Superintelligence Must Be Open Source: Hexo Labs
Techstrong.ai, Tuesday, June 2nd, 2026
Hexo Labs released an open-source multi-agent framework for self-improving AI that evolves beyond AGI.
Palo Alto-based Hexo Labs developed Self-Improving AI (SIA), a framework using three interoperating agents that learn through iterative cycles of hypothesis generation, experimentation, and refinement.
The system demonstrated a 350x performance improvement over human-guided learning on OpenAI's machine learning engineering benchmark. SIA uses two distinct models, one for scaffolding and instruction refinement and another specialized open-weight model for domain expertise that adapts based on task outcomes.
The company has partnered with Stanford, Oxford, and UC Santa Barbara to explore applications in scientific discovery. CEO Kunal Bhatia argues that superintelligent systems should remain open source to prevent power being concentrated in the hands of a few corporations.