Ideas under consideration
Questions still open, with the work done on them so far.
| The M-layer experiment | A matrix-exponential provocation became a controlled search for geometry a network can earn from observations. Self-context remains the durable default; continuation remains unsolved.Current question: Can a pointwise self-context prior become a set-conditioned learned connection without inheriting a fixed geometry? |
|---|---|
| Observe the boundary | A self-context MLP and transformer show exactly where nonlinear residual geometry should be observed—and why the rest of the network should remain on Matrix Transport. |
| Personal deblurrer | An exposure-level account of blur that asks which scene and camera motion could actually have produced the picture.Current question: How much of the latent scene and spatially varying motion can one observation identify without inventing certainty? |
| Photonic transport field | A renderer that compiles illumination into an adaptive sparse transport field, then lets the camera gather from already illuminated surfaces.Current question: Can exact visible source regions and camera-boundary reuse remove the remaining view-dependent reconstruction cost without restoring path enumeration? |
| Evolving denoiser | A denoiser that must account for which detail belongs to the scene, which belongs to corruption, and why.Current question: What joint law can carry radiance, residual, scale, and causal ancestry without silently deleting real detail? |
| The Kolmogrov question | A fixed-width candidate map that keeps exact identity outside the map, measures what, where, and when separately, and treats collisions as retrieval work rather than truth.Current question: Can bounded, independently addressable measurements preserve useful deformation neighborhoods better than simpler equal-bit sketches without impersonating exact identity? |
| File Manager | A file manager growing into a local computing system: native interface, exact search, durable state, policy, and network access.Current question: Can every layer remain inspectable and replaceable without giving up a coherent everyday tool? |
| The unembedding system | A language-model output head that keeps the usual parameter budget while allowing the backward vocabulary subspace to move with context.Current question: Can a context-rotating backward aperture improve useful vocabulary coverage and specialization without sacrificing the stability of an ordinary tied head? |