THE MOTIVATION
✧Make useful guidance easier to access.
I wanted people with fewer resources, less experience, or limited access to information to have a tool that could help them understand situations, reflect, and make better-informed decisions.
✧ SELECTED WORK / 03 — HUMAN × AI A REFLECTION ON DESIGN
XIAOBING NOTES / ORIGINAL APPROACH DISCONTINUED
An assistant
that knows when
to step back.
I wanted AI to make useful information and guidance accessible to more people. Through development, I came to see the danger in constant recording and AI advice. Today I am exploring whether personal understanding could be learned in a less intrusive way — without surrendering human choice.
ORIGINAL APPROACH STOPPED · RESEARCH CONTINUES
01YOUJUDGMENT
02AIASSISTANCE
✳
FROM INTENTION TO UNINTENDED CONSEQUENCESCROLL TO EXPLORE ↓
01 / THE QUESTION
“If an assistant helps me with every decision, am I still strengthening my own judgment?”
02 / HOW IT UNFOLDEDAN HONEST PRODUCT RETROSPECTIVE
My understanding
changed through building.
The chronology matters. My June 4 note was an early, exploratory thought written when I had only just begun developing Xiaobing Notes. The deeper concern about AI's potential harm emerged later, through the process of actually exploring the product.
JUNE 4, 2026 / AN EARLY THOUGHT
⌁The Whispering Earring.
At the very beginning of development, I wrote a short reflection inspired by a story about an assistant that could guide every part of a person's life. At that point, I was only beginning to consider the idea. I had not yet reached the understanding I would develop later.
Read the original entry and its English translation ↗BUILDING THE PROTOTYPE
⌘Explore an assistant that stays close.
I began experimenting with recording, transcription, organizing personal context, and context-aware suggestions across mobile and desktop. The prototype was an exploration, not a finished public product.
WHAT I REALIZED LATER
◌Constant assistance could become harmful.
As I explored the product more deeply, I began to recognize that continuous recording, analysis, and guidance could cross important boundaries. An assistant that is always observing and interpreting a person's life may intrude on privacy, influence decisions too heavily, and leave less room for independent judgment. The risk was not simply an inconvenient interface; the approach itself could be harmful.
THE CHANGE IN DIRECTION
■Stop the intrusive direction. Keep the question alive.
I paused the more intrusive direction of development and reconsidered what a healthy relationship between people and AI should look like. I have not abandoned Xiaobing Notes. I may return to it when I can build an assistant that offers useful help by invitation, respects personal boundaries, and gives people room to think and choose for themselves.
03 / AN OPEN PRIVACY PROBLEMRESEARCH IDEA / NOT YET IMPLEMENTED
Can an AI learn the person
without keeping the diary?
The earlier product relied on the idea of collecting and organizing personal context. I am now interested in a different direction: long-term personalisation in model parameters, rather than treating a permanent, searchable record of everyday life as the answer. This is a hypothesis to investigate, not a privacy guarantee or a working feature of the prototype.
Minimal, temporary context.
Only collect what the person knowingly permits, for a clear purpose, rather than treating constant observation as the default.
Learn preferences, not a life log.
Could a separate, locally controlled model adapter gradually reflect user-approved preferences and habits without needing to retrieve all the original conversations?
The freedom to disagree.
Let the user inspect, reverse, reset, or stop adaptation. Useful suggestions must never be treated as commands.
PERSONAL AI / ARCHITECTURE EXPLORERCONCEPTUAL COMPARISON · NOT A DEPLOYED SYSTEM
Recorded life data→External database→Retrieval→Model response
STRENGTH
Explicit, inspectable records.
When records are accurate and indexed, the system can retrieve specific events, dates, and original statements. Users may be able to inspect and delete the stored records.
RISK
A persistent record of private life.
Encryption and access control help, but do not remove the underlying concentration of sensitive information or the risk of over-collection, misuse, or breach.
User-approved patterns→Personal adapter+Base model→Model response
POTENTIAL
Personalisation without a full diary.
A per-user adapter might represent stable preferences, communication style, or recurring needs. It could reduce reliance on retrieving raw conversations for every answer.
RISK
Parameters can still encode private information.
Weights are not inherently private. They may memorize sensitive examples, can be difficult to inspect or selectively edit, and may be hard to make reliably forget. They are also not a dependable store for exact episodic recall.
User-selected detail→Temporary context→Single response→Discard
STRENGTH
Purpose-limited processing.
Information can be provided for a specific task rather than accumulated as a permanent personal history.
RISK
Less continuity, not zero risk.
The person may need to provide context again. Temporary processing still needs secure handling, clear consent, and a verifiable retention policy across every service involved.
Research boundary: these are design directions to investigate. I have not implemented a personal-weight system or proven that it protects privacy. Any approach would need threat modelling, extraction testing, user controls, deletion and reset tests, and independent review.
THE QUESTION I CANNOT ANSWER YET
What should an AI be allowed to know — and keep?
Personalisation is not only a model-quality problem. It is also a question of consent, ownership, security, correction, forgetting, and the freedom to stop. I want to understand these trade-offs before building another system around them.
The unresolved risk.
Information can leak from trained weights or personal adapters even if the original recordings have been deleted. This approach would still require serious work on consent, privacy evaluation, forgetting or reset mechanisms, and security. I do not yet know the right solution.
WHY I WANT TO LEARN FROM OTHERSI can build a prototype.
I can build a prototype.
I cannot solve this problem alone.
I admire what researchers and engineers at organizations such as OpenAI and Meta have built and learned. They are teachers to me, not collaborators on this project. I want to meet people who understand these problems far better than I do, question my assumptions, and work toward an approach that expands access while respecting privacy and human independence.
04 / INTERACTIVE PROTOTYPELOCAL DEMO / SAMPLE DATA
Explore the prototype.
See the original direction.
The supplied Xiaobing Notes prototype is embedded below, including its original mobile and desktop reference screenshots. Explore the navigation, recording cards, sample memory search, tasks, device and backup screens, and screenshot gallery. This is a local-only interface demo, not a connected recording product. The prototype documents an exploration, not a finished public service.
05 / WHAT I LEARNEDPAUSED, NOT ABANDONED
THE DESIGN PRINCIPLE
Be present when useful.
Step back when needed.
I still believe intelligent guidance can widen access to information and opportunity. But I am no longer willing to assume continuous observation is a fair price to pay for it. Xiaobing Notes is paused while I reconsider the privacy architecture, the degree of intervention, and how users can remain free to disagree or walk away. The project may restart when I can make those boundaries real.
Read the homepage reflection ↗LESSON / 001
MORE
ABILITY.
ABILITY.
NOT LESS
AGENCY.
HUMAN-CENTERED TECHNOLOGYAGENCY.