Experimental AI
The Little Bastard
A controlled runner for observing how local language models choose actions, respond to consequences, and fail in small generated worlds.
The work
Make behaviour measurable.
The project turns fuzzy claims about model agency and learning into bounded experiments with explicit conditions, observable actions, and recorded outcomes. Its architecture keeps the world, visible observation, memory, optional policy, model, and action layers separate so a result can be inspected instead of hand-waved.
The generated worlds are deliberately constrained. The model is not handed semantic hints or hidden progress, and controls can vary the world or strip it back to test whether an apparent behaviour survives the change.
Poor behaviour is valid evidence. The point is to understand what happens, not tune every run until the model looks clever.
- Deterministic generated worlds
- Separated world and observation layers
- Explicit memory, policy, model, and action boundaries
- Inspectable action outcomes
- Controls for world and prompt conditions
- Evidence-first experimental runs
Controls, observable actions, and recorded outcomes keep the work grounded. Unfinished questions stay unfinished until the evidence says otherwise.