Notes
Personal reflections, research intuitions, and questions I am still learning how to investigate.
Many of my research intuitions begin as thoughts that arrive unexpectedly, sometimes while I am working on something entirely different. I try to develop them into questions I can investigate. I also recognize that intuition is limited by what I understand. The more I learn, the wider that horizon becomes, and the more critically I can examine my own ideas.
A late-night reflection after negative findings in my representation-learning research: on Plato, the difference between knowing and comprehending, and whether increasingly convincing answers make understanding easier to judge.
When a physical system appears to bifurcate unpredictably, the randomness may belong to the observation rather than the full dynamics. A low-dimensional trace can look discontinuous even when its higher-dimensional source is smooth.
What we call noise may sometimes be geometry in coordinates we have not learned to observe.
I wrote a speculative formulation using latent manifolds, Koopman structure, and conditional flow matching to make the intuition precise enough to criticize. I have not yet implemented or tested the proposal. Read the six-page working note (PDF).
A person eating an apple can be recalled as an action, a category, a color, a shape, a place, or a neighboring event. That observation motivated PNEUMA's multi-aspect memory: one experience should remain addressable through several attributes rather than a single flattened representation.
Building PNEUMA made me ask how to test that intuition: which stored item supplied the evidence, which helped us reach it, and which apparent capability came from an upstream model? These questions now inform my work on evidence selection in ERL.
Autoregressive decoding commits locally. An alternative could refine a latent state iteratively, allowing a representation to settle before it is rendered into tokens. Related work on diffusion and continuous latent reasoning gives me a starting point. I want to understand when that extra refinement helps, and what would make me abandon the idea.
Reading an architecture can create recognition without operational understanding. Reimplementing small systems from primitive tensors exposes the hidden assumptions: tensor shapes, normalization, state, numerical error, and the exact point at which an abstraction stops matching the code.