How we got here
33 milestones since 2017, chosen because they still explain how today's systems work. Each one links to its primary source where one exists.
2025
- 24 Feb 2025 · Product
Claude Code preview
An agent that works in the terminal across a whole codebase. Coding assistants move from autocomplete to delegation.
- 20 Jan 2025 · Model
DeepSeek-R1
An open-weight reasoning model trained largely with RL, competitive with closed models at a fraction of the reported cost.
2024
- 25 Nov 2024 · Standard
Model Context Protocol
An open protocol for connecting AI apps to tools and data. Integrations become reusable.
- 22 Oct 2024 · Product
Computer use
Claude operates a desktop through screenshots, mouse and keyboard: agents leave the API sandbox.
- 12 Sept 2024 · Model
o1 and test-time compute
A model trained with RL to reason at length before answering. More thinking at inference buys accuracy.
- 13 May 2024 · Model
GPT-4o
One model natively handles text, audio and images, enabling real-time voice conversation.
- 4 Mar 2024 · Model
Claude 3
A three-tier family (Haiku, Sonnet, Opus) with long context and vision.
- 15 Feb 2024 · Model
Sora preview
Minute-long coherent video from text. Video generation joins the frontier.
2023
- 11 Dec 2023 · Model
Mixtral 8x7B
An open sparse mixture-of-experts model. Only a fraction of parameters are active per token.
- 6 Dec 2023 · Model
Gemini 1.0
Google's first model family trained multimodal from the start.
- 28 Jul 2023 · Model
RT-2
A vision-language-action model transfers web knowledge into robot control.
- 18 Jul 2023 · Model
Llama 2
Open weights with a commercial-use licence make self-hosted LLMs a serious option for companies.
- 14 Mar 2023 · Model
Claude launches
Anthropic's first public assistant model, trained with Constitutional AI.
- 14 Mar 2023 · Model
GPT-4
A step change in capability on professional exams, with image input. The bar every lab now aims at.
- 24 Feb 2023 · Model
LLaMA
Strong models trained only on public data. The leaked weights ignite the open-model ecosystem.
- 9 Feb 2023 · Paper
Toolformer
A model teaches itself when to call APIs: an early sign of native tool use.
2022
- 15 Dec 2022 · Paper
Constitutional AI
Harmlessness trained from AI feedback against a written set of principles, reducing reliance on human labels.
- 30 Nov 2022 · Product
ChatGPT
A chat interface on an RLHF-tuned model reaches a mass audience within weeks.
- 6 Oct 2022 · Paper
ReAct
Reason, act, observe, repeat. The loop at the heart of nearly every agent framework.
- 22 Aug 2022 · Model
Stable Diffusion released
A capable image model ships as open weights and runs on consumer GPUs.
- 27 May 2022 · Paper
FlashAttention
Exact attention made memory-efficient by being IO-aware. A key enabler of long context windows.
- 29 Mar 2022 · Paper
Chinchilla
Most large models were undertrained. Compute-optimal training scales tokens with parameters.
- 4 Mar 2022 · Paper
InstructGPT and RLHF
Reinforcement learning from human feedback makes a smaller model preferred over a far larger raw one.
- 28 Jan 2022 · Paper
Chain-of-thought prompting
Asking models to show intermediate steps sharply improves multi-step reasoning: the seed of reasoning models.
2021
- 29 Jun 2021 · Product
GitHub Copilot preview
Code completion from a large model, inside the editor. AI coding becomes a daily tool.
- 5 Jan 2021 · Paper
CLIP
Contrastive training on image–text pairs gives a shared embedding space. It becomes the glue of multimodal systems.
2020
- 30 Nov 2020 · Result
AlphaFold 2 at CASP14
Protein structure prediction reaches near-experimental accuracy. The clearest case of AI solving a long-standing scientific problem.
- 28 May 2020 · Model
GPT-3 and in-context learning
At 175B parameters, a model learns new tasks from a few examples in the prompt, without retraining.
- 22 May 2020 · Paper
Retrieval-Augmented Generation
A generator conditioned on retrieved passages. The pattern behind most enterprise AI today.
- 23 Jan 2020 · Paper
Scaling laws
Loss falls as a smooth power law in parameters, data and compute, turning model building into a forecastable investment.
2019
- 14 Feb 2019 · Model
GPT-2 and staged release
A 1.5B-parameter model writes coherent paragraphs. OpenAI withholds full weights at first, starting the release-norms debate.
2018
- 11 Oct 2018 · Paper
BERT
Bidirectional pre-training then fine-tuning becomes the default recipe for language understanding.
2017
- 12 Jun 2017 · Paper
The Transformer
“Attention Is All You Need” replaces recurrence with self-attention. Almost every frontier model since is built on it.