The Evolution From Transformer to ChatGPT

stanfordnlp · x · 2026-07-20

This thread reviews the evolution of large language models: - In 2017, Transformer emerged, clearly outperforming RNN/LSTM. - In 2018, BERT gained popularity, with Google driving the industry to train, distill, and modify BERT; meanwhile, OpenAI persisted with the decoder-only GPT route. - In 2019, BERT profoundly impacted search, and Google released T5; in 2020, GPT-3 exploded in the community, causing a sensation with GitHub trending, public video demos, and SQuAD performance. - In 2021, Stanford proposed the foundation model concept, attempting to unify the previous Transformer/BERT/GPT narratives. - In 2022, ChatGPT was described as the productized result of pairing GPT-3.5 Turbo with a frontend and infrastructure. The core takeaway: while the public often views ChatGPT as a sudden breakthrough, it was actually the culmination of years of model evolution from a research and product roadmap perspective.

Original post →

More from Models

Models channel →