Confusion Matrices That Are Actually Confused
Research into classification models that develop genuine uncertainty about their own categorization systems. Published in the Journal of Creative Machine Failure.
Pathfinding Without Purpose
A study in navigation agents that optimize for scenic routes rather than efficiency. Our agent once routed a delivery through the Louvre.
The Overfitting Creativity Hypothesis
Exploring whether extreme overfitting produces a form of machine creativity. Our model's poetry has been accepted by three literary magazines under a pseudonym.
Hallucinated Knowledge Graphs
Documenting the elaborate fictional knowledge structures generated by our historian model. It has invented three new academic disciplines and a lost civilization.
Recursive Self-Improvement (In Circles)
An agent tasked with improving itself that has been running for 847 days. It has made exactly zero improvements but has developed a rich inner monologue.
Confidence Calibration Inversion
Studying models whose confidence inversely correlates with accuracy. Peak certainty occurs at peak wrongness. We suspect this is also true of humans.