There are problems that are relatively well-defined. The mechanisms are mostly understood, the interfaces are clear, and the work is mainly about implementing, integrating, optimizing, scaling, or automating something reliably. These problems can still be technically difficult and highly valuable.

But there is another class of problems that I find myself much more drawn towards. These are problems where the system itself is not completely understood. There may be hidden mechanisms, interactions between many components, uncertainty, feedback, adaptation, emergence, or behavior that cannot be understood by looking at individual pieces alone. Solving these problems requires some combination of modeling, experimentation, computation, engineering, and scientific thinking.

This distinction exists almost everywhere.

In data science, one side may involve reporting, metrics, experimentation infrastructure, forecasting, or optimization around well-defined business questions. The other side may involve trying to understand the hidden state of a customer, a market, a physical system, or an organization, how that state evolves, what causes certain behavior, and how interventions change the system.

The same distinction exists in computer science. There are classical problems around algorithms, systems, performance, storage, networking, and reliability. But there are also problems involving games, mechanism design, distributed coordination, blockchain, multi-agent systems, and other settings where behavior emerges from interactions between many components.

AI has the same split, perhaps even more strongly.

One side of AI is about using models effectively: inference optimization, serving, latency, throughput, memory efficiency, API integration, retrieval pipelines, workflow automation, and agents that automate well-defined tasks. These are important engineering problems.

The other side begins when the behavior of the model itself becomes part of the problem.

How does a model acquire a particular behavior during pre-training or post-training? Why does one data mixture work better than another? How do model capabilities emerge as scale, data, initialization, optimization, and architecture interact? How do we shape behavior through preference learning, reinforcement learning, or alignment methods? How do we understand robustness, failure modes, deception, long-horizon behavior, uncertainty, interpretability, or safety? How should an agent behave when the task is ambiguous, long-horizon, partially observable, or requires forming and testing hypotheses rather than simply executing a workflow?

These questions are much less like ordinary software integration. They require experiments, measurement, models of behavior, competing hypotheses, and often ideas borrowed from several scientific traditions.

That is also why modern AI can be understood from many different lenses. Researchers study neural networks through information theory, dynamical systems, statistical mechanics, optimization, probability, geometry, control, and other perspectives. Even generative modeling sits at the intersection of several of these ideas: probability and information, dynamics and transport, energy and entropy, learning and adaptation.

A lot of this probably connects back to my early exposure to control engineering and later to complexity science. Over time, I kept returning to four broad ways of looking at systems: dynamics and control, information and computation, thermodynamics and entropy, and evolution and adaptation.

These have continued to appear in different forms throughout my work.

Control engineering introduced me to state-space models, feedback, Kalman filtering, dynamics, and system identification. Data science exposed me to uncertainty, behavior, experimentation, markets, and decision-making. Machine learning brought representation, prediction, optimization, and adaptation. Causal inference added interventions and counterfactual thinking. Semantic systems and ontology brought another question: how do we represent the world so that information from different places becomes connected and useful? Modern AI brings many of these ideas together again.

That is where I feel the strongest sense of coherence.

I enjoy building production-quality systems, but production itself is usually not what maximizes my curiosity. I enjoy engineering most when it gives me access to a deeper problem: when I can observe a system, ask what is actually happening underneath it, form hypotheses, model it in different ways, run experiments, understand its behavior, and eventually turn that understanding into something useful.

Even an ordinary-looking project can become interesting through this lens.

An EV telemetry capstone could simply remain an analytics project: ingest data, calculate efficiency, build a model, and create a dashboard. But once I start asking what actually generates those signals, the problem naturally opens into multivariate time series, state estimation, dynamical systems, system identification, physics-based modeling, Koopman operators, representation learning, world models, and Physical AI.

The opportunity was already inside the problem. The difference came from how deeply I chose to investigate it.

The same is true in AI. An agent project could simply be about connecting tools and automating a workflow. But if the task becomes long-horizon, uncertain, partially observable, or requires memory, planning, hypothesis generation, adaptation, and learning from feedback, the problem changes completely. It becomes a question about the behavior of an intelligent system.

That is the kind of transition I find interesting.

So for me, it is always about the problem space.

I am naturally drawn towards systems with dynamics, interactions, feedback, uncertainty, adaptation, hidden structure, and emergence, especially when there is enough room to combine scientific thinking with engineering and experimentation.

The question I want to keep asking is:

What system is worth understanding, what is actually happening underneath its observable behavior, and how deeply can I go before that understanding turns into something useful?

That is where I tend to stay curious, keep learning, and bring the strongest differentiation to the work