Mediator.ai pairs Nash bargaining with LLMs to systematize fairness
Mediator.ai hit Show HN April 21 (141 points) with one of the cleaner ideas of the week. They take Nash's 1950 result on fair cooperative bargaining — which requires each party to provide a utility function, the part that always made the math useless in real life — and use an LLM to estimate utility from natural-language conversation.
The trick is that LLMs are bad at producing absolute utility numbers but good at comparisons. So Mediator drafts candidate agreements, pits them against each other, scores each against both sides' needs, round after round, until no draft can do better. That is effectively running Nash bargaining over the latent preferences both sides expressed in plain English. Founder equity splits, shared-living disputes, contractor disagreements — anywhere two sides want a deal but neither wants to be outmaneuvered.
This is the kind of agentic application that is actually new. Not "ChatGPT but for X." It is pulling a 75-year-old game theory result out of cold storage and using LLMs as the missing utility-extraction layer. If it works at scale, it changes how a whole category of soft-conflict negotiation gets done.
Site at mediator.ai, HN thread at news.ycombinator.com/item?id=47835411. The economics paper that powers it is Nash's 1950 Bargaining Problem — worth a read either way.
← Back to all articles
The trick is that LLMs are bad at producing absolute utility numbers but good at comparisons. So Mediator drafts candidate agreements, pits them against each other, scores each against both sides' needs, round after round, until no draft can do better. That is effectively running Nash bargaining over the latent preferences both sides expressed in plain English. Founder equity splits, shared-living disputes, contractor disagreements — anywhere two sides want a deal but neither wants to be outmaneuvered.
This is the kind of agentic application that is actually new. Not "ChatGPT but for X." It is pulling a 75-year-old game theory result out of cold storage and using LLMs as the missing utility-extraction layer. If it works at scale, it changes how a whole category of soft-conflict negotiation gets done.
Site at mediator.ai, HN thread at news.ycombinator.com/item?id=47835411. The economics paper that powers it is Nash's 1950 Bargaining Problem — worth a read either way.
Comments