How did Anthropic measure AI’s "theoretical capabilities" in the job market?
Summary
Anthropic discusses a 2023 study that estimated AI's theoretical capabilities in the job market by mapping 22 occupations' tasks to potential performance by LLM-powered software. The blue 'theoretical capability' is based on speculative foresight rather than current model performance, relying on OpenAI/OpenResearch/UPenn work that used O*NET tasks and human labeling to judge potential 50% time savings; researchers acknowledge subjectivity and limitations, emphasizing that 'theoretical capability' does not equal replacement and that timelines are unclear.