These are the mental models I think with — the durable lenses behind the essays and the research, not essays in themselves. Each is a compact idea I keep returning to. The one they all sit under: to understand anything is to build a model of it.
To understand anything is to build a model of it.
Understanding isn't storing facts; it's building a structure that lets you predict and explain. Curiosity is just the standing drive to keep making that model better.
It's why world models pull at me, and why I think AI belongs to the same story: it is a model of intelligence itself, narrow for now. Modeling is the central act; everything else here is a mode of it.
Intuition is just implicit models.
Understanding comes in two registers. The explicit one runs through language and reasoning; the implicit one lives in skill, perception, and pattern — trained in the background through practice, consolidation, and sleep.
Know-how is an implicit model; so is intuition — compiled experience, not magic. Which also means it can be quietly, systematically wrong when it was compiled from biased data.
Don't choose between depth and breadth. Oscillate.
Breadth versus depth isn't a fixed trade-off, it's a loop. Go deep enough to earn real understanding of one thing, then step back and integrate it with everything else — then switch back.
The switching itself is the skill, and it's what keeps a holistic instinct rigorous rather than vague: integration without depth is just breadth with opinions.
Models operate at every scale; sliding between them is a way of seeing.
From atoms to the cosmos, from milliseconds to evolutionary time — moving across spatial and temporal scales gives a calm, holistic view, and usually a truer one.
It isn't only a worldview. In machines it becomes concrete: intelligence that learns on many time scales at once. That's the research below.
Dreams and stories are modeling run offline.
Dreams simulate scenarios safely, testing reactions and outcomes; fiction and film let you inhabit worlds you haven't lived. Both are the mind augmenting its own training data.
And both can overfit: when the day's experience is skewed or hard, the augmentation amplifies the bias instead of correcting it.
The dominant trajectory of AI treats learning and deployment as separate phases: pre-train on a vast corpus, freeze the weights, ship the artifact. For agents that live in static, predictable environments (chat windows, retrieval pipelines, code completion) this division has been spectacularly successful. For agents that have bodies, it has not.
The next frontier in embodied AI is not a larger pre-trained model, but learning that happens on many time scales at once — milliseconds to years.
Designing systems around a single time scale is the structural reason embodied autonomy keeps stalling at the demo-to-deployment gap. The jump from single- to multi-time-scale learning is qualitative, not incremental.
Useful learning in an embodied agent happens at four overlapping rates: millisecond fast updates, seconds-scale test-time adaptation, minutes-to-hours continual learning, and the days-to-years consolidation we call continual pre-training. Test-time adaptation is deployable today; consolidation is well-trodden; fast updates are an established primitive.
Continual learning at minutes to hours is where deployable progress has lagged — the regime where an agent is online long enough to accumulate genuine new evidence, but not so long that we can afford to ship the data back to a training cluster. It is where the open problems are sharpest, and progress here unlocks the rest of the stack.
Adaptation under real latency and energy budgets is not a deployment afterthought; it shapes which algorithms are even admissible. Co-designing the learning rule with the substrate it runs on is how the fast time scales become affordable at the edge at all.