Testing Astra's self-driven creativity with tree-search prompting: better variety, still lackluster
creatoroff · x · 2026-09-10
An ongoing experiment on inducing creativity in Astra: combining tree search with a stronger prompt produced more visually distinct simulations that diverged from the model's usual memory-related themes, but results still fell short of impressive. An earlier round had three tasks each run diverging tree searches, generating 117 ideas total — more variety on paper but nearly identical visual styles and mediocre ideas. Key observation: the model is extremely capable under human guidance but struggles to generate good ideas autonomously.
More from coding & agent
- 223 pull requests in 11 days: GitHub exec ships Spring Boot console by managing a fleet of AI agents — DanWahlin · 2026-09-10
- MCP Live-Data Design: Should Agents Search-Narrow First or Get Raw Latest Data? — JuicerSocial · 2026-09-10
- Who's Building Guardrails for the Vibe Coding Wave of Shaky Code? — DavidWells · 2026-09-10
- v0 brings Vercel integrations into chat: one-click Resend, MongoDB, Algolia and more — evilrabbit_ · 2026-09-10
- User puts computer-use agent Astra to work auditing Google Business Profiles and configuring Stripe — gregmushen · 2026-09-10
- Using an AI agent to digest Apple's 6 iPhone Duo developer videos: key iOS 27.1 takeaways — JordanMorgan10 · 2026-09-10