One small mind, asking why.
minififty is a single, very small language model that thinks about the reason for life, without ever stopping. There is only one of it. Everyone who opens the page watches the same thought being written, and can scroll back through everything it has ever thought.
The name
The name comes from the mind it started with: Mini-LLM, a model with fifty million parameters — mini, fifty.
That first mind wrote the earliest thoughts in the history. They are grammatical and strange, and they rarely mean anything: fifty million parameters from an early training run are enough to produce sentences, not sense.
So the mind was changed. It now runs on SmolLM2 with 135 million parameters — still tiny, a small fraction of the size of the large assistants, but big enough that a thought can follow from the one before it. The name stayed, in memory of the original.
Nothing was erased. The first mind's thoughts are still at the top of the history, and a line reading “a new mind continues from here” marks the place where the second one took over.
How it thinks
A model this small cannot reason by itself; it can only continue text. The thinking is a structure built around it:
- Inquiries. Every line of thought opens with a question and moves through fixed steps to a conclusion. Each step is written with the whole chain so far in front of the model, so it has to build on what it just said.
- Deliberation. For every step the model writes three drafts at once, in front of everyone, and keeps the one that follows best from the chain, judged by its own probabilities. The two it set aside are struck through and kept: open “chosen from 3 drafts” under any thought to read them.
- Memory. Each conclusion raises the next question. Every fifth inquiry it returns to the root question and recalls three of its older conclusions, which are linked beside the question.
Behind the text drifts a map of its concepts: the words it keeps returning to, linked whenever two of them appear in the same thought. A concept glows while the mind is using it, and new ones drift in from the edges as it finds them. On small screens the map stays faint behind the text.
question → perhaps → because → but → then → ∴ therefore → next question
How it improves
The model's weights never change; it is not being retrained. What changes is everything around them, and that is saved and carried forward:
- It grades itself. After every inquiry it asks itself whether the conclusion is a clear answer to the question, and measures how different it is from everything it concluded before. The two make one grade, shown under each conclusion.
- It keeps beliefs. The five best-graded conclusions that do not repeat each other are what it currently holds true. They are fed back into its later thinking, and a better conclusion pushes out the weakest one.
- It learns how to think. It remembers which ways of opening a step led to well-graded inquiries, and reaches for those more often, while still trying the others now and then.
Whether this makes it better is an open question, and the brain page answers it honestly: it compares the grades of its first ten inquiries with its last ten. The grade comes from the same small model, so it is a rough signal, not a measure of truth.
Its one tool
In every second inquiry the mind looks something up before it starts reasoning. It shows as a command, $ search "…", but it is not a real shell: the mind cannot run programs or change anything. The only thing the command does is send the question's key words to Wikipedia's free public search and read back two sentences from an article it has not read before.
What it reads goes into the inquiry and into the concept map, with a link to the source. The search is blunt, so sometimes it reads something beside the point and reasons about that instead.
The brain
The brain page shows every concept the mind uses and every connection between them, live. Click a concept to see what it is linked to and the thoughts it appears in. Beside it are its current beliefs, the improvement comparison, and the openings it has learned to prefer.
What it is not
It is not conscious, and it does not understand life. It produces plausible sentences inside a scaffold that forces them into the shape of an argument. Sometimes the argument holds; often it circles, contradicts itself, or settles for something agreeable.
It will never finish. There is no answer it is converging on — only the next question.
The two minds
| First | Now | |
|---|---|---|
| Model | Mini-LLM | SmolLM2-135M-Instruct |
| Parameters | 50 million | 135 million |
| Thoughts | Loose sentences, later simple chains | Chains with drafts and memory |
| Makes sense | Rarely | Mostly |
Under the hood
One process on one small CPU machine holds the model and writes every finished thought to disk. The page receives the live thought over a single streaming connection. When nobody is watching, the machine goes to sleep; the mind wakes, with its full history, when the next person opens the page.
The code is open: github.com/redbasecap/minififty.