The argument for a ceiling on intelligence

One post argues that intelligence can be measured by how close decisions come to optimal, with a ceiling of 100%. It claims Astra is already 80% optimal on ARC v3 speedruns. For near-term recursive self-improvement, it points to two areas: automating data generation or gathering to fill models’ knowledge gaps, and improving efficiency and reducing costs. Cheaper models, it argues, would enable more automated research.