Vaswani et al. (Google) replaced recurrence with self-attention—the base architecture of BERT, GPT, and nearly all modern NLP.
On June 12, 2017, “Attention Is All You Need” proposed the Transformer: an encoder–decoder with multi-head self-attention and positional encodings, without CNNs or RNNs.
Training parallelization beat recurrent seq2seq on translation. Soon the architecture spread to vision, audio, and multimodality.
Nearly all frontier LLMs of 2020–2026 are variants of this design. Few papers have so thoroughly reordered the AI stack.
Devlin et al. released BERT, popularizing language-model fine-tuning via masked LM and next sentence prediction.
Google merged Brain/DeepMind efforts into Gemini, a multimodal family natively trained for text, code, and image.
On September 2, 2026 Google launched Gemini 3.8 Flash at introductory pricing and Flash Cyber for trusted defenders via Fairwind, focused on long-horizon software engineering and patching.