Can intelligence
learn differently?
Most systems that appear to learn absorb a corpus once and then stop. This research asks whether a much smaller system can acquire explicit, inspectable knowledge by interacting with a world, and correct itself when that world turns out to be different than it believed.
A learner, a hidden rule, and no training data.
The world below has three switches and a gate. A hidden rule decides when the gate opens. The learner has never seen it. It holds all twenty-seven possible rules at once, picks the action that eliminates the most of them, and discards every explanation the world contradicts.
Press Change the world while it runs. The rule is replaced and the learner is told nothing. Watch it detect the contradiction and repair what it believes.
P = power, K = key, L = lockout. Gold is the leading explanation, violet still survives, struck through is eliminated by evidence.
What this is. A real elimination loop over a real hypothesis space, running in your browser. It converges in a handful of observations and recovers from a changed world without being told. What this is not. A result about large models, a replacement for anything, or evidence that the approach scales. The world is deliberately tiny so the mechanism is visible.
Eight steps, none of them decorative.
Nine properties worth having.
- Learning through interaction rather than absorption
- Explicit expectations, stated before the outcome
- Surprise treated as the signal, not as an error to suppress
- Competing explanations held at once, not collapsed early
- Active evidence gathering, chosen for what it eliminates
- Knowledge held outside the weights, where it can be read
- Contradiction detected rather than averaged away
- Belief repair when reality disagrees
- Continuous learning, with no separate training phase
Status
Know where
this breaks?
If you work on learning systems and can see the failure before I do, tell me. That is more useful to this work than agreement.