out-of-print FreeBSD black-cover book as a reward for translating computer books… From my earliest days as a software developer, to semiconductor process integration engineer, to technical manager, then to product and technology integration, and finally to entrepreneur—this journey has spanned decades. Although I’m still on the path (still on it, right?!), 2025 was a turning point: I no longer just “get things done with my hands,” but beyond hands-on work, I deeply think about “how to design systems that increase the probability of things being done well, invisibly,” while continuing to pull clients along (accompanying them) in hands-on implementation and live testing.

This transformation was, to some extent, pushed by the AI wave. When AI (LLM) can rapidly generate code, I realized my value couldn’t just be “thinking fast,” “coding fast,” or “speaking fast”—it had to be “thinking clearly,” “writing clearly,” and “communicating clearly.” Or more precisely: being able to transform vague requirements into structured specifications and workflows that both AI Agents and humans can understand and execute.

With helping clients address their pain points as the main axis, I used the Working Backwards methodology to start building an operational framework: starting from the customer’s problem, working backward to determine what kind of product might be needed, what kind of process, what kind of system, how to align stakeholders, how to reduce the cost of understanding, digesting, and absorbing information. This isn’t a new concept, but in the AI era, this mindset has become even more important—because AI can accelerate “doing,” but it can’t decide “what should be done” for us.

“Slow Down to Go Fast”—this phrase pretty much captures my main theme for 2025.

When everyone is chasing faster, more, more automated, more anxious, I—someone habitually pursuing efficiency—chose to deliberately slow down, to deliberately think: I know my team and I can handle, solve, and manage most types of tasks, but, what is truly important? What is truly essential? How should we choose? What can continuously accumulate value?

The answers gradually became clear:

  • Deliberate deep understanding of problems,
  • Deliberate cross-domain connections,
  • Deliberate ability to systematize tacit knowledge.

(Caption: If you’re interested, put on your work uniform and join me in reviewing 2025!)

1️⃣ The Balance of Technology and Product: Vibe Coding (Agentic Coding)

1.1 2025 Q1: Inspiration from YC RFS 2024 — Reimagining ERP

Last year, while researching Y Combinator’s Requests for Startups 2024, one direction particularly caught our team’s attention: NEW ENTERPRISE RESOURCE PLANNING (ERP) SOFTWARE.

YC’s Dalton Caldwell described it this way:

“ERP is the operating system of a business, but it’s universally expensive, painful to implement, and users don’t like it—yet it’s absolutely necessary and critical. We want to see startups build software that businesses love—flexible, easy to use, and beloved by customers.”

This description reminded me of scenes I’d witnessed in traditional industries over the past decade: complex ERP systems, data silos that were hard to integrate, and employees held hostage by processes. If AI can redefine software development, then could ERP—this “operating system of business”—also be reimagined? Over the past three years, in our spare time, we’ve been conducting in-depth interviews with two client partners, deconstructing organizational workflows, and finally integrating the fundamental objects that the entire organization can share across departments. Following Jeff Bezos’s 2002 internal memo “Amazon API Mandate,” we provided all of our client’s teams with (nearly) all workflows as API services. This has gradually evolved into one of the core product lines of Kyklosify Consulting, a boutique U.S. consulting firm—“Kyklosify BusinessSuite.”

As Q1 progressed and friends continued to refer clients, at least five electronics and traditional manufacturing companies inquired whether we could help update their legacy ERP systems. Carrying this question, although we knew that transitioning such core systems couldn’t rely on technology alone, if we could break through the technical bottlenecks of the past, might this be an opportunity that plays to our strengths in deconstruction, deep diving, and integration implementation? Alongside the curiosity about what vibe coding/agentic coding could actually achieve, we began conducting field research and various experiments with our partners.

(Caption: Q1, wishing everyone a happy new year with the bright red and gold-stamped Firefox red envelope.)

1.2 2025 Q2: Field Research on Traditional ERP with Partners

In the second quarter, we spent several weeks of our spare time deeply deconstructing a traditional ERP system—including its database structure, business logic, exception handling, and how customization was managed.

This process allowed our team to accumulate extensive experience and pain points using various AI vibe coding development tools. We divided the work, each exploring different tools. At the time, I primarily used Cline/RooCode with Amazon Bedrock and Cursor (thanks to AWS Taiwan for sponsoring research credits for the former; for the latter, I paid out of pocket), trying to have AI help us quickly understand complex legacy code and database schemas.

Honestly, the initial experience wasn’t smooth (remember, this was around early 2025, right when Amazon Bedrock had just made the Sonnet 3.7 model available and Claude Code had just been released). AI would often “get lost”—giving suggestions that seemed reasonable but actually deviated from the goal (and honestly, way too many suggestions and opinions). Or when handling multi-dimensional tasks, it would get lost in details and forget the original objective, with adjustments making things progressively worse (requiring a lot of patience, and yelling at it didn’t help).

There were still some smooth highlights and sense of achievement during the process. For example, in just a few weeks, while cross-referencing documentation from another ERP system (yes, you need to accumulate domain knowledge and knowledge bases regularly—you don’t realize you need the books until you need them, documentation too), we replicated the front-end interfaces of several core modules, exciting both our clients and the entire team.

These hard-won experiences actually helped my team understand more clearly: the value of AI tools isn’t in “providing answers,” but in “accelerating exploration,” “accelerating trial and error,” and “accelerating validation.” The key is that humans must be able to set the right boundaries, provide sufficient context, and know when to stop and rethink. (At this point, the various documents and process standards we’d accumulated over the years gradually became an advantage. Originally, humans might not have read them, but now we could hand them to AI Agents to read.)

(Caption: Q2, received a letter from France processed by Asendia. Asendia’s hub in Europe (like Paris, France) can print, sort, label, and ship for clients—essentially the AWS of postal services.) (As a fan, even receiving a letter has to be connected to AWS somehow… am I upgrading from iron fan to aluminum fan…)

1.3 2025 Q3/Q4: Practical Results in Embedded Systems and Cloud Development

In the second half of the year, our team began extensively applying this AI-assisted development methodology to embedded systems and cloud infrastructure development.

The results were quite significant:

  • Development progress significantly ahead of schedule
    • Features originally estimated to take months were completed in working versions within weeks.
  • Simultaneously advancing multiple products
    • Because development time for individual features was shortened, the team had capacity to handle multiple product lines simultaneously.
  • Faster iteration cycles with clients
    • Faster prototyping means faster feedback, faster corrections, faster convergence.
    • The prerequisite is that clients are willing and able to keep up with the iteration cycle. The most common bottleneck is providing feedback and making decisions within the cycle.
    • As a result, we mostly choose to work with the highest-authority holders in the organization as our client partners, such as the board, CEO, or business unit VPs.

But there’s an important prerequisite here: behind these results lies a massive amount of “spec writing” work. The time we spent writing various types of specs was no less than coding time. But this investment yielded higher quality, more predictable, and more maintainable output. Experience from years ago at the TSMC process integration engineering department, as well as participating in international specification development at the Bluetooth SIG Working Group, came into heavy use.

(Caption: A rare late-night work session at the end of Q4, but I still got up early to visit the local breakfast spots at Taichung’s Second Market. With delicious local breakfast, I was instantly energized, haha.)

1.4 Vibe Coding Tool Evolution and Lessons Learned

Looking back at 2025, my AI development tool usage went through a clear evolution:

  • Cline/RooCode → Cursor → Claude Code → Claude Code + Kiro
  • ChatGPT Pro
    • (USD 200/month; I distinctly remember 2025-02-03 when OpenAI announced Deep Research in Tokyo, I immediately upgraded from Plus to Pro. Later, Claude Code was too good, so I downgraded back to ChatGPT Plus and reallocated to Claude Max for Claude Code.)
  • Claude Max
    • (USD 100/month; my favorite, highly recommended!)
  • Gemini Pro, NotebookLM
    • (USD 16.8/month; included with Google Workspace Standard)
  • Amazon Bedrock, Google AI Studio
    • (for infra projects)
  • LM Studio
  • Apple Intelligence

Each transition wasn’t a direct switch, but more of a gradual transition with experimentation and validation along the way.

When I first started using Cline and RooCode, I was still stuck in the mindset of “treating AI as a smart autocomplete” and “how to write prompts to get better output.” Throw in a prompt, see what it generates. This approach was effective for small tasks but quickly hit walls in complex projects.

After switching to Cursor, I began to understand the importance of “context.” Cursor could index the entire project, allowing AI to better understand the relationships between code. But even so, when handling cross-module, cross-system tasks, AI would still lose focus. (Part of it was also that I was too quick to dive into fine-tuning details, and AI would follow me down that path…)

The real turning point was the emergence of Claude Code. It was no longer just an IDE plugin, but an agent that could “browse the repo itself, read files itself, trace dependencies itself.” The first time I used it, I thought it would be the same old story—manually copying and pasting code, waiting for it to guess—but Claude Code just started exploring autonomously.

That moment I realized: I wasn’t using a tool, I was working alongside an agent that could think and discuss with me. (So I immediately started shifting my mindset—the framework for collaborating with team members also needed to be understood by, and even iterated with, AI agents.)

By the second half of the year, when Kiro was released, we saw a more complete vision. Kiro isn’t just a coding tool, but an attempt to build a framework for “Intent-Driven Development.” It has Specs (structured specifications), Steering (project guiding principles), Hooks (event triggers)—these components combine to form a complete human-machine collaborative development environment.

(I recommend checking out Pahud’s video from 2025-08-13, “Multi-Role Development using Kiro”. This was just a few weeks before I headed to Seattle for AWS Heroes Summit 2025 to have in-depth discussions with the AWS CDK Team and other service teams. After visiting Silicon Valley’s South Bay, the Northern California Taiwanese Chamber of Commerce, and Seattle in late August to early September, seeing various industries and enterprises in the US already trying to implement various AI Agents in practice further confirmed our future direction.)

For more in-depth analysis of these AI development tools, check out my notes: Kiro AI: Agentic Integrated Development Environment and Claude Code Learning Notes. You don’t need to be an engineer—everyone can use these tools. Imagination is encouraged.

(Caption: Exploring the Kiro haunted house! The house of Kiro!)

1.5 From “Vibe Coding” to “Spec-Driven Development”

The most important conceptual shift this year was moving from “Vibe Coding” to “Spec-Driven Development.” (I think it doesn’t matter what you call it—don’t get too hung up on names. You’ll always encounter “your A isn’t their A.” Looking at the essence, what needs to happen still needs to happen, what needs to be done shouldn’t be skipped. But having a roughly consistent term might make it easier to consolidate and search for related experience sharing.)

Some say Vibe Coding is just throwing a vague prompt at AI and seeing what luck brings. This approach is useful for quick prototyping in small scopes. Even when providing very specific prompts or relevant documents to AI, output quality can be unpredictable due to the chosen model, context size limit, rate limit, and untouchable system prompts—making it hard to incorporate into production products. (At this stage, our control method is mainly to divide tasks into small enough scopes and combine with multi-shot.)

Spec-Driven Development represents a different mindset: first use human language to define as clearly as possible “what to do” and “why to do it,” then let AI handle the “how to do it” execution details. (Very suitable for product teams with relatively complete documentation, but professional service companies may need to rely heavily on the communication and articulation abilities of client-side contacts.)

This reminds me of my experience over a decade ago participating in international standard development at the Bluetooth SIG Working Group. Back then, we spent enormous amounts of time writing specification documents (FRD, specification), defining use cases, behaviors, boundary conditions, and error handling for every protocol. These documents seemed dry, but precisely because of clear specifications and test plans, manufacturers around the world could develop Bluetooth products that were compatible with each other.

These documents seemed dry, but precisely because of clear specifications and test plans, manufacturers around the world could develop Bluetooth products that were compatible with each other.

Specs are essentially the PRD (Product Requirements Document) and FRD (Functional Requirements Document) in product development. But in the AI era, these specification documents take on new meaning: they’re not just for humans to read, but instructions and context for AI to execute. (This should continue for a while in the Transformer-based LLM AI era, and terminology will likely change over time (e.g., Context, MCP, Skills, etc.) until the next-generation architecture or wave arrives, and then it will all repeat again (just get used to it, no need for anxiety—different terms are just for namespace)).)

This mindset shift aligns with the views of deep learning pioneer Yann LeCun. He emphasizes the paradigm shift from “Imperative Development” to “Intent-Driven.” Developers focus on clearly expressing “why we’re doing this” and “what we’re achieving,” while delegating the tedious “how to do it” to more intelligent systems.

Further reading: One of the frameworks for intent alignment (hasn’t been named yet, but the diagram I use most often is this one), included in my June 2025 Taiwan Product Conference presentation “Interoperate, Integrate, Iterate: A 10-Year PM Survival Kit for Traditional Sectors” slide deck.

1.6 Core Insight: Once the Engineering Bottleneck Is Removed, the Product Thinking Bottleneck Begins

This year’s deepest insight about vibe coding or agentic coding is probably captured in this phrase:

Coding is Easy, Context is Hard.

At an abstract level, this follows the same thread as what people used to say before AI: technology isn’t the most important thing.

When AI can rapidly produce code, the bottleneck is no longer “can’t write it,” but “can’t think it through” and “can’t connect the dots.” This aligns with what Amazon CTO Werner Vogels emphasized in his latest (and final, sadly) re:Invent 2025 Keynote. He introduced a key concept: "Verification Debt".

Werner pointed out: “AI generates code faster than you can understand it. Code arrives instantly, but 'understanding' can't keep up.” This aligns with Anthropic CPO Mike Krieger’s observation: over 70% of Pull Requests at Anthropic are AI-generated, and even their Claude Code tool itself was approximately 95% written by AI.

But paradoxically, when code output speed explodes, the team’s Merge Queue gets completely backed up. This is exactly like the “Theory of Constraints” mentioned in the classic management book “The Goal”: when you solve one bottleneck, other bottlenecks emerge. Werner cited ecologist Donella Meadows’ Systems Thinking to explain this phenomenon—just as introducing wolves to Yellowstone changes river flows, massively introducing AI code will fundamentally change the balance of software delivery systems.

The current bottlenecks have become:

  • Upstream (Decision-making and Alignment):
    • How do we reduce the ambiguity of natural language?
    • Werner mentioned the risks of “Vibe Coding” and emphasized the need to shift toward “Spec-driven Development.”
    • Rather than letting AI guess your intent (or complaining that AI doesn’t understand you), invest time first in writing clear specifications and requirements.
  • Downstream (Coordination and Delivery):
    • How do we ensure system resilience?
    • Werner emphasized “The work is yours, not the tools.”
    • The more AI writes, the more humans must transform from “writers” to “Owners” and “reviewers.”

The risk we face is: if humans no longer understand code line by line, will the Codebase become something only AI can understand?

This is indeed a huge challenge. But as the re:Invent Keynote advocated with the “Renaissance Developer” spirit, the solution isn’t to retreat to the era of “manual handwriting,” but to establish stronger “Mechanisms”—such as more rigorous specification documentation, more automated testing, and focusing human intelligence on cross-domain system design that AI (temporarily?) cannot replace.

True competitiveness lies at the intersection of “Domain Insight × Systems Thinking × Verification.” The battleground for future product managers and architects isn’t just feature planning, but—as Werner said—learning how to communicate with AI using precise specifications, and taking full responsibility for the final system behavior.

(Caption: Amazon CTO Werner Vogels said this was his last AWS re:Invent Keynote. I was right there in the audience, strongly feeling his reluctance to let go and his hopes for the future. There’s video—turn on the sound to feel it.)

1.7 Sharing and Exchange

This year I had several opportunities for public and private sharing, which allowed me to exchange these practical experiences with more friends. I’m very grateful to the organizations and enterprises that invited me to share, giving me the opportunity to share this first-hand information and receive cross-domain feedback. This enabled our team to validate ideas and accelerate iteration. Looking forward to our next meeting.

Here are some of the larger public sessions:

  • 2025-05: AWS Summit Hong Kong
    • I shared “Reinventing Programming: How AI is Transforming Enterprise Code Development” at the Dev Lounge.
    • Starting from each phase of the SDLC, I demonstrated how AI tools can assist in planning, design, implementation, testing, deployment, and maintenance. Exchanging with friends from software development, finance, and traditional industries, I found that everyone is exploring how to implement these tools in ways suitable for their teams (and exploring how to secure budget XDD).
  • 2025-06: Taiwan Product Conference

2️⃣ The Balance of Knowledge and Operations: Humbled Productivity (土炮工作法)

2.1 From Information Accumulation to Systematic Output

Another important experiment in 2025 was (slowly) attempting to package my knowledge management methodology as “Humbled Productivity (土炮工作法).” Don’t rush—currently there’s only partial experimental release. Hopefully 2026 will see gradually expanded scope. (But feel free to sign up as a guinea pig—we focus on essential analysis and won’t spread anxiety. (Hopefully.))

  • The term “土炮” (DIY/grassroots approach) comes from our team’s insistence on being “grounded” (perhaps a cultural thing?!). Not chasing the flashiest tools or newest frameworks, but seeking and validating solutions that best fit our clients and team.
  • “Humbled” emphasizes a humble attitude—when facing new technology, acknowledging our shortcomings (all kinds of shortcomings, all kinds of impostor syndrome, sigh), being willing to start from zero and relearn from the basics.
  • In the AI era, an easy trap to fall into is: when encountering a problem, just ask AI directly. But this habit gradually erodes our thinking ability.
  • Anthropic CPO Mike Krieger, when talking about educating children, mentioned a viewpoint I strongly agree with:

“Before asking AI, first ask ‘how can we figure out the answer ourselves?’

What experiments can we do? What can we observe?”

This is similar to my ongoing thinking about how to “think before AI.” AI can be a powerful tool, but if we outsource all thinking to it, we’ll gradually lose the ability to solve problems independently. AI can also be a powerful partner—through back-and-forth discussion with AI, we can exercise our thinking and strengthen our problem-solving approaches. Neither is superior—it depends on trade-offs and purpose.

For those interested in how to exercise thinking when stripped of AI, here’s some further reading:

2.2 The Connection Between Handwritten Notes and Context Engineering

This year I read a research paper about handwritten notes. The research shows that handwriting promotes brain connectivity more than typing. The theta/alpha frequency connectivity patterns during handwriting are crucial for memory formation.

  • This explains why when I encounter complex problems, I always pick up my reMarkable Paper Pro or Field Notes to think by hand—this isn’t being “retro,” it’s cognitive science best practice.
  • What I find additionally interesting is that this somehow echoes “Context Engineering” in the AI era. Context Engineering emphasizes how to effectively manage AI’s context—what information to feed to AI, what to omit, how to organize it so AI better understands our intent.
  • The process of handwritten notes is essentially doing “Context Engineering for the human brain”: filtering, organizing, building connections, strengthening memory. Knowledge processed through the human brain, when then input to AI, yields much higher quality than just throwing in a bunch of raw data.
  • Handwriting was also one of my ways to slow down this year. (Another way was having good meals with family and friends, eating mindfully. Bringing only a camera, not scrolling on my phone.)

(Caption: Taking handwritten notes with just the reMarkable Paper Pro during AWS re:Invent 2025 CEO Keynote.)

(Caption: Handwritten notes helped me go on a podcast with Jayson and Terry just an hour and a half after finishing the keynote.)

2.3 Integration Thinking for Knowledge Base and Workflow Orchestration

In practical work, I continued integrating knowledge management and workflows together.

  • Take Kiro’s Steering feature as an example. It allows development teams to write “team knowledge”—such as project architecture standards, design patterns, preferred libraries, and naming conventions—into files under the .kiro/steering/ directory. Before executing any task, AI will first reference these “guiding principles.”
    • This can be viewed as a kind of “Executable Architecture Documentation.” Past architecture documentation often became disconnected from actual code—documents said one thing, implementation was another. But when these documents directly influence AI behavior, they gain real binding force.
  • I stack various principles into these guidance files:
    • “Between Chinese and English characters, please add a half-width space”
    • “If you have concerns, ask me questions to clarify things I haven’t explained clearly”
    • “Please handle every detail completely, don’t skip, don’t omit”
    • Amazon Working Backwards thinking framework
    • Domain-Driven Design (DDD) design principles
  • These seemingly trivial guidelines accumulate to form a “team DNA.” When new people join (or when new AI Agents join), they don’t need word-of-mouth transmission—just reading these files helps them understand how the team works.
  • Next, I should gradually organize these into Skills.

(Caption: While listening to the AWS re:Invent 2025 CEO Keynote, Kiro was playing peek-a-boo in the corner of the projection wall!)

2.4 Three Pillars: People, Workflow, System

When sharing this framework with clients, I always emphasize: this is just an example template. You can modify it to fit your scenario, your experience, to whatever feels comfortable. The point isn’t to copy the framework, but to use the framework to guide thinking.

  • When implementing Humbled Productivity (土炮工作法), we use the Humbled Transformation Framework. The core design of the entire framework rests on three interconnected pillars:
    • People: Human-centered, solving problems, developing strategy, accountable teams
    • Workflow: Daily work, investigating details, breaking down steps, fact-based
    • System: Stepping back, seeing the big picture, integrating processes, creating value
  • All three are essential.
    • Only People without Workflow becomes “all talk, insufficient execution.”
    • Only Workflow without System becomes “missing the forest for the trees,” lacking feedback balance mechanisms.
    • Only System without People becomes “armchair strategizing, disconnected from reality.”

(Caption: Sharing the Humbled Transformation Framework in Hong Kong in May, exchanging ideas with developers and managers in attendance.)

2.5 Kyklosify Service Thinking

The practice accumulated over these years has gradually converged into the service framework of Kyklosify, currently releasing gradually as a side project to explore sustainability and problem-solving together with clients.

  • The name Kyklosify comes from the Greek “Kyklos” (cycle), representing the continuous cycle of deconstruction, understanding, integration, and emergence. This isn’t a one-time transformation project, but a process of continuous iteration, continuous learning.
  • The core thinking is “exploring essential workflows”, not just “implementing tools.”
  • According to reports released by McKinsey and Accenture in early 2025, 90% of large organizations plan to increase AI investment, and 80% are already using AI in at least one functional area. But only about 20% of organizations have actually redesigned workflows to implement generative AI.
    • This means most organizations are just treating AI as “a faster tool” without thinking about how to change the way they work. It’s like buying a treadmill but only using it to hang clothes—no matter how good the tool, if the usage is wrong, the benefits will be greatly diminished.
    • What our team considers “right” often involves, after layer upon layer of deep digging, returning to the company’s founding mission, combining current market parameters, fine-tuning the organization and workflows, and confirming that real pain points are being addressed. Throughout the process, we ensure that most work tasks can be called via API, ensure historical information is retained, paving the way for upcoming Agents and Digital Twins.
  • Worthwhile further reading:
    • Dcard CEO Kytu mentioned in his year-end review three obstacles for Agents entering organizations:
      • Tools: It’s not that AI isn’t smart enough, but that organizations haven’t given it “hands and feet to do things.” Beyond API connectivity, metadata, MCP, and Skills, what’s most often overlooked is the definition and implementation of governance that most Taiwanese enterprises lack.
      • Permissions: Needs a governance mechanism that allows comfortable delegation. This is one of the more challenging aspects in our communication process, which is why we can currently only choose clients who have the authority to adjust permissions.
      • Platform: Lowering the barrier to use, letting every team member (human) use and participate, integrating workflows and organizational knowledge into daily work.
    • Karpathy’s year-end review = 2025 LLM Year in Review

Kytu’s three points align almost perfectly with the challenges we encounter at our own site and at client sites. Technical problems are often the easiest to solve; the truly difficult parts are organizational change, permission design, and cultivating (or adjusting) usage habits.

Another viewpoint I strongly agree with is the role positioning of FDE (Forward Deployed Engineer), and in the Agent era:

“The scarcest thing isn’t the speed of getting things done, but knowing which direction to move forward.” – Kytu, Dcard CEO

In the Agent era, the most valuable talent may not be those who code the fastest, but those who can best “deconstruct and restructure requirements into executable workflows.” This capability isn’t limited to engineers—product managers, business partners, finance partners, and administrative partners can all cultivate it.

(Caption: About digital twins, I’m still imagining… The photo released this year that’s closest to digital twins is below. Guess—which side is reality? Who is whose reality? Or is all of this just symbols for conscious entities’ communication?)

3️⃣ The Balance of Health and Life: The Ritual of Slowing Down

3.1 Physical Health: Building a Weekly Fitness Habit

After talking so much about technology and work, one of the changes that actually affected me most deeply in 2025 was the weekly weight training habit I built starting in February.

  • At least once a week, rain or shine, weight training time became my “forced reboot” mechanism. No matter how late I worked the night before or how much project pressure there was, when training time came, I put everything down and focused on how my body felt. This “deliberate disconnection” actually made me more efficient at other times. When the body is exhausted, forcing yourself to work only produces low-quality output. Better to rest well, exercise well, and face challenges with more energy.
  • After working out, I’d have a big salad, bring the Paper Pro or iPad to handwrite and sketch some ideas, or exchange opinions with AI Founders Meetups friends. Although a lot of feedback said that at my age, once a week is too little, my small goal this year was to build an exercise habit. Even when traveling, I made sure to visit the hotel gym (I absolutely won’t say I only went there to refill my water bottle XDD)—I’m happy that I achieved the minimum standard of building an exercise habit.
  • Although I had a trainer watching to adjust and confirm exercise posture, the side effects of a sedentary profession caught up with me. After a Q2 business trip involving four hours of train travel each way plus walking around with a backpack, upon returning, while bending to pick up my backpack, I heard a “click” and experienced what it’s like to “throw out your back”… I won’t go into details, but thanks to W’s careful care and everyone’s recommended rehabilitation methods, I started paying more attention to adjusting desk and chair height, sitting posture, and duration.
  • After recovering from throwing out my back, in Q4 I almost wrecked a certain muscle on one shoulder from simultaneously carrying a Macbook Pro and Paper Pro on one shoulder. Fortunately, it was caught at the inflammation stage when I asked about it and received early care. Almost couldn’t travel and run my schedule—really grateful.

(Caption: Post-workout coffee ritual, um, this photo doesn’t show the coffee, but anyway, you understand XDD)

3.2 Mental Health: Reading and Consciousness Exploration

Among the books I read this year, the one that influenced me most deeply and that I loved most was Federico Faggin’s “Irreducible: Consciousness, Life, Computers, and Human Nature”.

  • Faggin is the legendary inventor of the microprocessor, but this book isn’t about technology—it’s about the nature of consciousness. From the perspectives of physics, information theory, and quantum mechanics (CIP), he explores why consciousness is “Irreducible”—meaning it cannot be reduced to pure computation or physical processes.
  • In today’s rapidly developing AI landscape, this book gave me a lot to reflect on. We easily equate AI’s capabilities with “intelligence” or even “consciousness,” but Faggin’s argument reminds me: the uniqueness of human consciousness may far exceed our current understanding (actually, I think we’ve never understood it, so it’s not really “far exceed” but more like “divided by zero”).
  • This isn’t to deny AI’s value, but to have a deeper recognition of “human” value. AI can process information, generate content, even simulate conversation, but “experience,” “feeling,” “meaning”—these may be things AI can never truly possess.
  • Recommended for those of you who are bombarded by information, social pressure, anxiety, or exhaustion but willing to take a small step to explore.
  • (I made an English-to-Traditional-Chinese translation for personal use. If you need it, have already purchased this book on Amazon, have a Kindle device or app, and we’ve met in person, we can discuss how to make it.)
  • Oh, and btw, during Black Friday I got an Amazon Kindle Colorsoft. I originally had no expectations, but after getting it and using it to read color magazines and make colored highlight notes, I quite liked it. Page-turning speed is also acceptable—it’s become my portable reading device.

Additionally, regarding sound, frequency, and hypersensitivity symptoms, I originally used noise-canceling headphones to alleviate them, but sometimes I still want to hear what the kids are doing (also they’re bigger and heavier, so unless I’m getting on a plane, I don’t necessarily bring them out).

  • Thanks to chatting about this concern with a friend while spending Thanksgiving in Seattle, they recommended loop earplugs, with different series corresponding to different noise reduction levels.
  • During Black Friday, I immediately ordered Loop Engage 2 Ear Plugs on Amazon, and fell in love with them immediately upon receiving them at AWS re:Invent.
  • Loop earplugs have a small hole on them so you can hear outside sounds. I wore them like this while attending sessions and browsing the Madison Square Garden-scale exhibit hall—very effective, no need to keep running to the reflection room anymore :)

(Caption: Figured I’d include a photo that should make everyone feel happy.)

3.3 Family and Friends

Looking back, these times are the most precious memories of the year. Work trivia gets forgotten, but time spent with family and close friends stays in the heart forever.

  • I love taking my family to explore various alleyways, or randomly picking a metro station to exit and wander around.
  • I absolutely love when family and friends take us on local itineraries during every outing—very grateful. Every gathering is cherished.
  • Actually, whether people are moving around or putting down roots, everyone encounters different challenges and emotional ups and downs. I really admire everyone cultivating their lives in various places—wishing everyone safety and health together.

(Caption: Plans couldn’t keep up with changes—the stone siblings got drawn away by the Sumida River riverside park.) (Originally stone dad wanted to take the whole family to a certain park on the other side of Sumida River to photograph the slide with Tokyo Skytree, but actually swings with Skytree isn’t bad either. The point is who you’re with, not where you are, right?)

3.4 Food and Kitchen

This hobby of playing with food (with respect) probably started when my mom taught me to fry my first sunny-side-up egg. First came the coffee experimentation during university years, then following mentors to visit French Bordeaux to taste grapes, exploring European cuisine and local snacks during self-travels and business trips, and in recent years more frequent visits to the US and Japan. With local family and friends leading exploration of local cuisine, I’ve gradually constructed West Coast Mexican cuisine and Italian-French cuisine in Japan. Bringing these flavors and textures back home to try cooking for family—that’s why you sometimes see hashtags like #kitchenlab or #softpluschrunchy in my photos.

(Caption: A bit sad that this “life goes on” magnetic plate that accompanied WE+we through meals and food photography during the pandemic years accidentally broke this year…)

3.5 Deep Gatherings

This year continued last year’s adjustment: 80% of time allocated to gatherings that allow deep exchange or one-on-one deep conversations, 20% of time participating in and observing gatherings for new domain information.

Monthly fixed gatherings cover the domains most needed for myself and work: Product, Technology, Knowledge, Infrastructure, AI.

  • Product People monthly meetup
    • A seven-person squad—this group is the strictest haha. We hold each other to rules of no absences and almost no new members. We’ve recorded over 25 consecutive gatherings.
    • Because everyone has certain aligned attributes, is super self-disciplined, and we have the longest time (we only set a start time, no preset end time), this is currently the deepest in discussion depth among all gatherings.
    • The format is somewhat similar to FounderSquad from several years ago—each person brings one question each month, and everyone discusses.
    • I highly recommend everyone form such small mutual discussion squads.
  • TGONetworks monthly dinner
    • Although sometimes jokingly called the CTO warming-up group, members from various industries and the entry threshold set before (e.g., department size over 15 people) make the overall discussion depth and experience value comparable, saving a lot of communication wait time and allowing direct problem-hitting.
  • AI Founders Meetups
    • Started with several old friends from the open source community mutually scheduling (reminding each other) weekly exercise. After exercising, check in on the group chat to encourage each other.
    • Gradually physical meetups emerged. Sometimes we meet at the gym—chat about recent progress after exercising. Sometimes at a coffee shop—one word leads to another and laptops come out, wait no, I mean we open laptops for live demos.
    • For example, demonstrating that with complete documentation, in 60 minutes, from zero, vibe coding a research institution’s static website, then connecting Cloudflare DNS and deploying to GitHub Pages. (Let’s not mention the years of blood and tears behind those 60 minutes…)
    • This group of founders has a wide geographic movement range. One time we also met up in Tokyo. Someone went to Europe to establish a company. Someone has experience operating engineering department offices in multiple regions. I roughly shared the experience of establishing and operating a US company.

(Caption: After meeting at BELP in 2013, hitting it off immediately, and then not properly sitting down for a meal and chat for over a decade, Richard, who just joined US-based Datastrato as VP of Ecosystem responsible for global open source ecosystem business, happened to both attend AWS’s annual developer conference re:Invent 2025 in Las Vegas, and we met up for a long chat at the MongoDB Private Lounge.) (Being able to have a long chat with Richard is also a kind of privilege XDD)

✳️ Conclusion: The Next Turn of the Cycle Continues

Looking back at 2025, if I had to summarize it in one word, it would be deliberate “Cycle (Kyklos).”

Deconstruct, understand, integrate, verify—then cycle again.

From the evolution of technical tools, to the systematization of knowledge management, to the establishment of life habits, every domain is going through such cycles. There’s no once-and-for-all answer, only the continuous process of iteration.

2025 helped me realize and review:

  • Tools change, but foundational capabilities don’t become obsolete.
    • Spec writing, systems thinking, cross-domain connection—these capabilities have value in any era.
    • But time is limited, so choices must be made—gaining these capabilities within limited time and accumulating experience to gain dimensions and parameters.
  • Efficiency is important, but direction and decision-making are also important.
    • Contemporary AI (LLM) can make us do things faster, but it also increases verification and inspection time.
    • Whether efficiency can be improved actually depends on the architecture, planning, and directional decisions before calling AI.
  • Slowing down isn’t wasting time.
    • Finance has TCO (Total Cost of Ownership) and ROI (Return on Investment) metrics; time does too.
    • Deep thinking, deliberate rest, quality companionship—these “slow” investments yield more lasting “trust and connection.”

What we saw in 2025 may just be the prologue of the AI era. In the coming years, the pace of change will only accelerate, the scope of impact will only widen.

Besides avoidance and various forms of lying flat, perhaps at this stage, I would still choose to face the flow of the current world with the approach of “Slow Down to Go Fast.” Maintaining clear thinking, clear direction, and steady rhythm in the accelerating torrent—perhaps that’s a form of competitiveness?! At least with the company of family, friends, and amazing teammates throughout the journey, strength will beget strength :)

That’s it, finalized, sending. Stop overthinking—reviewing 2025 is for taking solid steps forward, trampling 2026, oh wait, I mean, stepping into 2026!

  • 2026, I’m not chasing — I’m choosing.
  • Sometimes you’ve got to slow down to go fast, and let the cloud fireworks remind you, magic is already there.
  • Less noise. More signal. Same curiosity.
Cloud Fireworks, Magic, at Disney