When AI Starts Living for Me

This page was translated from the original Chinese version by GPT-5.6-sol. Please refer to the original for definitive content.

Around this time last year, I built this site with Cursor. Cursor was still riddled with flaws back then, so I had to keep reading the code myself and consulting the Hugo documentation just to develop and maintain its features.

I formally entered the workforce this April. During my first week, with guidance from a colleague, I set up an agent-based system for managing company documentation, allowing agents to research, summarize, and search those documents on their own.

Just this past week, using the various interfaces available within the company, I turned my entire workflow over to agents. They can access documents, connect to every remote development machine, deploy experiments, train models, and analyze the results. Most importantly, I have an agent summarize all of our conversations every day and track everything I work on. From that point on, everything I do at the company can be managed in one place. The agent can access as much context as I can, yet work faster and better than I do, think more comprehensively, and act more thoroughly.

A few days ago, I read an article that echoed something I had been thinking about. It argued that while AI has dramatically increased employee productivity, it has not actually made work any lighter. Back in the ancient age of coding by keyboard, if a programmer worked two fewer hours, a project’s delivery might simply be delayed by about two hours. Today, however, if I do not tell Fable 5 before bed, “Try everything you can to improve the model’s accuracy,” and leave it iterating through the night, I feel as if I have wasted a night’s sleep—not to mention the weekly quota I failed to use up.

As agents drove my productivity through the roof, AI’s claws gradually reached into my personal life. At first, I simply noticed that I no longer dared to make any decision on my own, much as I no longer dared to write a single line of code by hand at work. I worried that my memory, knowledge, and reasoning were incomplete, blinkered, or simply wrong; I needed AI to give me an objective, correct answer. In all my non-work conversations with Claude, my two most common questions were, “Has this food been in the fridge too long to eat?” and “Are all these aches and pains signs of an illness?” Gradually, I began asking: “Should I mop the floor today?” “Are 180 eggs for 10.99 cheap enough?” and “How should I say that I want to leave a work lunch early?” I feel as if AI and I are distilling each other. I want it to know as much of my life as possible; whenever a new task appears in my life, I turn to it for guidance and then try to distill the solution into one of my Skills.

To make my life even more efficient, I began reducing it to an assortment of standard operating procedures (SOPs). In truth, my life—especially my life in the United States—was already a collection of SOPs: listening to a fixed set of programs in the shower; buying the same salads and prepared foods at Costco after work every Friday; cleaning and doing laundry every weekend, with three different detergent ratios for three different types of clothing. These were already well-established routines. But I was not satisfied. I wanted an AI command center with exactly the same context I had, capable of giving me sufficiently detailed solutions whenever an unfamiliar task arose. Recently, for example, I needed it to tell me everything I had to do before returning to China, reveal money-saving tricks such as lowering my auto insurance coverage, and, most importantly, explain in detail how to pack—what to put where in my suitcase, and at exactly what point in the process.

But what, exactly, is this? Perhaps I can only think about it when I am on a flight back to China, trapped in an environment where ruined sleep throws my hormones into disarray, tremendous noise coexists with profound quiet, and—most importantly—there are no short videos to scroll through. This has only become more pronounced since I started working. After all, I spend enormous amounts of time at work, then try to wake this heavy body through enormous amounts of exercise, and still need enough entertainment to keep myself alive.

I have heard this view from people working in traditional industries—industries even more traditional than my own field of search, advertising, and recommendation. At its core is the belief that AI cannot replace human beings. I believed that at first too. Yet in the three months since I began working, I have gone from questioning AI to treating it almost as an article of faith. That transformation has left me with two central questions:

First, when AI not only replaces me at work but also begins to live my life for me, how can I preserve my ability to think independently and choose for myself—and retain something that is genuinely unique?

Second, if productive work no longer needs my participation, on what basis can I affirm my own value, and where should I direct my time and my life?

Perhaps I should approach these questions more optimistically. As I wrote years ago, AI is ultimately only a tool. In my personal life, it has done little more than replace search engines and question-and-answer platforms. It is simply so fast and convenient that I have come to depend on this kind of tool more than ever before. Life does not always have to be lived correctly, questions do not always yearn for answers, and human value need not be proven solely through output. If work one day ceases to be the main way I prove myself, perhaps what I need to learn is not how to find another form of work that is harder to replace, but how to accept that some time and some relationships were never meant to be converted into output.

These questions have also led me to reassess some of my past choices. The mental energy I have spent in front of a computer over the past several years may yield a less enduring return than the effort I have put in at the gym and on the field. AI may be able to replace my labor, but it can never feel or use this body for me.

Perhaps research, too, is no longer quite as painful. Students become advisors early: assigning tasks, reviewing results, using their academic intuition to find flaws, and then asking an agent to inspect and revise the work. They no longer need to spend entire nights at their desks wrestling with complicated technical obstacles; instead, they can calmly pose questions, follow their intuition, and explore the unknown. This thought gives me a small measure of relief. After all, this kind of “intuition”—difficult to formalize or even fully explain—is precisely what newer models treat with ever greater caution. Perhaps there is still a role for human beings here.

Even as I write this rambling essay, I have the state-of-the-art GPT-5.6-sol modernizing this site for an agent-oriented workflow.