Follow This Story · The Signal Archive

From Early AI Research to Generative Models

Generative AI did not appear overnight. This evolving history traces selected milestones from an early AI research proposal through attention, Transformers, language models, and publicly accessible conversational AI. It is a carefully selected timeline, not a claim that every event directly caused the next.

How We Got Here

1955 — An early research agenda. The Dartmouth proposal outlined a 1956 summer research project on artificial intelligence, asking whether aspects of intelligence could be described precisely enough for a machine to simulate. The archive establishes the year 1955, not an exact publication day. Original Dartmouth proposal.

September 1, 2014 — Attention in neural translation. A research preprint described allowing a translation model to attend to relevant portions of source text while generating words. This was one important development in attention-based methods. Original research preprint.

June 12, 2017 — The Transformer architecture. The paper Attention Is All You Need introduced a Transformer design built around attention mechanisms, offering a different approach to sequence modeling. Original Transformer paper.

October 11, 2018 — BERT and language understanding. BERT’s first preprint described pretraining bidirectional Transformer representations for language-understanding tasks. It is a milestone in AI language research, not itself a generative chatbot launch. Original BERT paper.

May 28, 2020 — GPT-3 and few-shot learning. OpenAI’s published research discussed how larger language models could perform some tasks from instructions or a small number of examples. Its results and limits were research findings, not evidence of general human-level intelligence. Original GPT-3 research note.

November 30, 2022 — ChatGPT research preview. OpenAI introduced ChatGPT publicly in a conversational research preview and invited feedback on its usefulness and limitations. The original announcement date differs from some later webpage metadata. Original ChatGPT announcement.

March 14, 2023 — GPT-4 announced. OpenAI introduced GPT-4, describing stronger performance in several tasks alongside limitations and safety concerns. Later changes to OpenAI’s website do not change the original announcement date. Original GPT-4 announcement.

What changed—and what did not

Some milestones introduced research methods; others represented product announcements or public research previews. BERT emphasized language understanding, whereas later generative systems emphasized producing text and interacting with users. Attention research and Transformers are important foundations, but this history does not reduce the development of modern AI to a single invention or organization.

What we’re watching

How research capabilities reach practical products, which limitations persist through successive generations, what independent evaluations reveal, and where the next shift in accessible AI capabilities occurs.

Historical source links lead to original research or announcements. Preprint submission dates and research-preview announcements are distinct from later revisions, releases, adoption, and demonstrated outcomes. Related TGO articles are added once their connections pass separate editorial review.