The transformer architecture (2017) enabled large language models; ChatGPT's release in November 2022 made general-purpose text, code and image generation a mass-market tool.
Attention Is All You NeedAshish Vaswani et al. · arXiv:1706.03762 (NeurIPS 2017) · 2017-06Introduced the Transformer architecture that underlies today's large language models.Peer-reviewedID vaswani-2017Original ↗RegisterGPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language ModelsTyna Eloundou; Sam Manning; Pamela Mishkin; Daniel Rock · arXiv:2303.10130 (later published in Science, 2024) · 2023-03Estimates around 80% of the US workforce could have at least 10% of their tasks affected by LLMs, and about 19% at least half of their tasks; exposure is higher for higher-wage occupations.Caveat: Exposure measures where LLMs could reduce task time, not whether jobs are lost. Authors include OpenAI staff.Working paperID eloundou-2023Original ↗RegisterGenerative AI at WorkErik Brynjolfsson; Danielle Li; Lindsey R. Raymond · NBER Working Paper 31161 (published in QJE, 2025) · 2023-04In a field deployment with 5,179 customer-support agents, an AI assistant raised issues resolved per hour by 14% on average and 34% for novice and low-skilled workers.Working paperID brynjolfsson-2023Original ↗RegisterExperimental evidence on the productivity effects of generative artificial intelligenceShakked Noy; Whitney Zhang · Science, 381(6654), 187–192 · 2023-07In a pre-registered experiment with college-educated professionals on writing tasks, ChatGPT reduced time taken by about 40% and raised output quality by about 18%.Caveat: Short, well-defined writing tasks in an experimental setting.Peer-reviewedID noy-zhang-2023Original ↗RegisterThe Impact of AI on Developer Productivity: Evidence from GitHub CopilotSida Peng; Eirini Kalliamvakou; Peter Cihon; Mert Demirer · arXiv:2302.06590 · 2023-02In a controlled experiment, developers with GitHub Copilot completed a JavaScript HTTP-server task 55.8% faster than the control group.Caveat: Single, well-specified task; authors affiliated with GitHub/Microsoft.Working paperID peng-2023Original ↗RegisterGenerative AI and Jobs: A global analysis of potential effects on job quantity and qualityPaweł Gmyrek; Janine Berg; David Bescond · International Labour Organization, ILO Working Paper 96 · 2023-08Finds clerical work has by far the highest exposure to generative AI and that the dominant effect for most occupations is more likely augmentation than automation; automation exposure is much higher in high-income than low-income countries.International organisationID ilo-2023-genaiOriginal ↗Register