Of Notes

Complex Systems Thesis

4 min read · revised 20 October 2026 · referred to by 1 note · ·

Initial Thoughts

  • Work on something in world models - finding a way to make them useful
  • Work on distribution shift of systems like one's house when a roomba is cleaning it or a mine on mars that a robot needs to navigate or mine in
    • Not just model drift tho
  • Finding ways of usefully deploying world models - JEPA style.
    • How do we turn observational world models into intervention models
    • Adding aleatoric uncertainty of multiple possible futures - similar to noise conditioning
    • Can we decompose world models into foundational components and sub components that are more malleable for shifts in the world we are operating on (eg. adding a wall to a roomba map).
  • Maybe in the spirit of continual learning
  • How can a learned world model construct and maintain a decision-useful internal representation of a complex environment that allows an agent to predict, plan, reason under uncertainty, and adapt as both the environment and the agent's interaction with it change?
  • Dark matter / astral topology brainstorm: 2026-08-29

Towards a Theory of Dynamical Transferability

  • Project idea discussed with Ekatarina Shatalina from LIHE King's.
  • Main question: Can shared dynamical structure tell us which datasets will help a model learn a new complex system from limited data?
  • Idea originated in neuroscience: Rather than pretraining exclusively on neuroscience datasets, we wanted to investigate whether models could benefit from data originating in completely different domains.
    • Originated from deep-mind pre-training principles for their cyclone tracking model.
  • Objectives:
    • Method for searching across systems for most adequate transfer-ability (open source package).
    • Demonstrate performance on neuroscience/energy cases.

Frontier Thoughts

So talking to a guy from OpenAI my previous thesis directions might not be the most forward looking and that maybe the industry has "moved on".

Some concerns are that AGI will replace most math based work like just proving theorems. I think that if we rae stil unable to keep systems running like cloud/banks or we are still in a world where solving it causes a lot of issues then we are not as far as we think. Rather than focusing on one problem and solving it I wanna focus on a more abstract idea which is still a very specific problem. The idea is to work on something systematic that does not die with innovation or gets forgotten. There is this idea of reactively being able to better solve problems with passive proactivity which can be done with AI. Take crises and other out-of-distribution singularities.

In the future we dont wanna encounter situations where we are caught off-guard. Although problem solving is critical for humans we need some help from AI.

Take this revolut case, if they had thought at some point to create redundancy with other cloud providers they night not have this current issue. If they had in the background kept track of multiple components of their system and had some intelligence analyzing this continuously they might be more crisis ready. If france had better models of their population they might be better able to react to the riots that broke out (we are not talking about prevention here) just readiness.

Fundamentally I think this is about building intelligent systems which are able to exhaustively explore some working-space and enact scenarios with this working-space as the base (enhanced with llms maybe).

An oranization/physical system/city/government are complex systems with interacting parts. Encoding them into latent working-spaces is tricky but possible.

Uncertainty-Aware Architecture for Action Planning

  • Reframing: not about world models per se, but about action planning improved by an uncertainty-aware architecture and representation learning.
  • Main question: How should a representation be structured so that uncertainty in it is decision-relevant, so planning stays good (and knows when it isn't) as the system changes?
    • Uncertainty only matters if it changes what the agent does - test is plan quality, not prediction accuracy.
  • Kinds of uncertainty (usually conflated):
    • Aleatoric - the world is stochastic, multiple possible futures.
    • Epistemic - the model has not seen this region.
    • Structural - the representation itself is wrong (a wall is added). Not just model drift, and the least studied.
  • Open design questions:
    • Where does uncertainty live: latent state, dynamics, or the planner's value estimates?
    • How does the planner use it: risk-sensitive objectives, deferring or exploring under epistemic uncertainty, replanning when structural uncertainty spikes.
    • Representation learning: separate controllable, uncontrollable and noise factors so uncertainty becomes actionable (controllability-aware latent).
    • Model-based, model-free or hybrid planner - leaning model-based/hybrid since structural shift is hard to handle without predicting consequences.
    • Modularity (local world change maps to local edit) only if it earns its place through planning performance.
  • Connection to the other threads:
    • Dynamical transferability becomes representation reuse: does a representation support good planning across systems with shared dynamics.
    • Crisis readiness is the motivating setting - out-of-distribution events are where epistemic/structural uncertainty should change behavior.
    • Roomba and Mars mine are planning problems with structural shift, so natural testbeds.
  • Evaluation:
    • Regret/return under controlled structural shift in environments with known ground truth first.
    • Calibration of when the agent knows it is out of its depth.
  • Risks:
    • Crowded area (safe RL, Bayesian RL, ensembles, distributional methods) - differentiator has to be separating structural uncertainty and acting on it.
    • Planning evaluation is harder than prediction evaluation.
  • TODO literature search: structural vs epistemic uncertainty in planning, object-centric/modular dynamics, causal representation learning.
/ or Ctrl K
search; empty, it shows notes to revise and lately read
m
mark this note to revise, or unmark it
c
cover maths, proofs and solutions to test yourself
n
uncover the next covered item
y
copy this note’s org-roam link
e
copy this note as markdown, LaTeX kept as TeX
p
the next problem to work; there h hints, s shows the solution, x passes it for today
→ ← j k space
turn the page, when reading in pages
l
on the map, trace the greatest distance, then the longest path
double-click
a formula, to copy its TeX
rest on maths
bind a part of a formula and open the wheel of notes it comes from; scroll or ↑ ↓ widens it, 1–6 goes
?
this card