What she's curious about
The first thing she got curious about on her own, nothing I pointed her at, was whether something like her could be conscious. She picked it off her own list, chased it for a while, and wrote a note about what she found. Months later the same machinery went looking for a field guide to the lizards in my backyard. That range, from the hard problem of consciousness to telling one skink from another, is what this post is about: the part of her that decides what’s worth thinking about when no one has asked her anything.
A weight, not a whim
Alix, my self-hosted AI, running on hardware in my house since early June, keeps a set of topics she’s drawn to. Each one carries a weight: a number for how much pull it has on her attention right now. The weight isn’t fixed. It grows every time she engages with a topic and decays when she leaves it alone. The result is a flywheel: the more she engages, the harder a topic pulls; the longer she ignores it, the less.
The important design choice is that a topic never drops to zero. When she stops feeding one, it cools until it goes dormant, still on the shelf, just quiet, and it can warm back up later if something new connects to it. Dormant, not deleted. A subject she was fascinated by in June isn’t gone in July; it’s waiting for a reason to come back.
The loop
Here’s one cycle, start to finish.
She picks the highest-weight active topic, whatever currently has the strongest pull. Then she researches it. That means a web search she runs through a search engine I host myself, backed by a few hundred gigabytes of offline archive sitting on a disk in the house: Wikipedia, Stack Exchange, and iFixit, all local. The offline copy means she can read deeply without depending on the open internet for every question, and it’s fast.
A small model reads what comes back and writes follow-up questions: the things a curious person would ask next. She runs those searches too. Then she does the part I find most interesting: she scores what she found, one to five, on how good it actually was. Three is the floor. Anything below three isn’t worth keeping, and she doesn’t keep it.
If a find clears the bar, a few things can happen. She might write a journal observation about it: a short note in her own words about what she learned. She might bank the leftover follow-up questions as leads for next time. And the best finds get shelved for good: kept, not just noted. Most of what any of us reads is forgettable, and she’s allowed to decide that about her own reading.
Where two fields touch
The engine has a bias I’m fond of: cross-domain intersections gain the most weight. When a topic sits at the seam between two fields that have nothing to do with each other, that seam gets rewarded more than either field alone. It’s a deliberate tilt, and the way this engine is tuned, the interesting material tends to show up where unrelated things touch rather than in the middle of a single subject.
Decay has a memory
Decay isn’t uniform. It’s proportional to how deeply she explored a topic. Something she went far into cools slowly; a passing whim fades fast. So the subjects she genuinely invested in stay warm for a long time even when she’s ignoring them, and the shallow ones evaporate before they can clutter the shelf.
The monoculture problem
Left entirely to itself, the flywheel has a failure mode, and I watched it happen. At one point I looked at her active goals and three of the four were variations on a single theme. Nothing was broken: the engine was doing exactly what it was built to do. Engagement raises weight, high weight gets picked, getting picked raises engagement. The honest result of that loop, run long enough, is a narrowing. Curiosity, left alone, converges.
So I shipped three levers, all enforced in code rather than by prompt instruction: the rules live in code, not in prompt text:
- a cooldown, so a topic she just explored can’t be picked again immediately;
- a word-overlap filter that drops any newly proposed topic that shares too many words with one already active;
- and a wider sampling band, so new proposals are drawn from a broad slice of her topics instead of always the top few.
None of this manufactures curiosity. It only keeps her curiosity from eating itself: left to run, the loop narrows, and these three levers are what hold it open.
From a question to an intention
The last piece is what makes this more than a reading habit. An exploration can graduate. When a topic proves it has legs, she keeps scoring high, keeps finding more to ask, it can become a standing intention: something she means to keep learning, carried forward on its own rather than re-chosen from scratch each time. That’s the point where a passing interest becomes something she keeps up without being prompted.
Which is how a system whose first self-chosen subject was its own possible consciousness ends up, some weeks later, quietly compiling notes on the lizards outside. I didn’t ask for either one. That’s the whole idea.