Import AI 468 delivers a packed briefing on the frontier of AI research and policy, covering recursive self-improvement (RSI) ideas, a new benchmark for autonomous post-training, and the delicate interplay between trust, transparency, and the race to develop advanced AI systems.
23 RSI Ideas for AI Development
The newsletter presents 23 distinct ideas for recursive self-improvement — the concept of AI systems that can improve their own capabilities. These range from architectural innovations that enable self-modification to training regimes that incentivize capability acquisition. The ideas reflect a growing recognition that RSI, once considered a distant theoretical concern, is becoming increasingly relevant as AI systems demonstrate the ability to assist in AI research and development.
PostTrainBench: Benchmarking Autonomous Post-Training
PostTrainBench is a benchmark that measures the ability of CLI agents to post-train pre-trained large language models.[reference:52] The benchmark tasks agents with improving the performance of a base LLM on a given benchmark, using bounded compute constraints (10 hours on one H100 GPU).[reference:53] Frontier agents such as Claude Code, Codex CLI, Gemini CLI, and OpenCode are evaluated on their ability to post-train base models like Qwen3, SmolLM3, and Gemma-3.[reference:54]
The leading systems reach roughly 23% weighted average on PostTrainBench versus 51% for official instruction-tuned models, revealing a significant gap that highlights the difficulty of automating post-training.[reference:55] The benchmark is designed to track progress in AI R&D automation and study the risks that come with it.[reference:56]
Trust and Transparency in AI Racing
The newsletter also examines how trust and transparency interplay with AI racing dynamics. As companies compete to develop and deploy increasingly capable AI systems, questions of transparency — about capabilities, safety measures, and incident reporting — become critical. The tension between competitive advantage and the collective need for safety and accountability is a central theme, with implications for regulatory approaches and industry self-governance.