Frontier AI systems now autonomously allocate capital to improve themselves
Source: Import AI
Researchers at FrontierML have built a benchmark showing that today's most advanced AI models can productively spend money—$10 to $200 per task—to hire computational resources and improve their own performance across 21 real-world ML tasks. The systems are moving from static artifacts to self-directed economic agents making resource allocation decisions. This complicates scaling predictions (they can now buy their way past bottlenecks) and safety assumptions (autonomous self-improvement via procurement introduces new control surfaces). The diversity of solutions agents discover—some fine-tuning diffusion models, others writing rendering programs—suggests these systems are already doing meaningful R&D work rather than simply executing human-designed recipes.