LessWrong (30+ Karma) “Thoughts by a non-economist on AI and economics” by boazbarak
Nov 5, 2025
Boaz Barak is an academic and blogger specializing in theoretical computer science and policy, offering non-economist insights on AI's impact on economic growth. He discusses the historical stagnation of living standards prior to recent prosperity surges. Barak elaborates on measuring AI capabilities through the METR framework, factors influencing economic growth, and the unique sigmoidal relation of task performance. He poses intriguing questions about AI's potential to revolutionize productivity and GDP growth, likening AI to an increase in the labor force.
AI Snips
Chapters
Books
Transcript
Episode notes
LLM Time Horizons Grow Exponentially
- Meta's METR graph shows LLMs' solvable task horizon growing exponentially over time on a log scale.
- Boaz Barak assumes a 6-month doubling time, implying task horizon quadruples each year.
Benchmarks Overstate Real-World Reach
- The intercept (absolute task lengths models can handle) depends on reliability, task type, and benchmark bias.
- Barak expects a substantial 'messiness tax' when moving from benchmarks to real-world tasks, lowering the intercept.
Inputs Drive The Exponential Slope
- Many exponential inputs (compute, staff, data, capital) drive the slope of capability growth.
- Sustaining exponential input growth is hard, but so far investment shows no sign of slowing and may have accelerated post-2024.
