thinkAbout

Leaderboards don't show what a model can actually do. I want to force the gap into the open visually and playfully — a task is worth building only if you can see who won at a glance.

• Four granularities, tested apart: context, tools, model, agent — each for the ability it owes
• Only the steps that open the widest gap; if every model clears it, it isn't a question
• Aimed at what resists quantification: the abilities no score reports are the ones I want to see
• Tasks are black boxes: no rules given, only traces — the model has to reconstruct the world
• The eval has to be fun to watch, or nobody watches — myself included

I build skills to fold my chain of reasoning: every vertical has its own core way of thinking, and once it's folded into a skill, that thinking can scale across domains.

• A skill is a folded chain of reasoning, not a manual of steps
• The core idea lives in the decisions: which path to take under which condition is the most valuable part of a domain
• The folding has to be general — move to another vertical and you swap the decisions, not the skeleton