About
What this course is, and the idea it is built on
Intelligent agent frameworks built on multi-modal large language models are moving the frontier of what technology can do in support of science. This seminar asks a specific version of that question: how can the abilities of such models — including “reasoning” — be used in support of hydrologic model conception, development, training, evaluation and improvement?
The format is reading, presenting and discussing papers from the recent literature — some assigned, some chosen by participants — alongside a group project that designs, implements and tests an agentic system for this purpose. Participants are encouraged to collaborate on a position paper for publication.
Hoshin’s framing: a career that has spanned the period from close to the origins of systems theory and information theory — the mathematical and linguistic characterization of dynamical systems — to the origins of artificial intelligence, the automation of analysis and reasoning.
No exams and no testing. Evaluation is subjective, based on engagement and contribution.
The opening lectures set up a chain of reasoning that the rest of the semester builds on. Everything from calibration agents to metric design comes back to these ideas.
An ordered set of characteristic descriptors, each built from symbols drawn from a finite but extensible alphabet. Data are descriptions. Metrics are descriptions.
Every description must be decoded to be used, and every description reflects representational choices: what to describe (purpose) and how (medium, format, computability).
Representation of persistent structure; storage of re-usable information; reasoning (predictive, inferential, counterfactual); and communication.
Interpretable, informationally relevant, distinguishing, compositional, generalizing — and compact, in the minimum-description-length sense.
But not just any compression: linguistically interpretable and readily decodable compression is preferable.
They learn the generative rules of a language from a large corpus, and excel where the rules are formal — but with enough good data they learn rules we cannot yet write down.
Intelligent algorithms that modify their actions based on feedback. Hoshin's claim: the most important invention of the century so far.
Physical constraints and symmetries → architecture hypotheses → process parameterization hypotheses → property-to-parameter hypotheses. Which of these can be agentified?