Meta's Muse Spark ran 20 self-improvement loops with 60+ agent runs in 2 hours for under $1
TinfoilTricorn · x · 2026-09-03
In a side-by-side test, Muse Spark 1.3 Ultra (Meta) and Fable 5.1 ran the same prompt for roughly 2 hours. The prompt required continuous self-improvement until independent judges scored the result above 9.5/10.
Muse Spark never stopped: it completed 20 self-improvement loops, launching 3 agents per loop — 60+ agent runs total — all in Ultra mode for under $1. Scobleizer, who shared the test, says Meta has sprinted ahead of Google here.
More from coding & agent
- Namespace partners with Cursor to give cloud agents native Mac/Linux Devboxes — dean_rie · 2026-09-03
- Open-Source MCP Tool Bridges iOS Simulator Context to Coding Agents — ivanzhaowy · 2026-09-03
- Meta's CORAL: An LLM-Native Harness That Continuously Optimizes Production Recommenders — _reachsumit · 2026-09-03
- My Allowlist Was Cited in Five Design Docs and Never Read at Runtime — Thirumalaiboobathi · 2026-09-03
- Developers are replacing --help with atuin's AI for unfamiliar shell commands — braelyn_ai · 2026-09-03
- Reef open-sourced: infrastructure for agents to continuously self-improve from inference-time signals — pliang279 · 2026-09-03