Ten eras of AI

Vaswani et al. publish 'Attention is All You Need'. BERT and GPT-1 arrive within fifteen months and the architecture of every subsequent commercial system is settled.

GPT-2 and GPT-3 demonstrate that capability rises smoothly with parameters, data, and compute. The field abandons architecture-tinkering for budget growth.

RLHF and SFT turn raw language models into something a non-researcher can address in plain English. InstructGPT and ChatGPT collapse the time between idea and prototype.

ChatGPT crosses 100 million weekly users in two months. GPT-4, Claude 1, Bard, and Llama mark the first proper four-model market.

Function-calling, JSON schemas, and assistants APIs make a model a runtime, not a chatbot. Bots begin doing things, not just answering questions.

GPT-4V, Claude 3 vision, Gemini, and the first native-multimodal Llama derivatives close the input gap. Document parsing stops being a separate product category.

Context windows pass 200K, then 1M, then 2M tokens. Retrieval pipelines are rebuilt on top of native attention and many of them shrink to nothing.

Cursor, Claude Code, Cody, Cline, Devin, Aider: each interprets 'agent' differently. The result is the first ten-billion-dollar agentic market without a settled definition.

OpenAI o-series, DeepSeek R1, Claude with extended thinking. Chains-of-thought become first-class outputs. Cost-per-correct-answer replaces cost-per-token as the operative metric.

Computer-use, browser-use, multi-agent orchestration, and the first credible autonomous loops. Where we are at the time of this entry.