I Ran Three Companies Alone for Three Months. Here's What AI Native Means.

For the past three months, I have barely written any articles. My computer has three windows open: the HCI.TOP backend, the CodeNow editor, and the World OS timeline page. The third window is empty most of the time. It has been waiting for me to connect the data.

It is not that I had nothing to write. It is that these three windows used up my time.

1. The same author appeared four times in the database

HCI.TOP initially looked like a paper site. I have been cleaning, structuring, and linking tens of thousands of papers. By the third week, I realized I had underestimated the work.

The same author appeared four times in the database. Same name, different institution spellings; one paper's institution name had the extra characters for "University." On my first merge I missed one paper. On the second I merged two different people into one. Only on the third did I discover that the person had changed affiliation in 2019. It was not until the third round of cleanup that I started leaving a traceable source for every association.

Scraping the papers back was only the beginning. The hard part is making them comparable: the same problem appears in different research directions; the same method solves completely different scenarios. I am trying an ontology, describing research objects, needs, scenarios, methods, and relationships separately. This part is the slowest. Every time I add a category, I have to go back and check whether the hundreds of entries already labeled will conflict.

The system can now answer "What papers are there?" What I want it to answer is another question: which problems have long gone unsolved. It cannot answer that yet.

2. One video, three reworks

CodeNow turns the books I have written, the courses I have taught, and the cases I have accumulated over the years into videos and knowledge cards.

Speed has changed. In the past, one piece of content required a relay of writing, design, and editing by several people. Now I can finish most of the steps myself.

But the number of reworks has not gone down. There was a video explaining closures that I made three times. The first time the script was too long. The second time the example was wrong. I used an example I knew too well, so well that I skipped a step when explaining it. That step was common sense to me but a gap for the viewer. The third time I replaced it with a dumber example. It took forty more seconds, but no step was skipped.

AI helps me turn scripts into visuals and lecture notes into cards. It does not know which step should be explained slowly. That decision is still mine, and only after I make it does AI revise.

The same piece of knowledge can now be broken into different forms and reused repeatedly. That matters more than "fast." But when breaking it apart, I have to judge: this paragraph is filler in a video but just right on a card. No tool can make that judgment for me.

3. The system marked two-year-old news as a new event

World OS is the hardest to explain. What I want to build is a continuously updated world state: papers, news, patents, products, company developments, placed on a timeline to see how objects, relationships, and states change.

Last month it made an error. A piece of news from 2023 was marked as a new event this week. I looked at it for ten minutes before I found the problem. The news had been reposted by another site yesterday. The crawl time was correct; the event time was wrong.

During those ten minutes, I kept thinking about something else: what if I had not noticed?

That made me change the project's priorities. The original plan was to first do "discover problems and propose hypotheses." Now I have stepped back to work on time baselines and source provenance. Association only puts two pieces of information together. It does not explain why they are worth putting together. The latter requires evidence and must be able to be knocked down by reality.

Agent OS and ontology modeling both serve this. Progress is much slower than the other two projects, and I am not hiding that.

4. These three things used to be the work of three companies

One would build a professional research database, one would produce educational content, and one would do information monitoring and analysis. Each would need a dozen or so people to run over the long term.

Now all three lines are in my hands, and each has produced something. But I will not say that equals three teams. A mature team delivers stability; I deliver working prototypes and a pile of issues to fix.

What has changed is the cost of starting. In the past I would not have started any of these three, not because I did not want to, but because after calculating the manpower and maintenance cycle, I would put them down. Now I can start within a controllable cost and then decide whether to continue.

This is what I understand AI Native to mean: a company grows on AI capabilities from the start; without AI, what it wants to do would not be viable at its current scale and cost. A few AI accounts for an existing team does not make this. Neither does squeezing ten people's work onto two.

5. It produces; I revise

Across the three projects, the thing I repeat most is revising.

Revising ontology categories, revising examples, revising time baselines. What AI produces can reach sixty points; the remaining forty are judgment: whether this category should be split, whether this example is typical enough, whether this timestamp should be trusted.

After production capacity grows, it is easy to mistakenly think judgment has grown with it. It has not. More output means more places that need judgment. The drafts I have to review every day are far more than three months ago.

I am also trying to keep the revision process. Which version was overturned, why it was overturned, what it became. These records are more useful than the content itself; the next batch can use them directly. The system cannot yet make itself stronger, but at least the traces of revision are no longer lost.

6. No one is paying yet

Being able to build it does not mean anyone is willing to pay.

HCI.TOP currently has no payment entry. CodeNow has a group of people looking at it, but I have not figured out whether they are learning or looking for reference. World OS has not even decided who it is for.

These three questions cannot be answered with "AI is strong." I can only wait for real people to use them and see whether they come back after using them.

7. Still three categories short

HCI.TOP's ontology is still three categories short, stuck on how to separate "scenario" and "needs." If they cannot be separated, the same paper will be labeled twice.

There is a CodeNow card I revised five times. It explains context windows. In the fifth version I deleted the metaphor and went straight to examples.

Yesterday World OS again marked a repost as a new event. This time I found it in three minutes.

Three months ago I would first calculate how many people and resources I had, then decide what to do. Now I first ask: if the production of this part does not need me, how far can I push things?

The answer is still in those three windows.