State of play of AI progress (and related brakes on an intelligence explosion)

The author argues that current AI progress is continuing along measured benchmarks rather than accelerating into an intelligence explosion, and that the "AI 2027" forecast overstates how fast AI research will speed up. The argument holds that many popular evaluations such as MMLU, GPQA, MATH and SWE-Bench are training targets for labs rather than held-out tests, so rapid score gains reflect hill-climbing and not general intelligence growth. It also contends that ML research is bottlenecked by messy, judgment-heavy work rather than the narrow compute and implementation tasks the forecast assumes, and it cites Paul Allen's complexity-brake argument and patent data showing diminishing returns.