A chronological record of the developments that shaped modern generative AI. Each entry links to the primary source, not a summary of one.

2017-06: Attention Is All You Need

Vaswani et al. introduce the transformer architecture, dropping recurrence and convolution entirely in favor of self-attention. Nearly every model on this page after 2017 is a transformer or a direct descendant of one. Paper

2020-05: GPT-3 and Few-Shot Learning

Brown et al. show that a 175-billion-parameter language model, ten times larger than any prior non-sparse model, can perform new tasks from a handful of examples in its prompt, with no gradient updates. This result is what made "just scale it up" a credible strategy rather than a hope. Paper

2022-04: Latent Diffusion and Stable Diffusion

Rombach et al. show that running the diffusion process in a compressed latent space, rather than directly on pixels, makes high-resolution image synthesis dramatically cheaper to train and run. This became the foundation for Stable Diffusion, released as open-weight software the same year and the reason image generation went from a handful of well-funded labs to anyone with a consumer GPU. Paper

2022-11-30: ChatGPT

OpenAI releases ChatGPT, a conversational interface over a GPT-3.5-class model. It reaches a million users within five days, becoming the fastest-adopted consumer software product at the time. This is the moment "generative AI" left research labs and became a mainstream, daily-used product category.

2023-03: GPT-4

OpenAI releases GPT-4, its first publicly available multimodal model (accepting both image and text input), scoring in the top 10% on a simulated bar exam among other professional and academic benchmarks. Technical report

2023-06

State of GPT by Andrej Karpathy A comprehensive presentation on the general state of Generative AI made possible by GPT.

Later developments

The pace after 2023 accelerated enough that a fixed list goes stale fast. For current-generation models (reasoning models, the latest frontier releases), see the model landscape section on the Overview page, which is updated as new releases ship rather than maintained as a historical record.

GenAI Chronology