Research
I am interested in what science costs, and in whether that cost can be studied rather than simply absorbed.
Where this comes from
I came to machine learning from particle physics, where a single measurement takes years, the instrument is as much an engineering problem as a scientific one, and the expensive part is deciding what to measure next. That experience shapes what I study now. Producing a correct answer is only the first step. The next challenge is scaling that success, and scale is often limited by resources. Resource use, in turn, depends on how a system represents and organizes its actions. In autonomous driving, for example, choosing a route, changing lanes, and controlling the steering wheel, accelerator, and brakes are actions at different levels of abstraction. A policy—cautious, standard, or aggressive—can shape decisions at any of those levels. I want to identify similarly layered action representations for scientific and engineering systems, so that policies can be defined and improved at the right level.
What I find interesting
A benchmark's scope, quality, methodology, difficulty, and lifecycle matter as much as its headline score. A benchmark can look hard today and saturate quickly; a narrow one can reward overfitting without measuring broad capability.
A Markdown file can suggest behavior. A durable capability needs a workflow, explicit specifications, and hooks that connect those rules to execution. The system must also remain modular enough for users to plug in their own standards. Structured Coding is one attempt to put that idea into practice.
Humans should own intent, judgment, and acceptance; agents should handle bounded execution and produce evidence. The interface between them must be easy for people to use and structured enough for agents to act on—the idea behind PairTeX.
Two systems may both succeed while taking different paths and consuming very different resources. In science and engineering, that difference determines whether a successful result can become a practical method.
Where I think this goes
Once we can benchmark how action and policy representations at different levels affect an agent's path and cost, we can improve the agent at those same levels. The available levers include post-training the model, allocating test-time compute, designing the scaffold, and structuring reusable skills. Improving performance is therefore a systems optimization problem: the objective is not success at any cost, but the best balance of capability, efficiency, cost, and task coverage.
My current bet for STEM is a domain-specialized, post-trained model paired with a lightweight scaffold and well-structured skills. A single general agent and a collection of specialized agents may both automate scientific work. The useful question is not simply which one succeeds, but which reaches the required performance and coverage most cost-effectively.