Research note

I'm researching how specialist AI agents produce better answers by disagreeing than any single model does by averaging. Inspired by Karpathy's LLM Council and the older economic idea that markets outperform planners, each experiment gives agents distinct personas, private incentives, and a visible budget, then studies what emerges when they're forced to argue instead of agree. The Roundtable is the first one.

Each agent is extended with personas, budgets, and an anti-convergence reward so the debate produces new ideas instead of consensus theater.

Research note

A single-page interactive prototype that lets you touch three ad units instead of reading about them: compute-as-currency (rewarded units priced in generations, not impressions), the intent auction (mediation that treats intent as a budgeted, depletable resource), and agentic / A2A (sponsored tool-calls, priced on outcome).

Each unit plays out inside a demo AI app. The left panel is what the user sees; the right panel is the mediation layer, live bids arriving, second-price clears, COGS debiting, the intent budget depleting. A persistent P&L ticker recomputes across every interaction and ends the demo contribution-positive without a single interruption.