This year concludes with “less is more.”
In product development, “thinking clearly and doing subtraction” is a form of less; in decision-making, “methodology doesn’t need to be numerous—even the most basic what, why, and how work remarkably well.”
Lack of focus leads to multiple simultaneous projects, multiple books being read, multiple games in progress… making some “completions” become abandoned halfway through, and more seriously, deceiving oneself into no longer needing reflection and appreciation.
Being too “greedy” and always choosing local optima may appear to preserve multiple possibility spaces for the future, but ultimately leads to overfitting—fast local responses but lost global objectives.
In 2025, from 0 to 1, I built multiple projects with AI, exploring the evolution and boundaries of applying SOTA models to coding agents. I developed personal workflows and insights, having clearer expectations for delegating certain tasks to AI and the ability to further improve completion quality.
Practicing from a coding perspective in the AI Native era, uncompromisingly keeping up with the times. In the first half of the year, I read several books such as “Python Deep Learning” and “Building Large Language Models from Scratch,” gaining a more rational and realistic understanding of the current AI represented by LLMs. Combined with NEXUS and AION, my understanding of agents became more grounded.
Because I’ve always been somewhat fascinated by abstract and philosophical musings, immersed in fictional worlds in my mind. Undoubtedly, throughout this year, I’ve appreciated many classic sci-fi films and sci-fi series. Coincidentally, standing on this current timeline, we have “sci-fi” AI, giving me some wonderful feelings while watching and secretly feeling fortunate. These include “2001: A Space Odyssey,” two “Blade Runner” films, “Memento,” “Arrival,” “Source Code,” “Westworld,” “Black Mirror,” “Severance”… Or rather, because of these films, my thinking and exploration of “intelligence” continued. Added to reading Jeff Hawkins’ “The Age of Intelligent Machines” and subsequent “A Thousand Brains,” along with some cognitive science and psychology books, overall, it felt very “sci-fi.” Fortunately, I didn’t indulge in intangible musings. I also realized and felt that subjective feelings and reflections after practical practice are often more concrete compared to random combinations of thoughts and abstract thinking.
What needs continuous practice next:
-
Don’t let “exploration” become an escape; exploration is not suspending judgment. Need to regulate “exploration”—what to explore? Why? Where are the boundaries? What are the exit conditions?
-
Don’t set limits, but also have clear objectives. Complex and profound things can be mastered through scientific methods. If difficult to comprehend, try a different method or textbook.
-
Don’t have too many methodologies; firmly remember the simplest ones and internalize them.
-
Use profound thinking to guide actual actions, then use feedback from actions to refine and deepen thinking.
今年以“少即是多”收尾。
项目产品中，“想清楚，做减法”是一种少；思考决策中，“方法论不追求多，甚至最基本的是什么、为什么、如何做就很好用”。
没有聚焦而使得多项目推进、多书目阅读、多游戏并进…让一些“完成”变得半途而废，甚至严重的是欺骗自我不再需要、思考与鉴赏相关。
太过“贪心”而总是选择局部最优，看上去在面向未来保留多个可能性空间，却最终过拟合了，局部反应快而全局目标迷失。
2025 这一年从 0 到 1 的与 AI 共建多个项目，探索体会 SOTA model 应用于 coding agent 的能力演进与边界，有一套个人的工作流与心得感受，对于某些任务委托于 AI 有较清晰的预期与进一步提高完成度的能力。
AI Native 时代的、从 coding 出发地践行，毫不妥协的跟进时代，上半年阅读几本书，如《python 深度学习》与《从零构建大语言模型》对于当前 LLM 代表的 AI 有了较为理性现实的认知，同时结合 NEXUS 与 AION，让我对于 agent 的认识更加落地。
缘由我一直对于抽象与哲思性的遐思有些着谜，沉迷于脑中的虚构世界，无疑，在这一年，鉴赏许多经典科幻电影与科幻剧，巧合的是，站在当前这个时间线上，我们有了“科幻”的 AI，让我的观看时总有些奇妙的感受并暗自庆幸，它们是《2001 太空漫游》、两部《银翼杀手》、《记忆碎片》、《降临》、《源代码》、《西部世界》、《黑镜》、《人生切割术》…或者说我也因为这些电影而对于“智能”的思考与探索也是不断，加之阅读到杰夫霍金斯的《智能时代》与后续的《千脑智能》、与一些认知科学、心理相关是书目，总的来说是倍感“科幻”，幸运的是，我并没有沉迷于虚无缥缈的 musings 中，我也意识并感受到，很多时候主观感受、实践践行后的反思所得与每每想法的随意组合与抽象的思考对比，前者是更加具体的。
需要接下来不断践行的：
-
不要让“探索”成为逃避，探索不是悬置判断。需要规范“探索”，探索什么？为什么？边界在哪？退出条件？
-
不要设限，但也要明确目标。复杂深奥的事物是可以通过科学的方法掌握，如果难以参透，换一个方法、教材试一试。
-
方法论不要多，牢记最简单的并内化。
-
用深邃的思考指导实际的行动，再用行动的反馈来修正和深化思考。