Back to How Do We Learn Dec 2025 Aliveline

Day 30

Second to last day! What would I like to get done before the end?

• Code wise
  • Re-run Kato shared
  • Run Atanas fidelity/stationarity tests
  • Update documentation/writeup with results
• Convert my various musings/thoughts/questions throughout this into draft-y Evergreen Questions, so I can refine & sharpen them later
• Reflections on this process while it’s still fresh
• Overall writeup, though I’ll let all of this sit for a little before doing this – mostly just want to make sure raw materials are in place while it’s fresh

Kato shared is running. Did some refactoring of Kato/Atanas code to avoid code going out of sync, it’s much cleaner and clearer to see what is going on now. Tons of documentation updates to get it to a finalized place (for now).

Thinking about this Evergreen Questions format… here’s the original Evergreen Notes:

• Evergreen notes should be atomic
• Evergreen notes should be concept-oriented
• Evergreen notes should be densely linked
• Prefer associative ontologies to hierarchical taxonomies
• Write notes for yourself by default, disregarding audience

What am I thinking for Evergreen Questions? Why do I need a separate format beyond Evergreen Notes?

• Original reasoning was that Evergreen Notes require a lot of maintenance – because they are densely linked, any semi major changes will require changing a lot of them. Evergreen Questions at least a priori seem like they may need less, because questions remain valid even if they are answered; how refined/sharp they are may change over time. Also loosely inspired from Questions are not just for asking
• I also like the idea of practicing the skill of converting questions into falsifiable / testable hypotheses and actual experiments. Each Evergreen Question is in a good state once I am able to sharply define an associated hypothesis/experiment (ideally just a single one to make sure questions are atomic)
• Probably I should just try it out and see what happens. Let’s do that

From Day 23 is my rough list of Evergreen Questions from this aliveline:

• Each neuron is a DD-DC
  • Can’t you test this by seeing what happens if you perturb a single neuron and see if it behaves like a feedback controller?
• Moving faster allows enables/strengthens learning through proprioception
  • Hypothesis: because each neuron is a DD-DC, speed increases quality & quantity of feedback
• Locomotion is implemented as loops in a low-D phase space in C Elegans
• Each brain has its own neural coordinate system to implement the same behavior
• Neural population activity encode tasks like motor control and even higher order cognition as manifolds
  • As things become more predictable, tangling decreases (manifold geometry becomes simpler)
  • hypothesis: As things become more predictable, some analog to Kolmogorov complexity goes down
• What might a useful KC analog definition look like?
  • Minimum number of neurons that need to be activated to achieve a particular neural state space trajectory?
  • what if we frame this in terms of feedback/control??
• Theories of the brain have shifted from a single-neuron view to global population dynamics due to advancements in measurement technology
  • What’s the natural extension here? What kind of measurement would enable this?
• The space of neuromodulators is massive, even in a small nervous system like C Elegans
  • Ex: Cytokines paper
• Can the Tracy-Widom distribution (or other universality class distributions) describe the transition of neural population activity before/after a rule is learned?
• Behavior is a top-down computational constraint on a neuronal network
• How does DNA bridge the micro (local neuronal rules) to the macro (evolutionarily relevant behavior)
  • hypothesis: perhaps in local neuronal rules are in the language of feedback control laws?
• What is the appropriate measure of ‘complexity’ or ‘representational capacity’ for a brain?
• How does the parallelism of the brain tie into the feedback control / nonlinear dynamical system view?
• This comment: In CN, the trend is to take algorithms from computer science and statistics and map them onto biology. What’s far rarer is extracting new ML algorithms from the biology itself.
• What if intelligence/the brain were simple and they seem complex because the environment is complex? simple in the sense of having a relatively small set of rules

Also asked codex to pull out Evergreen Notes/Questions from the ideas in this folder to see if it gives me interesting ideas. Results:

Process / Meta‑Research   - Constraints as a motivation engine — why “absurd objectives” reliably create momentum.   - Problem‑creator mindset reduces skill bottlenecks — when reframing beats solving.   - Fast iteration beats perfect framing in short timelines — when to pick a path vs keep exploring.   - Failure can be designed to be interesting — criteria for “good failure.”   - Research pace is gated by social bandwidth — author outreach as default step.   - LLM co‑working needs explicit loop design — prompts, checkpoints, and autonomy boundaries.   - Asking for help earlier changes research speed — how much is “enough” independence?   - Negative results are still structure — how to turn “didn’t work” into a useful note.

  Methods / Measurement   - Stationarity assumptions quietly dominate neural dynamics analyses — when they break.   - “Fidelity” vs “stationarity” as separate validation axes — why both matter.   - Freely moving vs immobilized data may be different systems — when manifolds disappear.   - Dataset conditions are first‑class confounds — stimulus regimes as regime switches.   - Small neuron subsets can recover global structure — what that implies and what it doesn’t.   - When a manifold is real vs epiphenomenal — criteria for “computational relevance.”

  Neuro / Theory   - Behavior as a top‑down constraint on network dynamics — not hardwiring, but shaping.   - Degeneracy vs universality — stable behavior with unstable micro‑implementation.   - Context‑gating as a memory strategy in tiny brains — minimal memory without overwrite.   - Manifolds as reused scaffolds — reuse without catastrophic forgetting.   - Learning as state‑space geometry change — rotations/contractions as signatures.   - Local rules → global loops — how local feedback might yield conserved dynamics.   - Genetic “reachability” instead of explicit encoding — behavior as a basin, not a script.   - Control‑theoretic language for neurons — feedback laws as a unifying micro‑vocabulary.

  AI / ML / Cross‑Domain   - Emergent low‑D structure in both brains and models — what’s the shared pressure?   - Macro behavior selection vs micro implementation freedom — analogy to model pruning or lotteries.   - When AI models should be trained near criticality — why structure appears at phase edges.   - “Simple rules, complex environment” as a brain hypothesis — where complexity really lives.