This is my third week with the new MacBook.

First, a few numbers: Codex's token counter has reached 220 million; I used half of my newly purchased Kimi subscription allowance in two days; almost all 48 GB of memory is occupied, with another 1.3 GB of swap in use.

Paintings? Very few.

At the end of Weekly 02, I said I wanted to do something in week three that the old computer could not: run a complete creative project with Codex and move my portfolio website a little further along.

In the end, the website did not progress, and painting was put aside too. The new computer was not idle; this week's "creative work" simply looked different from what I had imagined.

With a typhoon and high temperatures in Shanghai this week, I stayed home most of the time. Watching the World Cup also briefly turned my days and nights upside down. Each day, after waking, I first collected material for my WeChat publication and completed my daily writing. The remaining blocks of time went into listening to people talk about AI, watching AI coverage, and using AI.

I was not studying a long list of prompts or learning complicated techniques.

I was mostly trying to understand how far AI had developed, where it might go next, and how a freelance painter like me should build an AI workflow of my own.

If you are also wondering whether to subscribe to two AIs at once, or whether spending money to change your understanding is worthwhile, this may offer a point of reference.

After watching the World Artificial Intelligence Conference, Kimi appeared on my desktop

The 2026 World Artificial Intelligence Conference took place in Shanghai from July 17 to 20.

I did not attend in person, but over those days I watched livestreams from various sources and listened to the rolling series of interviews from Dedao, Luo Zhenyu, and Kuaidao Qingyi. People from different industries sat down to discuss AI and the changes they were seeing. As I listened, I began to feel quite specifically that the future they described might no longer be very far away.

The day before the conference began, Kimi released K3.

A guest on a livestream spoke very highly of K3, placing it among the strongest models in China at the time. That caught my interest. I bought a subscription that day and installed the Kimi client on my MacBook.

The number of AIs on my desktop that could do work went from one to two.

Codex was one place to hand over tasks, and Kimi was another.

My reasoning was practical: if I used up Codex's allowance, I could move a task to Kimi and continue. I would also give the same project to each of them separately.

I was not especially interested in arguing about which was stronger.

Like people, models do not look at a problem in exactly the same way. Codex gives me one approach, Kimi another. I choose the parts from each that make sense and work well, then combine them into my own version.

The final decision remains mine.

After a few days, I do find Kimi's interface comfortable. The way information unfolds and tasks are presented feels smooth. Codex is already familiar: for research, editing, or setting up a project, I naturally open it when something needs doing.

The two tools feel less like opponents in a ring and more like two different work areas on my desk.

After 220 million tokens, the tool starts to "disappear"

Two days after buying Kimi, I had already used half the allowance.

It reminded me of starting with Codex, when I often used up the allowance within a usage window and had to wait for it to recover. Now that the token counter has reached 220 million, an occasional allowance reset no longer excites me the way it did at first.

Perhaps every new tool goes through this process.

Just after installation, its presence is very strong. You want to throw everything into it and see what it can do and how far it can go. The allowance disappears particularly quickly.

I rather like that first period of getting carried away.

Using a tool a lot is how you quickly encounter its limits. Which tasks it handles well, which answers look complete but cannot be used as they stand, and which tasks you need to think through yourself first: these become clear only after running a few projects.

Once you truly know how to use it, the tool gradually starts to "disappear."

It becomes part of the everyday workflow and no longer calls for a separate expression of wonder each day. When I write now, I naturally open Dedao Brain and Codex. I do not first ask myself, "Should I use AI today?" There is a project, so I use it. Then I move on to the next thing.

You become less conscious of the tool, but the efficiency it brings has not disappeared.

The biggest surprise: AI listed 100 Shanghai sketching locations for me

My biggest surprise this week came from researching places to sketch and gather inspiration in Shanghai.

I wanted a list of places I could gradually visit, and asked Kimi to research it. It eventually listed 97 locations, along with addresses, costs, and opening hours.

I had not even asked it to make the number up to 100.

Previously, I would have had to open page after page myself, look up places, read descriptions, record addresses, and decide whether each was suitable. Before setting out, much of my energy would already have gone into preparation.

Now AI organizes the scattered information first. I only need to choose and go out.

It prepares the "where to go." What remains is for me to actually leave the house.

That is valuable to a freelance creator. We do not necessarily lack ideas. What we lack is someone to handle the long string of small tasks that stand between an idea and action.

Research, gathering information, sorting it, and making a first pass at analysis: none is especially difficult on its own, but all take time. If AI takes them on, I can save my energy for sketching, judgment, and creative work.

As for whether a particular venue's opening hours are accurate, I am fairly tolerant. When I am actually ready to go, I can check them separately. Venues adjust their arrangements; hours change. Things were never fixed forever. One outdated item would not make me dismiss the value of the whole piece of research.

What I ask of AI is not that "every word be absolutely correct," but whether what it gives me helps me keep acting.

What 48 GB of memory buys is the ability not to notice it

After Weekly 01 came out, a reader said it was forward-looking to buy 48 GB of memory before I had a clear need for high performance.

At the time, I did not know whether I would fill it. I only expected that I would probably need it in the future.

Now the answer is: I already have.

Activity Monitor shows almost all 48 GB in use, plus about 1.3 GB of swap. The browser has a pile of pages open, Codex and Kimi are working at the same time, and Dedao Brain, WeChat, and other everyday applications are sitting in the background.

They have not slowed the computer enough to bother me, so I do not even constantly notice that the memory is full.

What 48 GB gives me is not the feeling of "what a powerful computer," but the absence of a feeling: I do not have to decide what to close every time I open an application.

A month ago, when I bought the computer, I certainly did not predict that Kimi K3 would launch this week, or know how many AI tools I would run at once. But things do change quickly in the AI era. Models change, tools change, and the way I work changes too.

Sometimes being forward-looking does not mean accurately guessing what you will use in the future. It means leaving the future a little more room.

You do not need to "finish learning" AI before you start

Still, too much AI news can become irritating.

Since getting the new MacBook, I have been reading a greater quantity and variety of information. The algorithms have responded by sending me even more AI content: new models, new tools, new courses, and endless advice on "how ordinary people can seize the AI opportunity" or "how to avoid detours."

I would not call it anxiety. It is more a little restlessness and irritation.

Even as an AI beginner, I have found through actual use that it is not as complicated as it is often made to sound. I do not need to complete an entire body of knowledge to qualify to begin. Install the software, tell it in ordinary language what I want to do, and work can already start.

I understand that some people like to know steps one, two, and three before they begin. Tutorials certainly have value.

But my own approach has always been to get my hands on it first.

It is like painting. You do not have to understand all of perspective, color, and composition before making your first painting. Begin. When the perspective goes wrong, work on perspective; when the colors get muddy, investigate color.

AI is the same. Install it, open it, and use it boldly. If a task does not work, phrase it differently. If the result disappoints, keep adjusting. If you use up the allowance, treat that as the cost of learning at this stage.

Spending the money hurts, but it changes what I understand

Using half of Kimi's allowance in two days does feel expensive. The agent allowance refreshes monthly according to the subscription cycle. If I use it early, I either have to wait longingly for the next cycle or buy a top-up. I need to be a little more careful with it.

But when it produces something I am pleased with, I feel it was worthwhile.

At this stage, the money and tokens I spend buy more than a single result. They also change my understanding. Other people can warn me about many detours, but they cannot walk them for me.

Only after actually stumbling once do I know where to turn next time.

After two months of intensive AI use, two things need protecting

After a month or two of intensive AI use, I increasingly agree with this view: our own judgment and aesthetic sense will matter more in the AI era.

What a model produces often represents a kind of average level. With enough use, you can roughly predict how it will answer. You know how far it can go, and where you need to take over.

AI can instantly give me many topics, plans, and possible paths. But if I am already uncertain, those options do not necessarily make things clearer. They may leave me more confused.

Only when the direction and purpose are relatively clear can it answer the next question well: how do we make this happen?

AI is good at answering "how." I still need to decide "what, exactly, do I want to do?"

In the past, I was better at the collecting part of research. When I saw a relevant artist, example, or set of data, I saved it first. At the level of analysis, I still needed to build my ability, and I genuinely did not have that much time or mental energy.

Now AI can offer a first round of analysis. It may not explain every factor behind a popular piece of content, but it at least gives me a general direction. I then bring in my own experience to decide what is worth learning from and what does not suit me.

Let AI do the legwork, organizing, and analysis, while keeping the judgment in my own hands. By week three, this division has become clearer and clearer.

Some detours are ones you need to walk yourself

Last week, I said I would run a complete creative project with Codex.

As I originally imagined it, this week should have produced a website or a polished example of a project taken from research to finished writing.

The reality is that the website is not ready to be public, and the weather and my sleeping patterns have also interrupted a good deal of painting.

But perhaps I completed another kind of project: I began using two AIs together, finished the research into Shanghai sketching locations, and gained a clearer sense of where I want AI to sit in my workflow.

That may simply be how doing things works. When it does not run, adjust it. After adjusting, try again.

AI repeatedly tries, checks, and revises while carrying out a task. People do too. Failing to make it work the first time does not mean you cannot do it. Failing to deliver the originally planned result this week does not mean you have not moved forward.

Some detours are there to form judgment, rather than to be avoided.

Week three: not much painting, but the way is prepared

Recently, I have also been preparing an online service that offers companionship while people paint.

AI can tell you how to paint, but it cannot share the experience of a moment of flow in painting with you. It can help me research, organize needs, develop a structure, and explore plans. It cannot decide for me what kind of people I want to accompany, or how I want to help them keep painting.

Tools can save me preparation time. The actual action, judgment, and companionship between people still have to come from me.

I did not paint much in the third week, but I have already had AI organize 100 places where I could go and paint. Next, it is my turn to head outside.