Writing

To Garden This World from a Wasteland

On rebuilding my personal website as a place for writing, photography, daydreaming, and letting thoughts wander.

To destroy, and then to reestablish. This is the only principle of moving forward, as in software development, as in personal growth. And this is what I’ve decided to do to this website again, just three weeks after the last round of reestablishment.

The last reestablishment, ridiculous as it seems to me now, was such a quick, purposeful thing. I quickly organized all my things and quickly worked with ChatGPT to put everything on the site, eager to show someone that I had something. The show was a total failure, as it was destined to be from the very beginning.

It was a failure by my own measure, in my own struggle between fawning and freedom. Both begin with “f,” but they have totally different endings. How did I, someone who pursues freedom and has the talent to create things in the virtual world, end up building a fawning room? A room full of things I chose or made or wanted to show someone, hoping they’d be happy to see them and think well of me. My own room, my own space?

I’ve always dreamed of having a garden. A garden, a garden is a place for creativity, for flourishing, for things to grow as they will. As Yiyun Li writes, “Things in nature merely grow.”[1] I can’t have a garden right now. I can build a garden here on the website for myself. This is possible. So how could you call it anything but a failure that I turned the garden I dreamed of into a store selling things from elsewhere?

It feels like a place for everything but me. I forgot my instinct as a writer. By writer, I mean someone who cannot stop writing, rather than someone who writes perfectly. My instinct as a photographer, who takes photos to fight against time and everything passing away, rather than to produce great photos. My instinct as a daydreamer, who dreams up interesting things every minute, rather than valuable ideas. My instinct as a thought wanderer, who connects everything out of curiosity, rather than trying to focus deeply enough on one thing to become famous.

I, unfortunately, left no space for that part of myself on this website. So I can only clumsily and insincerely work with ChatGPT to produce more content until the day I don’t want to produce or pretend anymore. So this is only a store, a job, a dead place.

So I made a decision. To garden this world from a wasteland.


  1. Yiyun Li, Things in Nature Merely Grow (Farrar, Straus and Giroux, 2025). ↩︎

Writing

AI assistant and conversation data

The guide is an AI character, not me. Opening interactions such as the die use authored public material without calling a model. The chat status indicates whether live conversation is available.

When live conversation is enabled, sending a question submits it and the current article identifier to this site's backend. The backend sends recent conversation, an index of earlier exchanges, any recalled records and retrieved public material to Alibaba Cloud Bailian to generate a response. Cloudflare Turnstile verifies visitors when a session is created; no account is required. Do not submit secrets or private information. The model and verification providers' data-processing policies apply.

The current browser tab stores the last ten interactions in session storage to support returning between reading and chat. Clearing the conversation removes those local records. Server sessions last one hour and retain up to twenty completed exchanges. The latest three are included directly in context; earlier records are retrieved when needed. Expired records are cleaned up on subsequent requests. “Start over” abandons the current session but does not immediately delete its server records. Across HTTPS pages in the same tab, session storage holds a one-hour session credential so the backend can restore recent context and an index of earlier exchanges. This is not the model provider’s API key. Expired credentials or unavailable storage require verification again; displayed history is not automatically imported into a new session.

The server also retains quota, budget and short-term request-limit records; connection identifiers are hashed with a secret key. The application does not keep additional question logs, but the session store contains the conversation text described above. Failed model requests produce operational records containing only a random request ID, fixed error category, processing stage, call count and whether partial output was shown; these contain no questions, answers, IP addresses or credentials. The AI's tools access public site material, not private files, personal chats or credentials. Citations do not guarantee that every sentence is correct: check the original sources and distinguish my views, quoted views and model inference.

Hi.

An AI character, not Yuxin herself. Opening interactions are available; live conversation is not connected yet.