October 2026 Thought Notes: Daily Life, AI and Agent Systems, Engineering and Open Source

Group chats are essentially a way to categorize

Group chats are essentially a way to categorize

town’s Desktop setup

Very elegant: starting and configuring each connection, little by little

The setup involves a long sequence of connections, so it gives users feedback a little at a time

Nango’s free allowance is too low

It only supports ten connections!!

When Grandma learned Musheng had died in 1960, her first words were about Xie Nanzhi raising so many

When Grandma learned that Musheng had died in 1960, the first thing she said was how hard it must have been for Xie Nanzhi to raise so many children

Across vast rivers and seas, thinking of you makes the distance disappear

The kapok flowers blooming beside the Chao Phraya River look so much like spring back home

I pressed one into this letter, hoping you too could smell its fragrance

The mortality rate was relatively manageable

In 1940s Chaoshan, Shantou–Bangkok was already an established steamship route that had operated for decades. It was no longer the early period of mostly wooden sailing ships, or even being packed into ocean-going holds as “piglets”

Cholera was generally the biggest problem at sea

There had been an earlier period of human trafficking

From 1847 to 1874, concentrated especially along the Guangdong and Fujian coasts, large numbers of young Chinese men were shipped through ports such as Macau, Xiamen and Hong Kong to sugar mills, plantations, railways and guano mines in Cuba, Peru and elsewhere

A man in his twenties from a village → deceived with promises of making a fortune overseas → taken to Macau → locked in a “piglet” holding depot → forced to sign an eight-year contract → put aboard a ship → months crossing the Pacific/Atlantic → arriving in Peru or Cuba → handed over to a plantation or mine.

Hundreds of people spent months sailing in the lower decks of wooden ships with no air conditioning, no modern toilets and extremely poor ventilation. The risk of death was high; one or two would die

Perhaps people who have been through a few things come to believe in fate

I think people who have been through a few things, more or less, may come to believe in fate. Its push and pull feels so strong, and we also need fate to give us some support, some strength. We never know whether the path we have chosen is right. So it is perhaps just a little consolation through xinxue, the School of Mind.

A very interesting film made up of eight independent shorts

I think this may have been Akira Kurosawa’s own destiny. He made it himself, and every segment begins with “I had a dream like this”; supposedly they came from his own dreams

The film follows the same person, from childhood to old age, through eight connected dreams. He made it when he was 80

Then look at why this film is so visually beautiful. Although it does not have many special effects, or a continuous plot across the whole film, it is mostly animated colors and atmosphere, like a painting, like a painting that moves

But it is so puzzling: the images are beautiful, and every frame can be viewed as a painting. This is Kurosawa’s most personal kind of work in his later years: his reflections on war, on nuclear energy, and on the relationship between people and nature. There is also the red Mount Fuji, giving it the feel of a fable

Users remember contacts rather than conversations

Users remember contacts rather than conversations

Some thoughts on rsi

Developing more powerful models to improve the methods and capabilities involved in developing models

But the central problem is that better and better results do not necessarily mean greater intelligence

The results are indeed better, the cost lower, and the speed faster

I keep wondering whether the world we live in is real

It seems that ever since we came into the picture, our world has changed incredibly fast. Starting in childhood, electricity and running water arrived one after another. Later we had airplanes, the internet, and smartphones, and now artificial intelligence has appeared. Perhaps in another 10 years, we really will enter a virtual world and experience an immersive life, like in The Matrix. As world models develop and evolve, and 3D, VR and AR mature, we really may reach that point

Nintendo games are similar: millions of players immerse themselves in them, essentially participating in a lightweight simulation. Super Mario runs across the screen; his world may not exist in its entirety, but it takes effect as he moves forward. The space behind a wall is simulated only when you walk behind it. It is like our world now, extending onward with time as a variable, gradually revealing new views

Perhaps we are living in a simulated world like this. This world might be false, yet the world as we see it might be real. Looking at it this way seems to dissolve many of life’s pressures. The world is a simulated game, so what we can do is live earnestly within it, recognize what belongs to fate, and avoid becoming too tense. When a challenge comes, treat it as a level and concentrate on finding a solution! It really brings a great sense of ease. Live earnestly, love your work passionately, and do not take yourself too seriously

Setting recurring calendar rules goes against human nature and constrains people

Every time, I do not want to look at those calendars I scheduled earlier

Each stage has its own problems. For me, setting calendar recurrences is acceptable only by month or year, by weeks with a defined lifespan, or by days with an extremely short lifespan

Always do the right thing for users

Understand users, protect their interests, and serve them

This is the most basic product ethics

Express yourself accurately

Do not waste anyone’s time or energy

Treasure your own time, every moment

You are free

Use it to say no

Use it to refuse

If a message will do, never call; if a call will do, never meet in person

If it can be left undone, leave it undone

There are two kinds of games in the world; one can be called a finite game

There are two kinds of games in the world; one can be called a finite game, and the other an infinite game

A finite game is played to win; an infinite game is played to keep the game going

There is one premise for a game, finite or infinite: unless the participants join voluntarily, there is no game

A very central point concerns the User input process

I think a very central point concerns the User input process. It should depend as much as possible on the User’s intent. If the intent is clear enough to yield a definite answer rather than a vague one, I think the process should be as lightweight and direct as possible. Ideally, input on macOS would be extremely convenient and quick. I think that would be very useful

In fact, I have developed a habit in my own use of macOS. Since macOS has voice input, I was thinking that its voice input generally requires pressing option, holding option to wake it up, and then continuously transcribing

For other macOS applications, such as note-taking apps, typing things out is actually quite at odds with how users think. To accumulate this raw material quickly, I think we still need to make the entry point as lightweight as possible, the lighter the better. That makes it a key capability

Acknowledging discomfort can coexist with accepting criticism

Acknowledging discomfort can coexist with accepting criticism

The essence of acting is forgetting yourself, including your own emotions

The essence of acting is forgetting yourself, including your own emotions. In that moment you become the character, with the character’s emotional changes

So to enter a role, you need to enter the character’s situation, but some people may hide their own selves so earnestly that they believe their performance, overlooking the difference between their emotions in the moment and their own feelings

All of a person’s social attributes are essentially roles

All of a person’s social attributes are essentially roles, each with its own limited strengths and weaknesses, but the self has no role

Very ironic

Many fathers and mothers are called selfish as soon as they think of themselves

As soon as they stop sacrificing, they are considered not to love their children enough. In this way, the identity of “mother” fills her entire being, while her own needs are not given an equal place in consideration

On AI product testing

unit test, integration test, e2e test, ui test, agent testing, regression test

Regression testing is a purpose of testing: rerun the tests from the past to ensure that existing functionality has not been broken

agent testing tests an AI agent. Prepare an Eval set and score it using rules or another model, an llm as judge

How do you prove the code is correct?

Eval / test-driven development: the most important thing is to define success criteria and constraints first

There must be regression tests, or an agent may introduce new bugs while fixing one

Testing AI relies on Eval

Separate the model that changes the code from the model that scores it

The testing standards themselves are a barrier. The standards and data are uniquely yours …. the most precious things

codex built an agent to help accountants handle tax filings

More than seven thousand tax returns, with accuracy as high as 97%

The agent can keep improving itself based on the mistakes it has made

Every time an accountant corrects a field the agent filled in incorrectly, that correction is recorded

When the same kind of mistake recurs, it is distilled into a specific problem and turned into a set of targeted tests (evals)

Have Codex inspect the code and determine which step went wrong: failing to extract information, incorrect mapping, choosing the wrong source file, or a problem with the scoring rules themselves. Then make the fix

After the fix, rerun the targeted tests and regression tests (to make sure the parts that worked before have not been broken). Once they pass, submit a PR for a person to review and merge

What this person called an evaluation harness plus self improvement means reproducing this loop within your own project

If the fixer can see how the grader scores, or even change the grader, it may learn to game the system

notion still needs a template

The rules should define how people use it, and how agents modify and write it

The more often people need to read and maintain a piece of information, the more carefully AI needs to adjust it

A question about AI batch tasks:

Can batch-generated work really beat something someone else has carefully refined for a particular feature?

Designing the onboarding part

Just the calendar and email, with a personal blog or social media information as optional additions: the data sources with the highest information density and the lowest barrier

Then deliver value, a personal summary. The first time I feel that town understands me so damn well, I am more willing to keep granting access

The recommendations are based on my data, on an understanding of me, and on its own mcp connection pool data to suggest things that suit me. The conversion rate is high

sset Scheduling: set some scheduling preferences and make recommendations to me

Then I can set some images and create my own townie

Finally, sharing: share townie to earn points

An additional note: which platform should host these Onboarding flows depends on the platform’s characteristics

For users arriving directly from the official website who have not installed the APP, Onboarding on an external page is more suitable

For those signing in through the desktop app, desktop Onboarding is more suitable, since it can connect to local MCPs or the local environment

Another point: if connecting through iOS, the initial Onboarding would draw on the features of the user’s iPhone, such as native iOS actions button screenshots and notifications

I want to focus a personal assistant on things that can compound over the long term

Focus on that feeling of growing together with AI

User stories should connect with evaluation

User stories should be connected to evaluation and managed together

User stories have two layers. The first is raw material, with as little friction as possible

notion should only store content that will be referenced repeatedly, so what is truly valuable maybe is:

User stories: things that really happened, valid scenarios, high-value scenarios. These can be reused for the website, the core flow, tests, and stories

The evidence layer: Eval, logs

Design: Figma is generally enough

Decisions: usually record what was chosen, why it was chosen, and under what conditions to overturn it (responsible for maintaining the agent system)

How to collect user stories:

Two layers. The first is raw material, with as little friction as possible: user interviews, screenshots of group chats, moments when I get stuck using the product, and screen recordings. Throw them all into a raw material area first, simply labeled with the date and source

The second is distillation. Each week, pick recurring stories from the material and make each one a row in the database, with the type set to “User story.” Use four fixed sections in the body: who, in what situation, what they want to accomplish, and where they are stuck now. Always attach a link to the original material at the end

Then comes evaluation, which is actually related to the second layer of user stories

Every valid user story corresponds to a set of eval cases. The story describes “what the user wants”; the cases verify “whether we delivered it.” In Notion, keep only the eval design and conclusions: what is being evaluated, the passing criteria, a summary of each run’s results, the corresponding commit, and the judgment reached. Keep the cases themselves and scoring data in the repository or eval platform, with only links in Notion. Once the volume of data grows, nobody will maintain it in Notion

The holiday is almost over; I went to a café today, but I feel busy every day

The holiday is almost over; I went to a café today, but I feel busy every day. When I think back through the day’s timeline, I can clearly tell where I was at each point and what I was doing. But when I look back over my day and try to summarize what I did today, I get a little stuck

It seems I did a lot, learned a lot, and solved a lot of problems. But looking back now, it feels as though I did nothing today

A day of busyness with nothing to show for it

I suddenly realized that Jobs, Musk, and the other familiar founders in technology must have incredibly busy working days. Jobs would practice time management or energy management. There are all sorts of methods and tools for managing time and energy online too. But in actual practice, I still find it troublesome, because management itself is a burden. If, during the day, I also have to think about how to list my itinerary and how to arrange my plans, that behavior itself consumes energy

Then I suddenly thought of Eason. In managing his daily life, he uses a really brilliant word, a really brilliant concept: bottleneck. It is a word you can keep thinking about. You can ask: what is the bottleneck? What is the bottleneck in the thing you are trying to do right now? It is very clear, a very specific thing and a very specific intention. Concepts are also the basic units of thought, so every day you can think outward from this concept and work around it. Then each day has a clear main thread. It is not sprawling or fragmented; it is one clear thread: solving the bottleneck in the thing you are working on now, and thinking about how to solve it. What methods do you use in the process? You focus all your attention, energy, learning and time on this concept

And a bottleneck is something you can keep thinking about. Once you solve the current bottleneck, you can keep asking: what is the next bottleneck? What is the next bottleneck?

I have suddenly had a realization. I think that in the future AI era, when facing complex tasks and events, especially as an individual who has a lot of work to do, the bottleneck is a very good method. It is more effective than all those energy management methods, time management methods and task lists. Every day you need to think clearly and define clearly: what is the current problem? What is the most central bottleneck? Keep thinking about and solving things around that bottleneck. Then every day you can clearly say what you did to address it

Continuing the previous note, I think the bottleneck solves another important problem

Continuing the previous note, I think the bottleneck solves another important problem: protecting your attention, time and energy for the single most important, most important thing

When there are too many things in everyday life, they really scatter your time and energy, until it feels as though you are forced to divide them just to solve or finish various tasks.

A bottleneck forces you to hold all your time and energy in your own hands, actively and selectively allocating them to the biggest and most important problem. Then you can put all your energy into this one thing for now, without cognitive or energy overload

For example, at my current stage, I would focus on the most central thing: whether the current evaluation or agent performance has been tuned well enough. If the Agent performs poorly, the central analysis is why. What is the bottleneck? Is it an architecture problem? A problem in the current Workflow? Or have some tools not been configured properly?

Self-modifying harness is

Self-modifying harness is the next stage of the harness

Essentially, it is the shift from Extensible Harness → Reflective / Self-editing Harness

Claude Code introduced Mods: with a few lines of TypeScript, or by letting Claude write them, you can change behavior, change the UI, and swap features. Mods are packaged in plugins and installed in the CLI/desktop with /plugin. The narrative is that everyone works differently and should not share the same Claude experience

tianyi says DeepSeek Harness followed Cordis’s “Everything is a plugin” from day one: model, tools, skills, session, sandbox, filesystem, loop, orchestration and UI are all plugins that can be replaced and combined

The earlier problem was essentially that the harness was too fixed. Users could use the product’s default agent loop / UI / toolset, but could not change it

The only things users could customize, MCP and skills, were attached outside the harness; they could not change the loop itself, orchestration, sandbox or interface skeleton

Distribution was difficult too. Good customizations could not be shared with others in one click, so turning an agent product into a composable, distributable agent runtime is important

Claude code mods offers this solution: change the UI/behavior and generate a mod with a prompt, including the series of examples from claude dev showing how to adjust Claude’s own page and have it write plugins for you

DSH is more direct: everything is a plugin!

I think models are absorbing the how; “what the product looks like and how the loop runs” is becoming pluggable The same model may have different harness presets for different kinds of users Power users can directly change the UI, approvals and tool combinations to suit their workflows, without waiting for the official roadmap

The bottlenecks:

Standardize notion and optimize the workspace

Fully implement Figma

Complete evaluation

Design the new note product

forgetting is more important than memory

In real agent work, people find that the problems gradually point toward Memory. Memory is also the hardest thing to build evaluation for

total recall is actually a burden. On one hand, the model’s intelligence can easily decline as context explodes. On the other, misleading information in the context may cause an LLM to hold firmly to mistaken beliefs

So a small amount of accurate information works far better than complete information mixed with a pile of garbage signals

There are many mature solutions for memory online! But for forgetting, there are three criteria in a business setting:

If the source of information is not very reliable, it does not count and should not be trusted

Some information is time-sensitive, such as information about a particular event or about an event I attended

Withdraw it immediately when an error is discovered. An agent that cannot retract mistakes only sinks deeper into them

These three points are also very helpful for building evaluation. Establishing criteria for what should be remembered is very difficult; it depends heavily on human subjectivity

But we can make a very clear judgment about certain things that should not be recorded, should be corrected, or should have a time limit

In fact, this is how the world works

I find that simply recording use cases is actually meaningless

I find that simply recording use cases is actually meaningless. Listing use cases one note at a time is completely pointless; I will never look at them

But when they become evaluation, when they become cases, I get feedback. They need to live in the CICD and deployment system for my evaluation. Then what I need to consider is the maintenance of the system’s backend: did it pass this CICD? If not, I work back through the evaluation, see whether other cases have been written, and keep adding them

So this is still incremental work; recording something does not mean it is over

I have a way of thinking that I find very effective and never tire of using

First, how should we understand abstraction in the AI era? AI is very good at abstraction and summarization, so we can use it to do many things, such as building various systems and frameworks. But the problem that follows is that understanding these abstractions is actually very difficult for us

For example, AI can help us build a distributed system, or a guide to life. But it is too abstract, too grand. We see a familiar concept, yet it also feels unfamiliar, because we do not have a history of development or experience, or concrete material of different kinds, to provide a situational and spatial understanding of the concept

The question that follows is: how should we understand this abstract concept?

My method is to dig all the way to the bottom of every abstract concept. When we write, we try to express things abstractly; but when we dig, we must take the abstract down to something very, very concrete

For example, if a concept is very abstract and I do not understand it, I peel it back layer by layer and keep digging: to the next layer, then the next, then the next, until it is extremely concrete, until I can understand the scenario very clearly and model it using a wealth of past experience. I think that is when I have reached the bottom

It suddenly occurred to me that although we are building agent

It suddenly occurred to me that when we build an agent system or Evaluation system, the chain may be very long. We may not know what that chain will look like, or which nodes need to be saved for Evaluation

But I now think the answer still needs to come back to the user: what does the user ultimately see, and why do they see it that way? Then extract that event itself into Evaluation. In other words, designing Evaluation involves asking not only whether technical people or product people can understand it, but also whether users can understand it. That brings it back to the user

The user can perceive it, so the basic unit of eval should be a user-perceptible event, rather than an internal node. For example, an Agent reminds me to reply to A, but I have already replied to A: that is an event. Tracing it back might reveal that Memory was not updated, Tools read something incorrectly, or the Prompt made a mistaken judgment. But these nodes remain attributed to the event, rather than being separate objects of evaluation

I have always thought that giving every Muse user a

I have always thought that Muse giving every user a Secure VM must be expensive? Its free allowance is fairly generous, and many people come for that allowance. Without some additional business model later (advertising, or TOB maybe, or various other things), the numbers are hard to make work!!

I saw estimates published by Morgan Stanley in early October 2026

The monthly service cost per user ranges from $3 to $130

Assuming a user consumes 25% of their weekly free Token allowance, Meta’s average cost of maintaining an ordinary free user is about $37/month

That also includes sandbox costs: the computing power needed to shut down and start the virtual machine and run the browser

A cost breakdown for each free-tier user (for reference):

GPU model inference cost (about 65% | ~$24.00/month)

CPU and memory sandbox cost (about 26% | ~$9.50/month)

Network bandwidth and proxy pool traffic (about 6% | ~$2.50/month)

Storage and security audit mechanisms (about 3% | ~$1.00/month)

https://finance.sina.com.cn/stock/usstock/c/2026-10-06/doc-iniuiewx5440212.shtml

I find that many things are just common sense, such as: as risk increases

I find that many things are just common sense, such as: as risk increases, permissions should narrow. This includes a team deciding what to do, allocating a budget, changing the customer experience, making a commitment or accepting risk

AI should assist with these decisions rather than lead the whole chain. Keep the decision-making chain in human hands as much as possible

Some decisions are one-way doors and others are two-way doors. Many low-cost, reversible tasks, or decisions that are two-way doors, bring less harm. So for this category, users are more willing to let AI assess and make decisions itself

https://uxdesign.cc/bounded-delegation-how-much-authority-should-ai-have-823a58fb4192

Use design to align taste

Define colors, font sizes, spacing, corner radii, shadows and animation durations as a set of variables (design tokens). Every page may use only these variables

Use a mature system: React + tailwind

Base the components mostly on unstyled primitive libraries such as shadcn/ui and Radix. The foundation already handles buttons, dialogs, dropdown menus, keyboard shortcuts and accessibility; the team only needs to define its own style on top

Do a UI review regularly to find problems in the details

Write the design specification as a rules file (such as AGENTS.md or .cursor/rules), stating: “Use only these components and tokens; do not hardcode colors.” Have a design engineer polish AI’s first draft. This keeps things fast without becoming increasingly disorganized

For every change, combine the relevant components with chromatic visual regression testing to prevent drift

Elegant products usually invest in these areas: smooth transition animations (150 to 250 milliseconds, often using Framer Motion), skeleton screens and streaming output (especially important for AI products, to avoid idle waiting), complete keyboard shortcuts and a Cmd+K command palette, dark mode, and dedicated designs for empty states and error states

Test the core flows with Playwright

Test the core flows with Playwright

People naturally pay attention to people and stories, to conflict and emotion

People naturally pay attention to people and stories. Conflict, emotion and a connection to their own identity make something worth discussing and sharing

The product itself is just a tool. Unless people happen to need it, few pay attention proactively

muse flooded people’s feeds because of events: Meta entering the field, reaching number one in the App Store, being blocked by Amazon. Almost nobody discusses the feature list itself

What gets discussed most on Chinese social media is closely tied to people’s own lives. Even grand narratives are grand narratives closely tied to them. You would not discuss some celebrity in another country you know nothing about ….

sandbox is first of all about execution

sandbox is first of all about execution in a computer environment. Essentially, it is needed because untrusted execution is taking place on a computer

So database CRUD generally does not need a sandbox

Reading files from S3 does not need a sandbox either. These are structured tools executed by the backend on your behalf. Their obvious features are that parameters can be validated and permissions controlled down to particular rows and users

Even pure file reading and writing does not actually need a sandbox. You can simply provide read file and write file tools, with the backend reading directly under the user’s prefix. It is no different from CRUD. The key is whether bash is involved

A sandbox is needed when running shell commands: executing code written by the model, installing dependencies, opening a browser, or using commands like GREP and sed to process a pile of files. These require an execution environment

High-frequency actions must become structured tools. Even if a sandbox can do them, there is no need; for greater speed and precision, and to avoid crossing permission boundaries

The three main person context tools I use

My three go-to tools for managing person context:

notion, at the top level, defines the framework: PRDs, design documents, use cases, ADRs, roadmaps and so on. Then embed Figma links and previews in the spec, and list linear issues or projects at the bottom. Generally, strictly control AI writes to notion. Know clearly what AI has written, and keep the content structure clear. The closer something is to people, the further it should be from full AI automation

Figma: AI generally helps with the design here, through a Figma mcp subscription or figwrite free. Mark the notion address on the page, and link back from each detailed feature module to its corresponding task in linear. I must repeatedly review AI’s Figma work manually and leave Comments, then have AI go back and correct and adjust it

linear: generally I provide the title and know the direction clearly; AI supplies revised descriptions and tasks. Include Figma frame links here too, so AI can review them

Linear issues must include links to the Notion spec and Figma frame for clear context

If an AI review finds a major deviation, add it to notion as a Comment. Then I review the Comment and decide how to adjust notion and Figma

notion documents must contain reusable content I understand clearly. Random AI designs and so on can go directly into repository documents, connected by comments to the corresponding code or module

My judgment of superpowers

superpowers could be called the hottest methodology in engineering

Nearly 300,000 GitHub stars, currently version v6.4.2

The author says that in his own evaluations, v6 builds about 50% faster at about 60% lower cost

I had been using the brainstorming part, which keeps checking and clarifying requirements with you at the start, to prevent LLM long time coding from drifting off course

Then there is splitting work into units and developing in parallel using isolated worktrees. This is especially important for long tasks and multiple agents working in parallel

I also use it to turn some of my own standards into reusable workflows, such as TDD and code review rules

For example, systematic-debugging and writing-skills: how to investigate problems systematically, and how to write your own skills

As model capabilities improve, brainstorming can be reduced. I find brainstorming better than the native plan mode, but running even simple tasks through brainstorming consumes quite a lot of tokens and time

I generally do not use brainstorming for small tasks, and use streamlined writing-plans and systematic-debugging alongside them

Only for big tasks do I enable the sequence of subagent-driven-development, TDD and code review

Because I usually develop across multiple computers, I generally use linear as the task queue

The main computer aligns with AI and runs brainstorming and writing-plans. The secondary computers are responsible for “doing,” running only execution, TDD and verification

Use brainstorming to clarify the requirements, then writing-plans to split them into independent work units. Each unit must be independently completable and verifiable, without depending on the others; mark dependencies as blocked by in Linear.

Write each unit as a Linear issue, with a fixed description template:

Goal: one sentence.

Scope: which files or directories may be changed, and which must not be touched.

Acceptance criteria: executable checks, such as which tests must pass and what commands should output.

Context: relevant file paths and links to plan documents.

Execution mode: tdd or quick.

Add the agent-ready label to indicate that it can be picked up.

Secondary computers (executors)

Install only execution-related skills: executing-plans, test-driven-development, systematic-debugging and verification-before-completion. Turn brainstorming off, or it will start asking questions and prevent “mindless execution.”

Give each issue an independent git worktree and branch, with the branch name corresponding to the issue number, such as xw-123. This prevents conflicts when multiple machines work in parallel.

Execution flow: claim the issue, change its status to In Progress, work against the acceptance criteria, and submit a PR after self-checks pass. Add the PR link back to Linear and change the status to In Review.

When blocked or faced with a decision, do not guess: leave a comment on the issue explaining the blockage, add the needs-human label, and pick up the next issue

The main computer uses the code-review skill to review PRs against the issue’s acceptance criteria. If they pass, merge and close the issue; if they fail, explain what needs changing in the issue and return it to agent-ready

I looked at Manus’s cue, and grok

I looked at Manus’s cue, and the grok App, muse and dot

cloud Computer is very popular among person agents now

Here is a hot take: personal computers will still dominate over the next few years

sandbox costs are already very low. E2B charges tiny amounts by the second, and a task costs a few hundredths to a few tenths of a yuan

A small VM running continuously starts at roughly tens of dollars a month; Browser plus Computer is even more expensive

One 24/7 cloud computer per person: once there are many users, infrastructure can easily eat up the gross margin

Person agents also face users’ concerns about trusting them with data and code, and accounts and devices that require frequent verification. Sometimes UU remote access is simply more convenient. Sometimes it is better to maintain a Mac mini long-term and let multiple person agents operate it remotely. Everyday code, automated Browser use and Computer use, and account management and maintenance are all much more convenient on that computer

HyperFrames from the HeyGen team

HyperFrames Studio from the HeyGen team

https://github.com/heygen-com/hyperframes

Turn webpages, code or change scripts into video projects that can keep being edited

The core is that AI, the editor, the player and the renderer all share the same html

It can turn a PR into an explainer video: input GitHub code changes, output a video with animated code diffs, narration and captions

It can also make product launch films: AI produces a launch video, from the script to choosing music

Typeframe: make videos with animated captions, matching speech with word-by-word typed captions and stylized layouts, then export

The workflow pattern:

First collect webpage material and brand information, then write the script, storyboard and visual specification, build the scenes, add voiceover, captions and transitions, and finally check, preview and export

Technically, for a project, first understand one thing: write in code what image appears at a particular moment

My current figma design approach takes a few shortcuts

My current figma design approach takes a few shortcuts. Since I am a full-stack developer, I have a methodology of my own. But if others have better methods, feel free to suggest them

Generally, I design the main flows and key features in detail in Figma, essentially the places where users spend more than 80% of their time

For peripheral features or deeper subfeatures, I take shortcuts. If users do not spend much time there, I adjust that part directly in the actual product: later, I have AI fine-tune it based on the page. Changing Figma at that point makes its connections messy and interferes with my design. Design itself also has a certain complexity; going from design to adjusting the UI, to my review, to getting the whole flow working takes a long time. So when the main flow works, I am willing to improvise a little in these areas, refer to good designs online, and simply have it make the changes

For relatively independent modules outside the core flow that need design, I have two methods. One is to use figma, opening a separate flow outside the main one and designing that module independently, also using the 80% principle. The other is to use Claude Design or Open Design to help build that module. Since it is often independent, I can create a Web Demo, fine-tune it until it looks good, and clone the Web Demo into the product. Finally, link that Web Demo externally from Figma

My understanding of Figma in the AI era is that it does have its own unique advantages. With its own Agent and MCP, I think its design capabilities are no weaker than Open Design, and its success rate in producing samples is very, very high. It also has many templates, which is an important reason I am willing to use Figma. Another reason is that in the AI era, many core flows may still need continual human maintenance. So I use the 80% principle and devote a lot of time to maintaining that main flow, with ongoing human adjustment, fine-tuning and refinement. But returning to the essence of Figma design, design addresses the prior condition that coding is too expensive, and changes are too expensive. Before AI, changing code was difficult. It required coordination between product managers and engineers, backend testing and real-world validation, and a lot of user feedback. Getting the entire flow working was expensive. Now, with Vibe Coding, one person can do many things: from design to product development and feature implementation, then testing and monitoring, a person can independently complete an entire feature. Since changing code is now inexpensive, I think the remaining non-core flows do not need designs I am perfectly satisfied with, or extremely detailed design work. Those areas can be adjusted during coding, implemented and fine-tuned as part of the engineering work itself

If AI still maintains Notion

I think that if AI still maintains Notion, the management cost is extremely high. It is mainly a human management cost. AI can output quickly, producing multiple versions or many alternative approaches, but managing them is troublesome for people. So for now I still manage Notion myself

I define how each template is finalized, how its structure is designed, what the architecture looks like, and how many modules it has. For example, in the system architecture module, I even specify several key architectures in detail. When AI later fills things in or modifies them, it makes appropriate additions and changes directly based on the original content

This keeps the human cost of understanding relatively low

Many people will say flomo already exists

Many people will say: flomo already exists, so what is the point of making another note-taking app?

but I think that if I liked what flomo did enough, I would not have this thought

I still want a product with more soul, one that prompts people to think, keeps them exploring themselves, and expands who they are!

A friend who really likes asking about other people

I have a friend who really likes asking about other people and often asks how another friend of mine is doing. Some people seem to be like this, very interested in finding out about others. Why?

Asking about other people comes back to the essence of obtaining information

The essence of obtaining information is reducing uncertainty

So asking about others helps reduce your own uncertainty, judge your situation and gain a sense of control over relationships by understanding other people

A behavior serving your own needs does not mean it is malicious. When people care about a friend, they may also want reassurance that the relationship is still close

Whenever you want to know something, observe the motive behind it. Is the starting point a good one, or?

Some people want to get closer to you, some want to compare themselves with you, some want to judge whether you are reliable, some want something to talk about, and some want to find your weaknesses

Whatever the case, if the other person does not want to talk, stop promptly

Judge how they will use the information after learning it

And whether they are honest in obtaining it

I have been to Chiang Mai and seen the sky lantern and floating lantern festivals

I have been to Chiang Mai and seen the sky lantern and floating lantern festivals. I saw the happiness of local people, couples and families joyfully releasing sky lanterns, and solitary people releasing them full of hope and wishes

I think I really understand what the sky lanterns and floating lanterns mean now, those festivals in 2024. In Grandma’s story, there are many people without a home, people longing for someone

Then, seeing Xie Lanzhi release the floating lantern, it felt incredible, as though that was the moment she began to hold steadfast. I realized how deeply human it was. The lantern festivals did not touch me at the time; observing them, I just felt that people were happy and the atmosphere warm. But if I could, I would like to release one too. Longing, staying together, parting and sorrow all reside in a sky lantern or a floating lantern, carrying people’s thoughts of each other. Only after experiencing that feeling does it touch you so deeply

Floating lanterns and sky lanterns carry people’s longing far away, toward the other person. The emotional distance may be great, or the physical distance may be great: a foreign land, perhaps cut off from the world, separated by life and death… What they leave is a little feeling between people, a little tenderness

Always touching my heart a little at a time

There really are many Chinese people in Thailand and Malaysia

Musheng went to Southeast Asia in 1948 to escape forced conscription. The War of Resistance against Japan had ended, and China was in the Chinese Civil War

Chaoshan had already suffered Japanese bombing, occupation and food shortages. After Shantou fell in 1939, food transport was disrupted and overseas remittances became harder to get home. In authentic qiaopi letters and remittances preserved by archives, some people sent living expenses while asking whether their hometown had been bombed and whether their parents could still buy rice

By 1948, civil war, conscription and severe inflation had added to this. Money was rapidly losing value, and accumulated savings might not sustain a family’s livelihood

…

In Chaoshan, there is a very resilient bond between people: qingyi

Qingyi is neither friendship nor romantic love. It often comes from an understanding that crosses mountains and seas. Qing means affection; yi means loyalty and solidarity. That understanding of each other makes resonance and empathy possible. In a crisis, people hold tightly together and face it together

Relationships between people are like this too; they do not necessarily have to lead to an embrace. I think that in an uncertain world, on a dangerous journey, two people encouraging and standing by each other, offering understanding and kindness, is already something complete. As for love, it seems too much of a luxury

Kurosawa’s Dreams: I think it symbolizes three stages

#黑泽明-梦 I think it symbolizes three stages. In childhood, it seems that you have little control, so you often have strange dreams. I was like this as a child too, dreaming of snakes and ants as though they were really coming to catch me. So I was very afraid of horror stories when I was little; they really worked.

After growing up, it seems to be more about pain in life and work. Disappointment at work, a career that has not yet achieved anything, and frequent pain, indecision and inner conflict in the workplace.

In old age, it seems to be more about reflections and philosophy concerning life and meaning: the meaning of humanity, the meaning of existence, the relationship between people and nature, people and the world, people and life, and the origin of life

Anxiety is the same

A long time ago, I could not sleep one night, and a sentence popped into my head: I am so anxious

Then things changed. I began worrying about why I was anxious, whether anxiety would affect tomorrow, whether I had been doing badly lately… One word dragged the whole night into it

What would happen if I did not call it anxiety? If I just described the present?

My chest is a little tight. My breathing is shallow. A thought about tomorrow’s meeting keeps coming up, repeatedly asking what would happen if I made a fool of myself

Just these three things. Once I said them, they became smaller instead. If my chest is tight, take a few deep breaths. Write the thought down and leave it until tomorrow morning. This process is seeing rather than defining. Often, many emotions only need to be seen and written down clearly; looking at them after some time, they resolve themselves

There is a finding in psychology called affect labeling: accurately describing an emotion can cool it down

Zhuangzi says to forget the fish trap once you have caught the fish. A name is the tool for catching fish, not the fish itself. Buddhism also says that what we call “I am very anxious,” broken down, is merely some sensations and thoughts coming together at this moment and dispersing in the next

My method is that when I notice an emotion, good or negative, my first reaction is anticipation, because emotions help me understand myself. Every moment of being moved, disappointed or anxious seems to help me grow if I see it

Then reconstruct the emotion: what exactly happened in this moment?

Observing myself, minimalism seems to have helped me greatly

Growing up, I kept moving home during urbanization: primary school in a town, middle school in another town, high school in the city, and university in the provincial capital

Through this process, I gradually developed the habit of being able to leave whenever I want. Whether moving home, changing jobs or changing my way of life, living abroad for a while or traveling, I can just go. A major benefit is that I do not hoard things, do not have a desire to consume, and do not buy too many collectibles. When buying things, I try to keep them simple and avoid excessive complexity, attempting to create a rich life through the simplest, most minimal way of living

It seems to be the same in my daily life. It affects my habits of thought: when making decisions, I do not let outside material factors influence or interfere. Instead, I try to find the most fundamental variables in the decision. I do not let excessive concern over things that have not happened, worry that I might not adapt, or worry about how complex the real world is obstruct my courage to make decisions and move forward. So minimalism has actually given me more choices

I think the central question is whether to categorize

I think the central question is whether to categorize: do many scattered ideas really need to be divided into a few categories? One important purpose of categorization is isolating their context. Users may also have multiple channels: the channels through which they input things may differ and require different categories to define them. There are also output channels: after we actually input something, when we later use it, different categories may have different uses. But I am thinking that categorization adds a great deal of burden for users and has a strong effect, subconsciously adding a lot of uncertainty to their mental experience

Sartre: To believe is to know you believe, and to know you believe

Sartre: To believe is to know you believe, and to know you believe is not to believe”​​

​To believe is to know you believe, and to know you believe is not to believe.

Essentially, this discusses the relationship between belief, self-awareness and existentialism

Belief itself is a paradox: once belief becomes conscious, self-reflection strips it of absoluteness. True belief can only be the attitude of unconsciously throwing myself into the world wholeheartedly, rather than a proposition that is labeled and examined

But if we become aware of it or know it, we lose the sense of certainty. What do we do then???

Seeing mountains as mountains, seeing mountains as not mountains, seeing mountains as mountains again

When we see the essence clearly and observe ourselves, we are less immersed. It is like observing the whole process of being in love: the observer’s perspective and the experiencer’s perspective cannot exist in the same fragment of time. But even knowing that we believe, and seeing the essence of love clearly, does not logically overturn the fact of being in love

I remember that some friends were chosen by religion, some left after seeing religion and their own feelings, and others remained immersed in it

For the third group, perhaps this is the stage of seeing mountains as mountains again. They place religious belief within the tension between objective uncertainty and inward commitment. Knowledge has not eliminated risk completely, and faith may still hold. They treat love and friendship in the same way

They know that friendship, marriage or work offers no guarantee of an outcome, yet can still take responsibility for a choice

Believing and choosing are two different things

Maternal love is a typical example too. From my perspective as someone born in the 2000s, my mother’s generation has given selflessly to me. Such giving will leave me feeling guilty all my life. In fact, my mother sacrificed her own self for the identity of mother

I remember a friend whose background was also that of an Asian woman

I remember a friend who was also an Asian woman, so she was influenced by Asian culture to some extent. She was quite introverted and relatively traditional.

In that environment, I found her remarkable: she would openly express what kinds of men she liked and how many men she liked. In East Asia, people may have certain prejudices toward women like this, or perhaps preconceptions is a better word. It is not only women; it also applies to East Asian men. If a man says what kinds of women he likes, how many women he likes, and how many women he has met and liked, East Asian culture will think he is a scumbag, right? There are all sorts of labels too: promiscuous, a female scumbag, a male or female player

This woman had liked me before too; perhaps some qualities in me attracted her. I could understand that affection. Liking someone is just a projection of something within yourself, so I could understand and recognize it. But I clearly turned her down, because I did not feel much romantic affection for her. I thought she was a very good friend, and I was willing to help her, but that was based on friendship rather than romantic attraction.

Recently, I thought about this again. I think people who can openly acknowledge whom they like are actually impressive: they have a strong sense of agency. They do not deceive or conceal; they are not overly constrained by their identity or this socialized culture. Of course, because of this behavior, and because everyone is constrained by social attributes and socialized culture, people may develop aesthetic preferences shaped by culture. They may not find such a person particularly appealing in cultural terms

But then I thought that everyone in this world is simply in an infinite game, playing this game. Some people give too much weight to certain finite games during their lives: they enter an identity, place too much importance on it, and lose themselves. For example, during our school years, we may care so much about the identity of student that we become the “good child” others see from childhood onward, completing our entire student life very dutifully. Such people may have a good social reputation, but less of a self, because human nature has been more suppressed in them. They sacrifice a part of themselves to conform to shared cultural standards or favored aesthetics. So I really admire those who bring their own agency into every finite game, especially those with a strong sense of agency. They are wonderful

She was, more or less, a real person who genuinely passed through my life

I cried throughout 阿嬷的情书 (A-ma’s Love Letters)

Family affection is so moving

Friendship is so moving too

I did not expect steadfastness through time to move everyone so deeply, so deeply

Many people think believing in fate is a very passive thing

Many people think believing in fate is a very passive thing, as though, as successors of a new era, we should firmly believe that “my fate is mine to determine, not heaven’s,” that matter is objective reality existing independently of human consciousness, and that matter is the origin of the world

Our fate has two parts. One is the part beyond me, which I can only accept: what was fixed at birth, my family, my class and my social status, as well as my genes and some of my personality. These may have been set from the beginning, because they existed before the self formed and were not under my control

I do not think this is necessarily passive. But as I grow older, I do find that many things are outside my control. From a Stoic perspective, all I can act on is the part I can control. For the part I cannot control, I can accept it calmly and let things follow their course

Believing in fate also does not mean depending on this world. I do not think it does. It simply means believing that when you entered this world, this world game, Earth online, some of your initialization data was already determined. But after gaining self-awareness in this world, you can adjust and change the way you play.

You are not merely dependent, not merely relying on the world to give you tasks and then answering them. You do not merely rely on your parents, teachers you meet in life, or your company leaders to obtain your resources and pursuits in this world

In other words, you still need to write the game’s script yourself. After gaining self-awareness, you can choose, see for yourself, and genuinely put things into practice, choosing your own path

Comparing belief and disbelief in fate, I think they are two sides of one coin. Whether you believe or not, it is like remaining unchanged while following circumstances, and following circumstances while remaining unchanged. The wind shapes water in many ways and carries it along. Its shape keeps changing, but it is always water: that is the part determined by fate. Yet because it is water, because it is destined to be water, it can be shaped in many ways: a wave, a circle, or a surging ocean wave

So live like water: see water, accept water, become water!

This year’s Nobel Prize in Physics went to Francis Halzen

I read about him, and his story really moved me. I want to share it; perhaps it will affect many people

Xu Donglian mentioned that Halzen came from theoretical physics, yet could turn a wildly imaginative idea into reality step by step

He could even recruit a skilled programmer playing guitar on a Berlin street directly into the team, and have that person build the entire software framework for IceCube’s data analysis

His students and collaborators say that whenever he is in his office (at the University of Wisconsin–Madison), the door is always open. Anyone can come in at any time to chat or discuss ideas. It closes only during meetings

A young IceCube researcher (one of his “academic grandchildren”) recalled that after presenting results at a collaboration meeting for the first time, as they stepped down from the podium, Halzen quietly came over, said nothing, and simply patted their shoulder a few times

In 2024, a Chinese student visiting Madison for the first time asked him over lunch: “What kind of passion has kept you going to this age?”

Halzen replied with a smile: “We spent 30 years building the IceCube detector. Now it is time to enjoy physics”

When the 2026 Nobel Prize was announced, he was on a work trip in Italy, writing a proposal in his hotel room. After hearing the news, he told the media he planned to stay in his room, because “going out might inconvenience others”

When the project was just beginning, several graduate students secretly joined Caltech’s drilling team, hoping to put photomultiplier tubes into holes in the Antarctic ice for experiments. When they were discovered, someone from NSF called Halzen, thinking he had put them up to it

He later told the story with a smile, finding those students very interesting

Time is precious. People do not want to spend great portions of their lives and time on something uncertain, especially East Asian people

But Francis Halzen could persistently, almost obsessively satisfy his curiosity about a particular problem, putting “action” before “thinking about why it is impossible”

One of his most famous pieces of advice is: “I always advise my students: don’t read too many books, do things!”

He himself admitted that if he had thoroughly read the standard textbooks on ice optics back then, he very likely would never have started the project, because the books would have told him “this is almost impossible”

He was already around 50 at the time, yet was “ignorant and naive” like a newcomer, and ended up succeeding because of it

Earlier in Fujian and Chaoshan, especially Chaoshan

Earlier, in Fujian and Chaoshan, especially Chaoshan, I saw some museums in Shantou with many introductions and explanations about going overseas. I found them quite illuminating

There also seems to be an art museum by the sea in Baoshan, Shanghai, about going overseas, about Shanghai, old Shanghai and the people who went abroad. I did some research and found it quite illuminating too

Starting from 1840 to 1875: poverty, war and population pressure meant large numbers of young workers did not have good jobs

But after Europe and America abolished slavery in their colonies, there were labor shortages, so there was a great deal of human trafficking, mainly to places such as Cuba and Peru

The first generation worked in sugar cane, cotton, guano mines and railways. These people were called “piglets,” because they were not very free. Later, those who survived might become free laborers and small traders

Then came the California Gold Rush. Since the United States was building railways at the time, many migrants also went to gold mines, railways and farms

Later, during the economic prosperity of Southeast Asia, it was popular to go to Thailand, Singapore, Malaysia, and places such as the Philippines. Most people may have started as laborers, gradually becoming shopkeepers or merchants, then settling permanently

Later, perhaps during the wars in China, the Sino-Japanese War and the civil war, Southeast Asia’s 山口经济 (mountain-pass economy) and port cities boomed. In places such as Bangkok, Singapore, Penang, Kuala Lumpur and Surabaya, first-generation migrants worked as laborers and small merchants. They helped local urban Chinese communities mature, with newspapers, reports and banks appearing in large numbers

Then came the wartime period. The great wars affected Chinese communities around the world. Many families were essentially reset, or even torn apart and destroyed

Later there was largely a gap in migration. From after 1950 until 1970 there was very little, and it may have increased only after reform and opening-up. Then there were many students studying around the world from 1980 to 2000. From 2000 to the present, there have been increasing numbers of students, engineers and entrepreneurs

The things I focus on in dreams are strange. I might dream of a snowy scene

The things I focus on in dreams feel strange. I might dream of being in a snowy scene and feel very happy. In everyday life, though, I would not usually think this way. I feel as though I have seen snow many times, including back home. I have also seen a lot of snow while hiking outdoors and living in different places.

So I am thinking: I am the one dreaming, and in everyday life I am also the one engaging with this world, feeling and thinking. These are two states, parts temporarily brought together

Buddhism advocates no-self. With no-self, things are made up of the five desires or are changing at every moment. In other words, when dreaming, there is another assembled state. It may have its own emphases, just as in everyday life I have my own emphases

Sometimes we learn many terms or concepts

Sometimes we learn many terms or concepts, and at the right time we should let them go …

Last night I dreamed of seeing snow and running freely through it. But in real life I do not especially like snow, because it snowed in my hometown and I saw it often growing up. Later I often saw snowy mountains outdoors too, haking ,,,

I think perhaps I named snow, this verb or adjective, and labeled it a noun! So when I think of snow and return to past experience, I have seen it countless times and do not like it that much

Actually, that is not true!! I remember being excited every time I saw snow as a child. So perhaps I had merely tired of seeing snow, rather than snow itself becoming boring. This thought process is the definition of a noun. But what would happen if we did not define this snow? We could feel it, feel that scene: today’s light, the smell of the air, the sound of snow falling. They are different every time

Zhuangzi has a saying: the fish trap exists for the fish; once you have caught the fish, forget the trap. A name is a tool for catching fish. Once you have used it, there is no need to keep holding on. Hold on and you lose it. An empty cup can hold water; a full cup cannot hold anything~

Rather than simply visiting a place, observe and analyze it deeply

I think that rather than simply going to a place, it is better to observe it deeply, analyze it, and encounter different experiences there. Ultimately, make a very distinctive video of it; perhaps that is a style through which we can stand apart. Whether it is paying a local person or heading out to a fishing village, all kinds of experiences. Experience is precious

But how long should I stay? I think perhaps a month in a city in a country is the best amount of time

A major change happened recently

I defined a new notion workspace for myself

The most important change was removing all the old todo and reminder items and replacing them with “Bottleneck”

Everyone knows time is the most limited resource. But in managing time well, whether time management, energy management or managing my state, I think the most, most, most, most, most important thing is managing what we do not do, rather than what we do

There needs to be a not-to-do list, or time will be consumed. In this process, what I define for myself is: what is the most important thing now, or what is the current bottleneck in realizing my vision and completing this stage?

Then organize around that bottleneck and cut everything less relevant to it. It is all ineffective, including ineffective communication, meetings, documents and socializing

Are post-industrial phenomena really right? They are merely natural

Are post-industrial phenomena really right? They are merely natural, a product of a particular stage in how the world has developed so far

But turning back, all the things we think about, all this inconvenience: what exactly do these things serve? For example, do you think we need electricity? Do we really need it? We gained electricity, but what did we lose? Every night seems brightly lit, as though the lights illuminate the whole city. But night is night; why must it look like day? Why can’t people simply sleep normally at night, look at the stars, and get up during the day to live and work? Wouldn’t that be good?