BERT sweeps the NLP benchmarks
Google's bidirectional encoder sets new records on 11 tasks and becomes the workhorse of applied NLP.
2017 – 2019
The Transformer Era
One paper rewires the field. Attention replaces recurrence, and scale starts to whisper.
GPT-1 shows generative pre-training works
OpenAI demonstrates that unsupervised pre-training plus fine-tuning beats task-specific architectures.
“Attention Is All You Need”
Eight Google researchers introduce the Transformer. Every frontier model since is a descendant of this architecture.
GPT-2 deemed “too dangerous to release”
OpenAI staggers the 1.5B-parameter release over misuse concerns — and starts the public conversation about AI safety.
2020 – 2022
The Scaling Era
Bigger is different. Language models become few-shot learners and images fall to diffusion.
AlphaFold 2 solves protein folding
DeepMind cracks a 50-year grand challenge in biology, later earning the 2024 Nobel Prize in Chemistry.
GPT-3: language models are few-shot learners
175B parameters. No fine-tuning needed — just describe the task in the prompt. The API economy of AI begins.
Chinchilla rewrites the scaling laws
DeepMind shows most large models are undertrained — data, not just parameters, is the bottleneck.
GitHub Copilot ships AI pair programming
Codex-powered completions land in the editor. Developers become the first profession transformed by LLMs.
DALL·E draws what you describe
Text-to-image arrives. An avocado armchair becomes the most famous chair on the internet.
Stable Diffusion goes open source
Weights on the internet, image generation on consumer GPUs. An entire creative ecosystem ignites overnight.
Late 2022 – 2023
The Chat Era
AI meets everyone. A research demo becomes the fastest-growing product in history.
Amazon commits up to $4B to Anthropic
The cloud wars become AI wars: every hyperscaler now backs a frontier lab.
Mistral 7B punches far above its weight
A French startup's first model beats Llama 2 13B — Europe enters the frontier race with €105M and seven weeks of training.
ChatGPT reaches 1M users in 5 days
A “low-key research preview” becomes the fastest-growing consumer product ever — 100M users in two months.
GPT-4 passes the bar exam
Multimodal, top-decile on professional exams, and the new benchmark every lab chases for two years.
Claude debuts with Constitutional AI
Anthropic's assistant launches, trained with explicit principles instead of pure human preference.
Microsoft invests $10B in OpenAI
The largest AI investment to date locks Azure in as the compute backbone of the frontier.
LLaMA escapes the lab
Meta's weights leak within a week of release — accidentally founding the open-weights movement.
Gemini 1.0 launches across Google
Google's answer to GPT-4 ships natively multimodal, in three sizes, across every Google product.
2024 – 2025
The Multimodal & Reasoning Era
Models learn to see, hear, speak — and think before they answer.
o1 thinks before it answers
Reasoning tokens arrive: test-time compute becomes the new scaling axis, and benchmark ceilings shatter.
MCP standardizes agent tooling
Anthropic open-sources the Model Context Protocol — the “USB-C of AI” connecting models to tools and data.
GPT-4o talks in real time
One model, native across voice, vision and text — with sub-300ms conversational latency.
Coding agents go autonomous
Claude Code, Codex and Cursor agents move from autocomplete to closing tickets end-to-end in the background.
DeepSeek R1 shocks the market
Frontier-class reasoning, MIT-licensed, trained for a fraction of Western budgets. Nvidia sheds $600B in a day.
Sora makes minute-long video from text
OpenAI's video model stuns with physics-plausible scenes and sets off the generative-video race.
2025 – today
The Agent Era
From answering to acting. Models use computers, write software and do real work.
Claude 4.5 Opus leads agentic coding
84.2% on SWE-bench. Long-horizon software work — plan, build, test, ship — increasingly runs unattended.
GPT-5.2 ships native computer use
4M-token context and first-party desktop control. The model doesn't just advise on work — it does the work. This week.
AI capex passes $400B a year
Datacenter buildouts hit nation-state scale; compute, power and chips become geopolitics.
GPT-5 unifies chat and reasoning
One model routes between fast answers and deep thinking — and agentic reliability becomes the headline metric.