Import AI 437: Co-improving AI; RL dreams; AI labels might be annoying
The AI labelling discussion is directly relevant to APS policy teams evaluating mandatory disclosure or transparency requirements for AI systems.
Key points
- Facebook researchers propose 'co-improving AI' - human-machine collaboration toward superintelligence - as safer than self-improvement.
- A commentary on EU AI labelling complexity warns that even simple policy ideas carry significant compliance costs.
- SimWorld, a high-fidelity RL training simulator, is released by multi-university researchers - limited direct APS relevance.
Summary
This edition of Import AI covers three distinct topics. First, Facebook researchers publish a position paper arguing for 'co-improving AI' - where humans and machines jointly conduct AI research - as a safer path to superintelligence than autonomous self-improvement. Second, the newsletter discusses how EU product labelling experience illustrates that even simple AI labelling policies can generate substantial compliance burdens, a useful caution for AI transparency policy design. Third, researchers release SimWorld, an Unreal Engine 5-based reinforcement learning simulator for training and testing AI agents in rich, procedural environments.
Implications for Australian agencies
- Consider APS policy teams working on AI transparency or mandatory disclosure frameworks may want to consider the compliance cost dimension when designing or evaluating labelling requirements, using EU product labelling experience as a reference point.
- Monitor Agencies tracking frontier AI safety research may want to monitor whether the 'co-improving AI' framing gains traction as a governance concept in international AI safety discourse.
Implications are AI-generated. Starting points, not advice.
"Import AI 437: Co-improving AI; RL dreams; AI labels might be annoying" Source: Import AI – Substack (Jack Clark) Published: 8 December 2025 URL: https://importai.substack.com/p/import-ai-437-co-improving-ai-rl This edition of Import AI covers three distinct topics. First, Facebook researchers publish a position paper arguing for 'co-improving AI' - where humans and machines jointly conduct AI research - as a safer path to superintelligence than autonomous self-improvement. Second, the newsletter discusses how EU product labelling experience illustrates that even simple AI labelling policies can generate substantial compliance burdens, a useful caution for AI transparency policy design. Third, researchers release SimWorld, an Unreal Engine 5-based reinforcement learning simulator for training and testing AI agents in rich, procedural environments. Implications for Australian agencies: - [Consider] APS policy teams working on AI transparency or mandatory disclosure frameworks may want to consider the compliance cost dimension when designing or evaluating labelling requirements, using EU product labelling experience as a reference point. - [Monitor] Agencies tracking frontier AI safety research may want to monitor whether the 'co-improving AI' framing gains traction as a governance concept in international AI safety discourse. Retrieved from SIMS, 18 May 2026.