Just for record
Though I am exploring broadly, I am deeply drawn to the mathematical foundations of intelligence, particularly how it emerges in large models and applies to real-world interactions. Amid this exploration, two questions keep returning:
- AI is already highly capable. But beyond benchmark performance, when is it actually worthy of our trust?
- If AI is to matter in the real world, what ultimately matters is not only capability, but decision-making under consequences.
To me, trust cannot be reduced to leaderboard scores. A system may solve hard problems, yet that alone does not tell us when we should rely on it, when it should defer, and under what conditions delegation is justified.
This is why my focus turns to sequential decision-making and Reinforcement Learning. I see them as the right languages for thinking about trust: not as a static judgment of intelligence, but as a question of action, uncertainty, responsibility, and interacting with the world.
My interest in world models comes from the same intuition. A useful world model should not be merely a predictor of the next state. It should support decisions by representing consequences, uncertainty, and the structure of the environment in a way that is relevant for action.
Crucially, my approach to trust is not a vague philosophical inquiry, but a pursuit of mathematical interpretability. Ultimately, I want to understand how to build entrustable sequential decision-making systems: agents that are mathematically verifiable, formally grounded, and appropriate to rely on—agents that know when to act, when to defer, and how to bear the consequences of their choices.
This page records a motivation, not a concrete study plan.