Discovery

What if a model did not reread its past in words? That question led to the largest completed experiment in this repository: compress long documents into latent vectors, project those vectors into soft tokens, and compare the result with a text summary buffer. The experiment began with an attractive hypothesis: Latent Pager Memory can preserve useful information with less generation cost than a text buffer. The data supported that hypothesis and exposed a dangerous price....

The dangerous answer is not always the wrong one. It is the wrong one delivered with enough confidence to stop the search. This month I revisited two experiments in the repository. One measures whether models admit uncertainty across factual, reasoning, ambiguous, boundary, and impossible questions. The other samples the same model repeatedly to measure agreement and entropy. Together they test a practical claim: Uncertainty becomes useful when we measure both confidence within one answer and disagreement across several answers....

A fact remains the same for everyone who sees it. A social fact changes with the observer. Eve thinks the book is in the cupboard. Henry knows it moved. Bob saw Henry watching. One room now contains several incompatible realities. The repository’s social cognition suite tests whether models can keep those realities separate. I combined two recorded experiments around one claim: Modern language models can track explicit nested beliefs, but their broader social inference depends strongly on contextual evidence....

When there is no correct answer, what remains to measure? Taste sounds private and slippery, but it leaves observable traces: repeated choices, confidence, sensitivity to framing, and disagreement between judges. The repository contains an experiment across art, poetry, music, design, and prose that turns those traces into data. The claim under investigation is deliberately limited: Language models produce stable, model specific aesthetic preference profiles, even when no option is objectively correct....

Text lets an incorrect explanation remain elegant. A simulation is less polite. The bridge falls, the orbit escapes, or the ball passes through the floor. I wanted to test a narrow version of a larger idea: Can a model with fewer than one billion parameters learn enough structured code to generate small interactive physics worlds? The repository contains a completed Qwen3 0.6B LoRA run built from synthetic p5.js examples. It also contains a developmental learning proposal for MuJoCo....

A camera moves left. Or perhaps the subject moves right. The pixels alone do not tell us which coordinate system the sentence meant. That ambiguity became the experimental question for this month: Does adding explicit camera coordinates make a movement label meaningfully more reconstructable than ordinary cinematic language? The larger camera dataset project in this repository proposes generated environments, depth estimation, scene reconstruction, scripted camera paths, and captions derived from those paths....