The problems we're trying to solve
LLMs excel when the world is explicit. The interesting problems are implicit.
Benchmark leaderboards reward one class of problem. LLMs and the systems built on them excel when the world is explicit. They write code, summarize documents, answer questions and follow instructions. They solve problems where the objective is defined and success can be verified. Those capabilities transformed software development.
They are not the problems this work is about.
The implicit
The work lives where things are ambiguous: belief systems, enterprise world modeling, reasoning under uncertainty. The most interesting problems are rarely explicit.
- Organizations hold conflicting beliefs.
- Information is incomplete.
- Evidence arrives over time.
- Assumptions evolve.
- Decisions are made long before certainty exists.
Decisions are made long before certainty exists.
A different class of challenge
Explicit problems reward a better answer to a well-posed question. Implicit problems need something else: a model of what is believed, by whom, on what evidence, with what confidence, and what has changed since. That is a different challenge, not a harder version of the same one.
It is also the lens for reading model releases. Leaderboards matter less than what a model makes possible to build for systems that operate under ambiguity.