Notes on things that resist neat answers.
About Me- I'm Haochi Zhang. I work in financial model risk and build applied AI systems.
- The interesting problems rarely have clean answers.
- Models are abstractions. Data is messy. Real systems have to live with both.
- A prototype proves something can work. Building something people can rely on is a different problem.
- Most things here are attempts to think through that gap.
When Intelligence Becomes a Feature
When Copilot gets more capable than ever, where should enterprise AI use cases live?
Orthogonality — 02: Agent Design as Basis Selection
A tool set is not a list of what an agent can do — it is a basis for its action space, and the useful question about a new tool is what direction it adds.
Orthogonality — 01: Choosing Coordinates for Complexity
An architecture is not just a way to organize code — it is a choice of coordinates for a system's variation, and a bad basis makes one conceptual change move five files at once.
LLM 101 — 04: Prompt as Representation
Patching a prompt rule by rule means controlling consequences as if they could be adjusted independently — the deeper choice is which coordinates represent the task at all.
LLM 101 — 03: When Examples Are Necessary
Specifying a task and calibrating its semantics are different jobs — when a category word does not pick out a unique position on its abstraction axis, a few labeled examples locate it through relative structure.
LLM 101 — 02: Why Few-Shot Fails
An example is not a rule shown by instance but a complete observation — an extremely dense conditioning signal that can narrow not just the form of the answer but the model's distribution over hypotheses.
LLM 101 — 01: Prompt as Conditioning
A prompt doesn't command a model — it conditions a probability distribution, reshaping which continuations become likely.