Updated July 5, 2026
41 checkpoints, from the first formal machine to whatever we are calling the present. Scroll sideways.
Foundations
1947
Alan Turing begins treating machine intelligence as something that can be built, tested, and reasoned about.
The pre-AI era starts with a shift from philosophy to machinery: could a programmable computer exhibit intelligent behavior if it had enough memory, search, and learning?
Foundations
1950
Turing publishes "Computing Machinery and Intelligence" and reframes the question of whether machines can think.
The paper gives AI a public thought experiment: judge machine intelligence by behavior in a controlled interaction, not by introspection into consciousness.
Foundations
1956
McCarthy, Minsky, Rochester, and Shannon propose a summer research project on artificial intelligence.
The field gets a name, an agenda, and a founding myth: intelligence might be described precisely enough that machines can simulate it.
Symbolic and Early Learning
1958
Frank Rosenblatt introduces a trainable pattern-recognition machine inspired by neural computation.
The perceptron turns learning from examples into hardware and math, giving neural networks their first major public wave.
Symbolic and Early Learning
1966
Joseph Weizenbaum builds a pattern-matching chatbot that imitates a psychotherapist.
ELIZA is technically simple but culturally important: people respond to conversational interfaces as social systems before the systems truly understand them.
Symbolic and Early Learning
1969
SRI builds an early mobile robot that reasons over a simplified world and executes plans.
Shakey becomes a canonical example of embodied AI: sensing, symbolic planning, and physical action stitched together in one system.
Expert Systems
1972
Stanford researchers build a rule-based medical expert system for infectious disease consultation.
MYCIN proves that narrow expert knowledge can be encoded into rules, explanations, and recommendations, even when deployment remains difficult.
Expert Systems
1980
DEC uses a rule-based system to configure computer orders, showing commercial value in expert systems.
The lesson is pragmatic: AI reaches production first where the world is constrained, expensive mistakes are common, and rules can capture expert work.
Neural Revival
1986
Rumelhart, Hinton, and Williams popularize learning internal representations with backpropagation.
This becomes a technical hinge: neural networks can now learn useful hidden features instead of relying only on hand-designed inputs.
Neural Revival
1989
LeCun and collaborators demonstrate backpropagation for handwritten ZIP code recognition.
Modern computer vision starts to take shape: local filters, shared weights, and gradient learning applied to real image data.
Benchmarks
1997
IBM Deep Blue defeats Garry Kasparov in a six-game match.
The victory makes AI visible to the public as a specialized system that can exceed elite human performance in a formal domain.
Deep Learning
2006
Researchers show practical ways to train deeper neural networks before today's data and compute scale.
The field starts moving away from brittle feature engineering and toward representations learned by stacked models.
Deep Learning
2009
A massive labeled image dataset creates a shared arena for measuring visual recognition progress.
ImageNet matters because it pairs scale with competition, making model progress legible and comparable year after year.
Deep Learning
2011
IBM Watson defeats champion players by combining search, language processing, and evidence scoring.
Watson shows that AI systems can integrate many imperfect subsystems and still win in a fast, language-heavy environment.
Deep Learning
2012
A GPU-trained convolutional network wins ImageNet by a large margin.
The result changes the allocation of research attention: data, GPUs, depth, and learned representations become the default path for perception.
Generative Models
2014
Generative adversarial networks train a generator against a discriminator.
GANs make image synthesis feel dynamic and competitive, opening a new era of generative modeling before diffusion becomes dominant.
Reinforcement Learning
2016
DeepMind combines deep networks, search, and reinforcement learning to beat a Go world champion.
Go had long represented intuition and astronomical search space. AlphaGo makes self-play and policy/value networks central to frontier AI imagination.
Transformers
2017
"Attention Is All You Need" introduces the architecture that scales into modern language models.
The transformer is the pivotal infrastructure idea: parallel training, long-range token interactions, and a model family that keeps improving with scale.
Transformers
2018
BERT and GPT-style models show that large-scale pretraining can transfer broadly across language tasks.
Instead of building task-specific systems from scratch, teams pretrain general models and adapt them, setting up the foundation-model era.
Transformers
2019
OpenAI releases a larger transformer language model with surprisingly coherent long-form generation.
GPT-2 turns language modeling from a benchmark exercise into a visible product and safety conversation.
Foundation Models
2020
A 175B parameter language model demonstrates strong few-shot behavior through natural-language prompts.
The interface changes: users can specify tasks in ordinary text, and the model adapts without a bespoke training run for each workflow.
Foundation Models
2020
DeepMind's system reaches breakthrough accuracy on protein folding.
AI progress becomes visibly scientific, not just linguistic or perceptual: learned systems can compress years of biological structure work.
Foundation Models
2021
Text-conditioned image generation becomes a mainstream AI capability.
Generation starts moving from demos to creative tools: prompts become a bridge between language, visual concepts, and controllable outputs.
Foundation Models
2021
OpenAI Codex connects language models to code generation and developer workflows.
The coding assistant category appears: models can draft, translate, complete, and explain code well enough to reshape developer tooling.
Scaling and Chat
2022
DeepMind argues many large language models are undertrained relative to their parameter count.
The frontier shifts from bigger-only thinking to compute-optimal tradeoffs across parameters, tokens, and training budget.
Scaling and Chat
2022
A latent diffusion model becomes widely available for local and community experimentation.
Open image generation changes the culture of AI: model weights, extensions, fine-tunes, and creative workflows spread beyond closed labs.
Scaling and Chat
November 2022
A chat interface turns language-model capability into a daily tool for millions of people.
The important move is product shape: instruction following, memory of a conversation, and low-friction access make LLMs legible to non-specialists.
Multimodal and Agents
March 2023
OpenAI releases a stronger multimodal-capable model with broad performance gains.
GPT-4 becomes a reference point for model capability, safety evaluation, enterprise adoption, and the first wave of serious agent experiments.
Multimodal and Agents
February 2023
Meta releases LLaMA, helping open-weight language-model experimentation spread quickly.
A parallel ecosystem forms around smaller, adaptable models that can be inspected, fine-tuned, and deployed outside a single hosted API.
Multimodal and Agents
March 2023
Anthropic introduces Claude as a constitutional-AI assistant focused on helpfulness, honesty, and harmlessness.
The market becomes multi-lab and safety-positioned: assistant behavior, context length, and reliability become product differentiators.
Multimodal and Agents
December 2023
Google introduces Gemini as a family of models designed around multimodal reasoning.
The field moves from text-only chat toward systems that operate across images, audio, video, code, and tool use.
Multimodal and Agents
February 2024
OpenAI previews a video generation model capable of long, coherent scenes from text prompts.
Video makes the generative leap visceral: temporal consistency, physical plausibility, and creative control become frontier concerns.
Multimodal and Agents
May 2024
OpenAI releases an omni model designed for real-time multimodal interaction.
The assistant interface starts to feel less like a text box and more like a real-time collaborator with speech, image, and tool surfaces.
Reasoning and Agents
September 2024
OpenAI introduces o1, emphasizing deliberate reasoning for math, code, and hard multi-step tasks.
The product language changes from instant completion to thinking time, evaluation, and reliability on tasks that require sustained reasoning.
Reasoning and Agents
January 2025
DeepSeek releases open reasoning models that make frontier-style reasoning feel less exclusive.
R1 intensifies the global model race around efficiency, open weights, reinforcement learning, and how fast strong capabilities can diffuse.
Reasoning and Agents
February 2025
OpenAI launches a research agent that searches, reads, and synthesizes sources into longer reports.
The assistant moves from answering from model memory toward using tools, citations, and multi-step source gathering as part of the work product.
Reasoning and Agents
April 2025
GPT-4.1 focuses on API coding performance while o3 and o4-mini combine reasoning with tool use.
The model menu fragments into useful roles: faster coding models, deeper reasoning models, cheaper small models, and multimodal tool users.
Reasoning and Agents
May 2025
Anthropic introduces Claude Opus 4 and Sonnet 4 with emphasis on coding and extended task performance.
The agentic frontier becomes practical: models are judged by whether they can hold a software task, inspect context, edit, test, and recover.
Reasoning and Agents
2025
OpenAI presents GPT-5 as a broadly available model with built-in reasoning and strong coding performance.
By this point, users expect a single assistant to route between fast answers, deeper reasoning, multimodal inputs, and tool-driven work.
Integrated AI
February 2026
Deep research expands toward authenticated apps, trusted-source constraints, and real-time progress control.
The direction is clear: serious AI work is not only model capability, but permissioned context, connectors, interruption, provenance, and workflow fit.
Integrated AI
2026
Stanford HAI reports rapid capability growth, broad adoption, infrastructure strain, and policy gaps.
AI is no longer one technical story. It is an industrial, educational, scientific, political, and infrastructure system moving at different speeds.
End of the line
Every entry here was somebody’s whole career. The interesting part is not any single one of them — it is how short the gaps have become.