Import AI 436: Another 2GW datacenter; why regulation is scary; how to fight a superintelligence
A mixed-signal research newsletter covering agentic AI training infrastructure and existential risk scenarios - useful background for AI safety and governance practitioners.
Key points
- OSGym enables cheap, scalable training of AI agents to operate computers across multiple applications.
- RAND paper models options for neutralising a rogue superintelligence - all carry severe collateral risks.
- Regulation commentary is US-focused and editorial in nature; limited direct APS policy signal.
Summary
This edition of Import AI covers four threads: OSGym, an open-source platform enabling cheap parallel OS-level AI agent training from top US universities; a $900m raise by Luma AI to build a 2GW compute supercluster in Saudi Arabia; an editorial argument that over-regulation creates vetocracy, using US hardware startups as examples; and a RAND paper modelling how humanity might counter a threatening superintelligent AI system. Each thread is technically or strategically interesting but the newsletter is a curated commentary digest rather than primary policy guidance.
Implications for Australian agencies
- Monitor AI safety and long-horizon risk teams may want to monitor the RAND superintelligence threat-response paper as a reference point for extreme-risk scenario planning.
- Monitor Agentic AI practitioners could track OSGym as an indicator of how quickly agentic computer-use capabilities are maturing outside frontier labs.
Implications are AI-generated. Starting points, not advice.
"Import AI 436: Another 2GW datacenter; why regulation is scary; how to fight a superintelligence" Source: Import AI – Substack (Jack Clark) Published: 24 November 2025 URL: https://importai.substack.com/p/import-ai-436-another-2gw-datacenter This edition of Import AI covers four threads: OSGym, an open-source platform enabling cheap parallel OS-level AI agent training from top US universities; a $900m raise by Luma AI to build a 2GW compute supercluster in Saudi Arabia; an editorial argument that over-regulation creates vetocracy, using US hardware startups as examples; and a RAND paper modelling how humanity might counter a threatening superintelligent AI system. Each thread is technically or strategically interesting but the newsletter is a curated commentary digest rather than primary policy guidance. Implications for Australian agencies: - [Monitor] AI safety and long-horizon risk teams may want to monitor the RAND superintelligence threat-response paper as a reference point for extreme-risk scenario planning. - [Monitor] Agentic AI practitioners could track OSGym as an indicator of how quickly agentic computer-use capabilities are maturing outside frontier labs. Retrieved from SIMS, 18 May 2026.