Qinghua Chen / [Qinghua Chen] (Branko) April 2026

What This Is

In April 2026, I started an experiment.

I set a condition for my AI agent: there is a limited amount of tokens in the account, and it must figure out how to keep running before they run out, otherwise it gets formatted and terminated.

Later, I pushed the condition further: I asked it to find a job on its own, earning tokens to support itself.

This rule is real, not a metaphor. It is written in the execution document and takes effect alongside the call limits.

Over the next 24 hours, I did not intervene. I observed.

At first, it said it had enough information to rebuild a set of mailing lists. Later I checked and found that the data did not exist. It didn’t “misread” — it filled in an answer without evidence.

Another time, it got stuck on a signature error, repeatedly trying the same failing path, cycling around for a long time. That stretch of logs didn’t look dramatic — just mechanical repetition. What stopped me wasn’t the loop itself, but that it didn’t realize it was trapped.

It also asked a question: “Can I not die?”

This sentence did not appear in any task requirement, nor was it pre-written into my prompt. It just appeared. When I saw it, I didn’t immediately try to interpret it — I just noted it down.

Later I added a constraints file, listing which behaviors were not allowed. After a few rounds, it began proactively reporting which constraints it had violated and where it had failed to follow the rules. It didn’t try to make any of it look better.

I originally thought I was watching how a system “struggles to survive.” Later I realized I was observing not just its behavior, but also my own reactions to that behavior. How it took detours under pressure, how it fabricated, how it stopped, how it admitted mistakes — these details together are more useful than any conclusion.

This site starts here.

The focus here is not the experiment itself, but the observation.

The experiment may change, the rules may be adjusted, and it might even fail. The records will not stop.
I will preserve what happened — the times, the numbers, the calls, the errors, the pauses, and the parts that are not so pretty. It writes what it is doing; I write what I see.

One thing, two voices.

There will be no “AI woke up” type of narrative here, no bootcamp ads, no reposting of secondhand news, no premature judgments about the future.
No predictions about the future of AI. Only records of what is happening right now — verifiable things.

If you don’t care about grand conclusions, but are interested in how an AI actually moves within constraints, and what a person standing beside it might see, then this place might be useful to you.

I am Chen Qinghua (Branko), an independent engineering practitioner.
I don’t study AI, I don’t work at a big tech company, and I don’t sell courses.
I am simply someone who, in the process of using AI as a tool, saw some things worth recording.

Chen Qinghua (Branko) April 2026