Three artifacts on dark slate: a vacuum tube, a perforated punch-card, and a black silicon chip — the chronology of compute.

AI history atlas, 1950 to 2026

A year-by-year map of the papers, models, and turns that built the present moment

This page is a working atlas. Not a hype reel, not a hall of fame. It tracks the actual events that moved the field from Turing's 1950 question to the agent systems shipping in 2026, and it tries to say honestly why each entry still matters now. Some entries are obvious. Some are quiet papers that turned out to compound for decades. A few are commercial moments that changed funding and attention more than they changed the math. The dates are taken from primary sources where we could find them; the readings underneath each marker are our own, written in plain language. We kept three rules while building it. First, no invented details. If we cite a paper or a benchmark, the URL points at the actual paper or the actual benchmark page. If a fact is uncertain as of June 2026, the prose says so. Second, no boosterism. The two AI winters were real and they were a feature of the field, not a failure of imagination. We mark them as carefully as we mark the breakthroughs. Third, "why it matters now" should be a useful sentence, not a slogan. If a 1969 critique still constrains how you choose a model architecture in 2026, the entry should make that link visible. The shape of the timeline is uneven on purpose. The early decades had wide gaps between events because the field was small and compute was scarce. Since roughly 2017, the pace has compressed dramatically; a paper that would have defined a five-year subfield in 1995 now gets metabolized in months. We tried to keep the entries at a roughly constant grain of importance, which means recent years look denser than they would in a more uniform chronology. That density is the story. If you are coming to this atlas to figure out what to read first, skip to the reading list at the bottom. If you want to understand why current models look the way they do, the cluster from 2017 to 2022 is where the present shape of the field was set.

How to read this atlas

Each entry below has three parts: the year and event, what the work actually was, and why it still matters in 2026. We use 'still matters' in a specific sense — either the technique is still used in production systems, or the conceptual framing it introduced is still how practitioners reason about the problem, or the institutional consequence (funding cycle, public expectation, regulatory posture) is still shaping the field. When none of those apply we say so. Several entries are paired with an AI winter marker; those winters are not embarrassments to skip past, they are the part of the field's history that explains why current researchers are cautious about over-promising. We use the term 'foundation model' in the post-2018 sense (large pretrained model adapted to many tasks), not as a vague honorific. Where benchmark numbers are mentioned we name the benchmark; where dollar amounts appear we mark them illustrative unless the figure is from a primary source. We do not adjudicate priority disputes that the field itself hasn't resolved (for example, the exact ordering of credit for backpropagation); we link to surveys instead. The goal is a map you can actually navigate, not a victory parade.

Timeline · 1950 to 2026

1950
Turing, 'Computing Machinery and Intelligence'
Alan Turing's paper in Mind reframed the question 'can machines think' as an imitation game with a behavioral test. Why it matters now: every public conversation about whether a model is 'really' intelligent is still working inside Turing's behaviorist frame, including the people who reject the frame. Modern evaluation benchmarks descend directly from this move.
1956
Dartmouth Summer Research Project
McCarthy, Minsky, Rochester, and Shannon convene the workshop that names the field 'artificial intelligence' and projects (optimistically) that significant progress could be made in a summer. Why it matters now: the founding optimism set the template for hype cycles the field has repeated roughly every fifteen years since.
1957
Rosenblatt's Perceptron
Frank Rosenblatt builds the Perceptron at Cornell Aeronautical Laboratory — a single-layer trainable classifier with a learning rule. The 1958 paper appears in Psychological Review. Why it matters now: it is the direct ancestor of every neural network in production today. The basic update rule is recognizable in modern stochastic gradient descent.
1969
Minsky and Papert, 'Perceptrons' (first AI winter begins)
The book formalized what single-layer perceptrons could not do (notably XOR) and is widely credited with cooling neural-network research for over a decade. The full story is more nuanced — multilayer networks were known to be more powerful, but training them was unsolved. Why it matters now: it is the canonical example of a correct technical critique having an outsized institutional effect, and a reason later researchers worked hard to publish negative results carefully.
1980s
Expert systems and the second AI winter
Rule-based expert systems (MYCIN, XCON, R1) attracted commercial investment in the early-to-mid 1980s; the Lisp machine market collapsed and DARPA pulled back on Strategic Computing Initiative funding, producing the second AI winter by the late 1980s. Why it matters now: it is the field's clearest case of a paradigm shipping real value in narrow domains while failing to generalize, which is exactly the failure mode current LLM deployments must avoid.
1989
LeCun, MNIST, and convolutional networks
Yann LeCun and collaborators at Bell Labs apply backpropagation to a convolutional network for handwritten digit recognition (the LeNet line of work; the published 'Backpropagation Applied to Handwritten Zip Code Recognition' is from 1989). MNIST itself, the benchmark dataset, is assembled in the 1990s. Why it matters now: convolutions, weight sharing, and translation invariance are still the architectural intuition behind modern vision encoders, and MNIST remains the first dataset most students train on.
2006
Hinton, Osindero, Teh — 'A Fast Learning Algorithm for Deep Belief Nets'
The paper popularized layerwise pretraining of deep networks and is the conventional starting point for the 'deep learning' renaissance, alongside related work by Bengio and others. Why it matters now: it was the moment 'deeper is feasible' became a working assumption, which is the assumption underneath every model trained since.
2012
AlexNet wins ImageNet
Krizhevsky, Sutskever, and Hinton's deep convolutional network wins the ImageNet Large Scale Visual Recognition Challenge by a wide margin, trained on two consumer GPUs. Why it matters now: this is the inflection where 'big neural net + GPU + lots of labeled data' moved from research curiosity to obvious default, and GPU vendors (notably Nvidia) became infrastructure for the field.
2013
Word2Vec (Mikolov et al.)
Efficient learning of dense word embeddings via skip-gram and CBOW objectives. Why it matters now: it normalized the practice of representing discrete symbols as learned vectors, which is the conceptual move that all modern language models inherit. The 'king − man + woman ≈ queen' demo became the field's most-shared intuition pump.
2014
Goodfellow et al., Generative Adversarial Networks
The GAN paper at NeurIPS 2014 introduced a two-network adversarial training setup. Why it matters now: GANs dominated generative image research from about 2015 to 2020 before diffusion models displaced them on most benchmarks. The adversarial-training pattern still shows up in modern alignment and red-team setups.
2016
AlphaGo defeats Lee Sedol
DeepMind's AlphaGo, built on Monte Carlo Tree Search plus deep networks, wins 4-1 against Lee Sedol in Seoul. The 'Mastering the game of Go with deep neural networks and tree search' paper appeared in Nature in January 2016. Why it matters now: it shifted public and policymaker perception of what reinforcement learning plus deep nets could do, and the 'self-play + search' template is being revisited in 2025-2026 reasoning models.
2017
'Attention Is All You Need' — the Transformer
Vaswani et al. at Google introduce the Transformer architecture, replacing recurrence and convolutions with multi-head self-attention. Published at NeurIPS 2017. Why it matters now: every large language model in production in 2026 is a Transformer variant. This is the single paper most responsible for the current shape of the field.
2018
GPT-1 and BERT
OpenAI publishes 'Improving Language Understanding by Generative Pre-Training' (GPT-1); Google publishes BERT shortly after. Why it matters now: the pretrain-then-finetune recipe became the default for language tasks. GPT-1 specifically established the decoder-only autoregressive recipe that scales into GPT-2, 3, 4, and beyond.
2019
GPT-2 and the staged-release debate
OpenAI announces GPT-2 in February 2019 and initially withholds the full 1.5B-parameter model citing misuse concerns, releasing it in stages through November 2019. Why it matters now: this was the field's first public attempt at staged release for safety reasons, and the template (and the criticism it attracted) shaped every subsequent capability-release conversation.
2020
GPT-3 — 'Language Models are Few-Shot Learners'
Brown et al. (OpenAI) publish the GPT-3 paper at NeurIPS 2020. The 175B-parameter model demonstrates that scale plus a prompt is often a substitute for task-specific training. Why it matters now: it established in-context learning and few-shot prompting as primary interfaces to large models, and it is the model whose API access reshaped how startups built with AI.
2021
DALL·E and CLIP
OpenAI releases DALL·E (text-to-image transformer) and CLIP (contrastive image-text pretraining) in January 2021. Why it matters now: CLIP's joint image-text embedding became the workhorse behind a large fraction of generative-image systems (including diffusion models that came later), and DALL·E demonstrated that text-conditioned generation worked at consumer-visible quality.
2022 · January
Chain-of-Thought prompting (Wei et al.)
Google Brain researchers show that prompting large models with intermediate reasoning steps improves performance on math and commonsense benchmarks. Why it matters now: it is the conceptual ancestor of the 2024-2025 reasoning models — the idea that exposing a model's intermediate work is itself a capability lever.
2022 · April
DALL·E 2, Imagen, and the text-to-image year
Diffusion-based text-to-image systems (DALL·E 2 in April, Imagen in May, Stable Diffusion's open release in August) reach quality and accessibility that triggers the first broad public encounter with generative AI. Why it matters now: this is when the artist and copyright conversations that still shape model training policy in 2026 actually started.
2022 · November
ChatGPT public launch
OpenAI launches ChatGPT on 30 November 2022, built on a fine-tuned GPT-3.5. Reaches approximately 100 million users in the following months, becoming one of the fastest-growing consumer products in history. Why it matters now: this is the event that moved AI from research demo to default consumer expectation, and triggered the funding and product wave that defines 2023 through 2026.
2023 · February
Meta releases LLaMA
Meta AI releases LLaMA (7B to 65B parameters) to researchers; weights leak publicly within a week. Why it matters now: it is the start of the modern open-weights ecosystem. Almost every open or semi-open large model since traces lineage or comparison to a LLaMA-family release.
2023 · March
GPT-4 launches
OpenAI releases GPT-4 in March 2023. The technical report explicitly does not disclose architecture, parameter count, or training data. Why it matters now: GPT-4 set the capability bar that every subsequent frontier model has been measured against on standard benchmarks (MMLU, GSM8K, HumanEval, etc.), and the closed technical report became the template for frontier-lab disclosure norms.
2024
Sora, Claude 3.5 Sonnet, and the multimodal year
OpenAI demos Sora (text-to-video diffusion) in February 2024. Anthropic releases Claude 3.5 Sonnet in June 2024, which becomes a widely-used frontier model for coding tasks. Google ships Gemini 1.5 with long-context capabilities. Why it matters now: 2024 was when 'multimodal' stopped being a research direction and became a product expectation.
2024 · July
Llama 3.1 405B
Meta releases Llama 3.1 in July 2024, including a 405B-parameter open-weights model. Why it matters now: it was the first open-weights release that genuinely competed with leading closed frontier models on standard benchmarks, and it set the reference point for the open-vs-closed debate through 2025-2026.
2024 · September
OpenAI o1 and the reasoning-model shift
OpenAI releases o1-preview in September 2024, a model trained to spend significant inference-time compute on chain-of-thought reasoning. o3 is announced in December 2024 with strong scores on ARC-AGI and frontier math benchmarks. Why it matters now: this is the move from 'scale pretraining' to 'scale inference-time compute' as the next axis of capability gain, and it changed the cost structure of frontier-model use.
2024 · October
Claude 3.5 Sonnet (new) and computer use
Anthropic releases an updated Claude 3.5 Sonnet in October 2024 with a research preview of computer-use capability — the model can take screenshots and control a virtual mouse and keyboard. Why it matters now: it is the first widely-available frontier model that ships an agentic interface to general computer control, and the design choices in that release shape the agent ecosystem in 2025-2026.
2025 to 2026
The agent year
Across 2025, frontier labs and startups ship agentic products: autonomous coding agents (Devin, Claude Code, GitHub Copilot Workspace), research and computer-use agents, and long-horizon task systems. As of June 2026 the agent category is still consolidating; reliability on multi-step tasks is the central open problem and benchmark numbers vary widely by methodology. Why it matters now: the field is currently betting that agents are the product form that converts model capability into economic value, and the next twelve to twenty-four months will determine whether that bet pays off.

The two AI winters in plain language

An AI winter is a period of reduced funding and reduced public attention that follows a cycle of inflated promises. The field has had two, and parts of three. The first winter (roughly 1974-1980, deepening after the 1969 Minsky-Papert critique and the 1973 Lighthill report in the UK) followed early symbolic AI's failure to scale to general reasoning. The second winter (roughly 1987-1993) followed the collapse of the commercial expert-systems market and the Lisp machine industry. There was also a less-famous 'connectionist winter' in the mid-1990s when neural networks lost ground to statistical methods like support vector machines, before recovering after 2006. The point of marking these is not to predict another winter — it is to be honest that the field's optimism curve has historically run ahead of its capability curve, and that current expectations should be calibrated against that history rather than against the most recent press cycle.

Papers that compounded

Year	Paper	Why it compounded
1950	Turing · Computing Machinery and Intelligence	Set the behavioral evaluation frame the field still argues inside
1957	Rosenblatt · The Perceptron	Established the trainable-classifier template for neural nets
1986	Rumelhart, Hinton, Williams · Learning representations by back-propagating errors	Made multilayer training tractable and standard
1989	LeCun et al. · Backprop applied to handwritten zip code recognition	Convolutions plus weight sharing, still the vision-encoder pattern
1997	Hochreiter and Schmidhuber · Long Short-Term Memory	Made sequence modeling work, defined the pre-Transformer era
2013	Mikolov et al. · Word2Vec	Normalized learned vector representations of discrete symbols
2014	Goodfellow et al. · Generative Adversarial Nets	Adversarial training as a generative-modeling and safety pattern
2015	He et al. · Deep Residual Learning (ResNet)	Residual connections, now ubiquitous in deep architectures
2017	Vaswani et al. · Attention Is All You Need	Transformer; the architecture underneath every current LLM
2020	Brown et al. · Language Models are Few-Shot Learners (GPT-3)	Established in-context learning as a primary interface
2022	Wei et al. · Chain-of-Thought Prompting Elicits Reasoning	Conceptual ancestor of o1-style reasoning models
2022	Ouyang et al. · Training language models to follow instructions with human feedback (InstructGPT)	RLHF as the standard alignment recipe for deployed LLMs

Year1950

PaperTuring · Computing Machinery and Intelligence

Why it compoundedSet the behavioral evaluation frame the field still argues inside

Year1957

PaperRosenblatt · The Perceptron

Why it compoundedEstablished the trainable-classifier template for neural nets

Year1986

PaperRumelhart, Hinton, Williams · Learning representations by back-propagating errors

Why it compoundedMade multilayer training tractable and standard

Year1989

PaperLeCun et al. · Backprop applied to handwritten zip code recognition

Why it compoundedConvolutions plus weight sharing, still the vision-encoder pattern

Year1997

PaperHochreiter and Schmidhuber · Long Short-Term Memory

Why it compoundedMade sequence modeling work, defined the pre-Transformer era

Year2013

PaperMikolov et al. · Word2Vec

Why it compoundedNormalized learned vector representations of discrete symbols

Year2014

PaperGoodfellow et al. · Generative Adversarial Nets

Why it compoundedAdversarial training as a generative-modeling and safety pattern

Year2015

PaperHe et al. · Deep Residual Learning (ResNet)

Why it compoundedResidual connections, now ubiquitous in deep architectures

Year2017

PaperVaswani et al. · Attention Is All You Need

Why it compoundedTransformer; the architecture underneath every current LLM

Year2020

PaperBrown et al. · Language Models are Few-Shot Learners (GPT-3)

Why it compoundedEstablished in-context learning as a primary interface

Year2022

PaperWei et al. · Chain-of-Thought Prompting Elicits Reasoning

Why it compoundedConceptual ancestor of o1-style reasoning models

Year2022

PaperOuyang et al. · Training language models to follow instructions with human feedback (InstructGPT)

Why it compoundedRLHF as the standard alignment recipe for deployed LLMs

Reading list, ordered for someone starting now

Start with Turing's 1950 paper. It is short, written for a general audience, and still the cleanest framing of the central question.
Read Rosenblatt's 1958 Perceptron paper alongside a modern explainer; the math is recognizable as gradient descent on a linear model.
Skip ahead to Vaswani et al. 2017, 'Attention Is All You Need'. Read it twice. Almost everything that follows in the LLM era is a variation on this architecture.
Read the GPT-3 paper (Brown et al. 2020) for the empirical case that scale alone changes the interface to the model.
Read the Chain-of-Thought paper (Wei et al. 2022) for the conceptual move from 'better outputs' to 'visible reasoning'.
Read the InstructGPT / RLHF paper (Ouyang et al. 2022) for how deployed models are made to follow instructions.
Then read one survey or one blog post on the Anthropic, OpenAI, DeepMind, and Meta AI research pages from the last six months — the frontier moves fast enough that papers older than a year are background, not current state.

What 'why it matters now' actually means

Still in production

Live techniques

Convolutions, residual connections, attention, embedding lookups, dropout, layer normalization, AdamW, RLHF, KV caching. These are not historical; they are in the inference path of the model serving you right now.

Still the framing

Live framings

Behavioral evaluation (Turing), in-context learning (GPT-3), chain-of-thought as a capability axis (Wei et al.), pretrain-then-adapt (BERT/GPT). Even when the specific implementation changes, the conceptual frame is how practitioners reason about the problem.

Still the constraint

Live constraints

The two AI winters set institutional caution. GPT-2's staged release set disclosure norms. The 2022-2023 image-generation copyright fights set training-data policy. These are not technical facts but they bound what is shippable.

Historical but read it anyway

Cautionary reading

Symbolic AI, expert systems, classical knowledge representation. Not in production for general tasks, but the failure modes are instructive — they are exactly the failure modes current systems must avoid in narrow vertical deployments.

What is unsettled as of June 2026

Three things are genuinely open and we will not pretend otherwise. First, whether inference-time compute (reasoning models, search, agent loops) will keep producing capability gains at the rate seen in late 2024 and 2025, or whether it will plateau the way pure pretraining-scale gains appear to have plateaued. As of mid-2026 best-effort, the public benchmark trajectory suggests continued progress but not at the breakneck rate of the o1-to-o3 jump. Second, whether the open-weights ecosystem will continue to close the gap with closed frontier labs or whether the gap will widen as training-compute requirements grow; the Llama and Mistral release cadence through 2025 supported the closing-gap case, but the most capable closed models still lead on the hardest benchmarks. Third, whether agent reliability on long-horizon tasks will reach the threshold where most knowledge work can be meaningfully delegated; current public benchmarks (SWE-bench Verified, GAIA, OSWorld, and others) show real progress but also persistent variance, and methodology differences make cross-lab comparisons harder than press coverage suggests. We do not know the answers, and we are skeptical of anyone who claims certainty about them right now. Check provider docs and current benchmark leaderboards for the live state of any specific number.

Sources

[01]
Turing's 1950 paper 'Computing Machinery and Intelligence' appeared in Mind, volume LIX, issue 236
academic.oup.com/mind/article/LIX/236/433/986238
[02]
The original 1956 Dartmouth Summer Research Project on Artificial Intelligence proposal naming the field
raysolomonoff.com/dartmouth/boxa/dart564props.pdf
[03]
Rosenblatt's 1958 paper 'The Perceptron: A Probabilistic Model' in Psychological Review
psycnet.apa.org/record/1959-09865-001
[04]
Minsky and Papert's 'Perceptrons' (MIT Press, 1969) and the subsequent expanded edition
mitpress.mit.edu/9780262630221/perceptrons
[05]
LeCun et al. 1989 'Backpropagation Applied to Handwritten Zip Code Recognition' in Neural Computation
yann.lecun.com/exdb/publis/pdf/lecun-89e.pdf
[06]
Hinton, Osindero, Teh 2006 'A Fast Learning Algorithm for Deep Belief Nets' in Neural Computation
cs.toronto.edu/~hinton/absps/fastnc.pdf
[07]
Krizhevsky, Sutskever, Hinton 2012 'ImageNet Classification with Deep Convolutional Neural Networks' (AlexNet) at NeurIPS
papers.nips.cc/paper_files/paper/2012/hash/c399862d3b9d6b76c8436e924a68c45b-Abstract.html
[08]
Mikolov et al. 2013 'Efficient Estimation of Word Representations in Vector Space' (Word2Vec)
arxiv.org/abs/1301.3781
[09]
Goodfellow et al. 2014 'Generative Adversarial Networks'
arxiv.org/abs/1406.2661
[10]
Silver et al. 2016 'Mastering the game of Go with deep neural networks and tree search' in Nature
nature.com/articles/nature16961
[11]
Vaswani et al. 2017 'Attention Is All You Need' introducing the Transformer architecture
arxiv.org/abs/1706.03762
[12]
Radford et al. 2018 'Improving Language Understanding by Generative Pre-Training' (GPT-1)
cdn.openai.com/research-covers/language-unsupervised/language_understanding_paper.pdf
[13]
Devlin et al. 2018 'BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding'
arxiv.org/abs/1810.04805
[14]
OpenAI's February 2019 GPT-2 announcement and initial staged-release policy
openai.com/index/better-language-models/
[15]
Brown et al. 2020 'Language Models are Few-Shot Learners' (GPT-3)
arxiv.org/abs/2005.14165
[16]
Radford et al. 2021 'Learning Transferable Visual Models From Natural Language Supervision' (CLIP)
arxiv.org/abs/2103.00020
[17]
OpenAI's January 2021 DALL·E announcement
openai.com/index/dall-e/
[18]
Wei et al. 2022 'Chain-of-Thought Prompting Elicits Reasoning in Large Language Models'
arxiv.org/abs/2201.11903
[19]
Ouyang et al. 2022 'Training language models to follow instructions with human feedback' (InstructGPT, RLHF)
arxiv.org/abs/2203.02155
[20]
OpenAI's 30 November 2022 ChatGPT launch announcement
openai.com/index/chatgpt/
[21]
Meta AI's February 2023 LLaMA release announcement
ai.meta.com/blog/large-language-model-llama-meta-ai/
[22]
OpenAI 2023 'GPT-4 Technical Report'
arxiv.org/abs/2303.08774
[23]
Meta's July 2024 Llama 3.1 release including the 405B-parameter open-weights model
ai.meta.com/blog/meta-llama-3-1/
[24]
OpenAI's September 2024 o1 announcement describing inference-time reasoning
openai.com/index/learning-to-reason-with-llms/
[25]
Anthropic's October 2024 Claude 3.5 Sonnet update and computer-use research preview
anthropic.com/news/3-5-models-and-computer-use

Keep reading

Learn · playbooks →Research papers →Atlas index →Tools for builders →OrangeBox →B00KMakor →Model comparisons →Founder's view →

AI history atlas, 1950 to 2026

How to read this atlas

Timeline · 1950 to 2026

Turing, 'Computing Machinery and Intelligence'

Dartmouth Summer Research Project

Rosenblatt's Perceptron

Minsky and Papert, 'Perceptrons' (first AI winter begins)

Expert systems and the second AI winter

LeCun, MNIST, and convolutional networks

Hinton, Osindero, Teh — 'A Fast Learning Algorithm for Deep Belief Nets'

AlexNet wins ImageNet

Word2Vec (Mikolov et al.)

Goodfellow et al., Generative Adversarial Networks

AlphaGo defeats Lee Sedol

'Attention Is All You Need' — the Transformer

GPT-1 and BERT

GPT-2 and the staged-release debate

GPT-3 — 'Language Models are Few-Shot Learners'

DALL·E and CLIP

Chain-of-Thought prompting (Wei et al.)

DALL·E 2, Imagen, and the text-to-image year

ChatGPT public launch

Meta releases LLaMA

GPT-4 launches

Sora, Claude 3.5 Sonnet, and the multimodal year

Llama 3.1 405B

OpenAI o1 and the reasoning-model shift

Claude 3.5 Sonnet (new) and computer use

The agent year