Economica (Coase, 1937) · 1937
Why do companies exist at all? Coase's answer — coordination inside is sometimes cheaper than transacting outside — is the question AI just reopened.
American Economic Review (Baumol, 1967) · 1967
The cost-disease paper: why services whose productivity can't grow get relatively more expensive forever. AI is the first credible attack on it.
Datamation (Conway, 1968) · 1968
The origin of Conway's law: systems copy the communication structure of the organizations that build them. Your org chart ships.
American Economic Review (David, 1990) · 1990
Factories electrified in 1890 and saw no productivity gain for forty years — until they reorganized work around the new power source. The single best lens on enterprise AI today.
European Review (Strathern, 1997) · 1997
The source of Goodhart's law as popularly stated: "When a measure becomes a target, it ceases to be a good measure."
Journal of Economic Perspectives (Autor, 2015) · 2015
Automation substitutes for tasks and complements the ones that remain — the task-level lens every 'AI takes jobs' claim needs to survive.
Science (Suri & Jack, 2016) · 2016
M-Pesa access lifted ~194,000 households — about 2% of Kenya — out of poverty. The measured case that payment rails are leverage.
arXiv (Google) · 2017
The paper that introduced the Transformer — the architecture underneath every modern AI model, and the starting point for understanding why this wave happened.
NBER (Brynjolfsson, Rock & Syverson, 2018) · 2018
Why transformative technologies depress measured productivity before raising it: the intangible investment comes first and is invisible on the books.
arXiv (Anthropic) · 2022
Anthropic's method for training AI against a written set of principles instead of armies of human labelers — the foundational idea behind how Claude is trained.
arXiv (DeepMind) · 2022
The 'Chinchilla' paper that showed model size and data should scale together — it quietly rewrote the cost equation for the entire industry.
NBER (Brynjolfsson, Li, Raymond) · 2023
The landmark field study of 5,179 support agents: AI raised productivity 14% on average and 34% for novices — the best real-world evidence that AI helps less-experienced workers most.
arXiv (OpenAI) · 2023 · cited
arXiv (OpenAI / UPenn) · 2023
The study estimating ~80% of US workers could see at least 10% of their tasks affected by LLMs — the canonical early framework for AI and jobs.
arXiv (Microsoft Research) · 2023
Microsoft researchers' deep dive into an early GPT-4 — controversial, influential, and a marker of when expert opinion started to shift.
arXiv (Microsoft / MIT) · 2023
The controlled experiment behind the famous '55.8% faster' statistic — the most-cited number on AI coding productivity.
arXiv (DeepSeek-AI) · 2024 · cited
arXiv (Meta AI) · 2024 · cited
Stanford Digital Economy Lab · 2025 · cited
arXiv (DeepSeek-AI) · 2025 · cited
METR · 2025
METR's finding that the length of tasks AI can complete doubles roughly every seven months — the single most useful trend line for planning around capability growth.
NBER · 2026 · cited