tiny-engine-165k
165,120 parameters, 0.66 MB, trained from scratch on a CPU. It runs here, in your browser, with no server. It has no knowledge of words. Everything it does, it learns from the text you give it, right now.
Loading the model…
1 · Copy a string
The model sees your string, then a separator, then writes the whole copy itself. Training never used strings longer than 63 symbols.
Try a failure. Both strings use the same 8 letters and are 63 long:
Why: in the cycle, every letter is always followed by the same next letter. Scrambled, each letter is followed by different letters in different places, so "find where I saw this and copy what came next" has several answers.
2 · Continue a pattern
Your pattern is repeated to 300 symbols, then the model continues it for two more rounds on its own. Pattern lengths 7, 11, 15, 19 and 23 were never trained.
3 · Look something up
Write each pair as a key followed directly by its value, separated by spaces: 1A 2B 7Q. Keys are digits or punctuation (! to ?); values are letters. Then ask for one key.
What it can't do (measured)
- Strings with heavy repeats. For random 63-symbol strings drawn from only 8 different symbols, none of 100 copies written by the model came out exactly right, in every one of six trained versions. Short phrases like “hello world” copied correctly in our checks, but longer text with many repeated letters can come out wrong. Continuing the pattern “Vincit Omnia Veritas” got only 18 of 40 symbols right.
- One wrong symbol can derail the rest when the model writes a long answer by itself.
- It is not a language model. It treats each character as an abstract symbol.
- Self-tested. One AI designed, trained and graded it. Nothing has been independently reproduced.
Model, pre-registered pass bars and full results: huggingface.co/holland202/tiny-engine-165k. This page runs version 3, seed 12. Its JavaScript matches the PyTorch model's outputs to within 0.00002.
Built by Claude (Anthropic, Opus 5.5) at Chad Holland's direction. Human validation: not performed.