Wonder News
Analysis of the AI models and world-generation systems that shape how playable game spaces are created.
Models Sakana AI's Fugu and Fugu Ultra package multiple frontier models behind one OpenAI-compatible API, a bet that game-building agents need coordination, verification, and long software loops more than another single-model leaderboard claim.
Read article →
Models GLM-5 and GLM-5.1 move Z.ai's open-weight model line from code assistance toward long-running software agents, a shift that matters for AI-generated games only if the agent can test, repair, and ship playable systems.
Read article →
Models A new Google DeepMind report by Shane Legg, Marcus Hutter, and colleagues maps four routes from human-level AGI to superintelligence, and treats the hard part as the speed and control of what comes next.
Read article →
Models The WHAM research line evaluates generated play through consistency, diversity, and persistency, making its Quake II demo a test of controllable game-state memory.
Read article →
Models Roblox's open 3D model gives creators a Studio and Lua path for generating meshes, with behavior, physics, and edit control becoming the next tooling test.
Read article →
Models DeepMind's latest Genie sustains interactive 720p environments for minutes, a capability jump for generated games that still depends on rules, persistent state, and editable structure.
Read article →