convert all of your problems to text. this is what we do at vetinari and just continues to hold true. recreationally i’ve been building an engine designed for agents to oneshot an entire networked multiplayer game. there’s a lot of bizarre abstractions.
first: a three layer divide. the core engine is harshly separated from all game functionality via a constrained scripting API. game functionality is then also divided from the client. contrast this e.g. with [networkvariable] and client code side by side in for a unity playercontroller.cs
second: json/websocket networking transport layer. just give up on FPS or fighting games. once you give up on those two incorrigible genres, you can implement abstractions that make game design incredibly favorable to agents. if every message is a json request, you can now define a CLI that allows agents to design scenarios, test mechanics, and implement its own feedback loop.
third: level design in ascii. if you want to represent 2d or 3d spaces in text, tarn adams solved this decades ago with dwarf fortress. this would be painful for a human, very easy for an agent. now procedural generation can take ascii level files and construct worlds accordingly.
fourth: game clients require juice; game feel. hit a minecraft zombie, it vibrates a bit, flashes red, gets knocked back, makes a sound. juice can be described best as feedback density. for most of gaming history, juice is hand tuned ad hoc by designers/devs for a particular mechanic. turns out, agents really aren’t bad at juice. skip the ad hoc. define 50 juice primitives, wire them up as a juice package, and voila.
routinely the question to ask is how to convert this historically difficult, manual process, to being text first. if you unlock text, you unlock the profound capabilities of language models.
Source by khaled messai
