Poetiq

Builds harnesses that raise base-model performance without retraining — scaffolding around other labs’ models rather than models of its own. In this corpus its harness on GPT-5.2-xhigh took ARC-AGI-2 to 75% at under $8 a problem without touching the base weights, it later orchestrated Gemini, GPT and Claude to 55% on Humanity’s Last Exam, and in August it unveiled a self-optimizing Metasystem that upgrades its own harnesses, prompts and code, arguing recursive self-improvement is the fastest path to superintelligence.