Attention Is All You Need (2017) ☆

Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser and Polosukhin replaced recurrence entirely with self-attention, making sequence models parallelizable — and therefore scalable on GPUs. Cited hundreds of thousands of times, it is the single most consequential paper in this database: GPT, Claude, Gemini, Llama and nearly everything else instantiate it. All eight authors left Google, mostly to found companies (Cohere, Character.AI, Adept, Essential AI, Sakana).