Timeline of Intelligence
From Turing’s question in 1936 to DeepSeek shocking the compute narrative in 2025, seven stages thread together the rise and fall of near ninety years of AI development. Tap a figure’s badge to jump to its card.
Ⅰ · "Can Machines Think?" — The Birth of AI
The core of this stage is not any particular technical breakthrough but the first serious posing of a question never before taken seriously: can a machine think? In 1950 Turing proposed an operational test that sidestepped the philosophical dispute; at the 1956 Dartmouth conference McCarthy formally named this new discipline; and Newell and Simon, with a program able to prove mathematical theorems automatically, showed that "symbol manipulation" might just be the essence of intelligence. Meanwhile the science-fiction writer Asimov had, as early as 1942, in a fictional world begun thinking ahead of the real researchers about the core question, lasting to this day, of "what rules a machine ought to obey."
Turing’s Question and the Dartmouth Conference
In 1950 Turing published a paper proposing the "imitation game" (the Turing test), giving the vague philosophical question of "whether a machine can think" an operational alternative. In 1956 McCarthy hosted at Dartmouth College a two-month workshop that formally proposed the term "artificial intelligence," and Minsky, Shannon, Newell, and Simon were all participants in this meeting thereafter acknowledged as the founding mark of the discipline.
The Naming of "Machine Learning" and the Prelude of Cybernetics
Samuel proposed the term "machine learning" in 1959, and the checkers program he wrote, improving its play by repeatedly playing against itself, is the earliest practical prototype of this idea; almost at the same time, Wiener’s cybernetics had already systematically expounded how the "feedback" mechanism explains at once the self-regulating behavior of organisms and of machines, and was the first to warn of the social impact automation might bring.
Science Fiction Shines Into Reality: The "Three Laws of Robotics"
As early as a 1942 short story Asimov proposed the "Three Laws of Robotics," a fictional ethical framework designed for literary creation that thereafter continued to inspire real AI-safety and ethics researchers to think about how behavioral constraints ought to be set for intelligent systems, a question to which there is still no settled answer.
Ⅱ · The Perceptron, the Winter, and the Revival of Connectionism
The perceptron Rosenblatt proposed in 1958 for the first time let a machine show the ability to learn autonomously from data, but the media’s overly optimistic reporting laid the ground for Minsky’s later mathematical critique — the book Perceptrons directly plunged neural-network research into the "first AI winter" of over a decade. In this trough backpropagation was independently discovered more than once, statistical learning theory and Bayesian networks gave the question of "how to learn reliably from data" a more rigorous mathematical foundation, and behavior-based robotics also quietly rose at this time.
The Hope and the Disillusion of the Perceptron
The perceptron Rosenblatt proposed in 1958 was the first neural-network model in history able to learn classification rules autonomously, and the media’s overly optimistic reporting drew academic criticism. In 1969 the book Perceptrons co-authored by Minsky and Papert used mathematical proof to point out the structural limits of the single-layer perceptron, a conclusion thereafter widely thought to have directly caused the great shrinkage of neural-network research funding, known in history as the first "AI winter."
The Rediscovery of the Backpropagation Algorithm
Werbos’s 1974 doctoral thesis had already proposed a method highly similar to the principle of backpropagation, but it drew little attention; in 1986 the paper co-authored by Rumelhart, Hinton, and Williams systematically showed that this method could let multi-layer neural networks learn useful internal representations, and only from then did backpropagation truly win wide academic application; the Boltzmann machine that Hinton and Sejnowski had proposed earlier also gave the connectionist research of this period an important theoretical foundation.
Statistical Learning Theory and Bayesian Networks
The statistical learning theory and support vector machine Vapnik proposed gave the generalization ability of machine learning a rigorous mathematical explanation, and for over a decade before the rise of deep learning were the mainstream method for classification tasks; the Bayesian networks Pearl proposed gave AI systems an equally far-reaching graphical-model tool for reasoning under uncertainty.
Behavior-Based Robotics and Robots That "Read the Room"
In 1986 Brooks proposed the "subsumption architecture," arguing that a robot need not first build a complete symbolic understanding of the world and could, by simple reactive behavioral layers alone, show complex adaptive behavior, an idea that thereafter gave rise to iRobot’s Roomba; his student Breazeal pioneered social robotics, designing the Kismet robot able to interact emotionally with humans through facial expression.
Ⅲ · Human-Machine Play and the Practical Turn of Machine Learning
At this stage AI moved from a relatively niche academic topic gradually into public view and industrial practice. Deep Blue’s 1997 victory over chess world champion Kasparov was the first true public milestone in the history of AI; meanwhile the education-outreach movement driven by Google Brain and by figures such as Andrew Ng brought deep-learning research systematically into the view of large tech companies and learners worldwide; the LSTM architecture Schmidhuber proposed gave sequence-data modeling a mainstream tool for over two decades; and the DeepMind founded by Demis Hassabis and others also quietly planted at this time the seed of the AlphaGo that would thereafter burst upon the scene.
Deep Blue Beats Kasparov
In 1997 IBM’s chess computer "Deep Blue" played a six-game match against world champion Kasparov, at last beating him by two wins, one loss, and three draws, the first time in history a top human player lost to a computer in a formal match, thereafter widely seen as the first true public milestone in the history of AI.
The Founding of Google Brain and the Education-Outreach Movement
In 2011 Dean co-founded the Google Brain research team, bringing deep-learning research systematically into one of the world’s largest tech companies; almost at the same period Andrew Ng put his Stanford machine-learning course online, and Howard founded fast.ai, the two together opening the education-outreach movement that brought deep learning out of the elite lab and onto the computer screens of learners worldwide.
The Proposal of the Long Short-Term Memory Network (LSTM)
The LSTM architecture Schmidhuber and Hochreiter proposed in 1997 solved, by introducing a gating mechanism, the vanishing-gradient problem of traditional recurrent neural networks in handling long-sequence data, and for over two decades thereafter was the mainstream tool for sequence-modeling tasks such as speech recognition and machine translation, only gradually yielding its mainstream place after the rise of the Transformer architecture.
The Birth of DeepMind
In 2010 Hassabis with Legg and Suleyman together founded DeepMind, the company’s core goal summed up as "first solve intelligence, then use intelligence to solve everything else," acquired by Google in 2014 and thereafter gradually growing into one of the world’s most important AI research institutions.
Ⅳ · The Deep-Learning Revival and the Shock of ImageNet
This is one of the most critical turning points in the whole history of AI. Hinton held on through the decades of "winter" in which neural-network research was widely written off by the mainstream, and in 2012 his students proved with AlexNet, on the ImageNet competition Fei-Fei Li created, the overwhelming advantage of deep learning, thereafter triggering the collective turn of the whole field toward deep learning. The generative adversarial network Goodfellow proposed opened the earliest technical path for generative AI; meanwhile China’s AI industry too quietly set out, driven by figures such as Robin Li and Kai-Fu Lee.
Persistence Through the "Neural-Network Winter" and the Shock of AlexNet
Hinton held to deep work on neural networks through the decades of "winter" in which the AI field widely turned to other technical paths, and his 2006 deep-belief-network research is seen as the starting point of the deep-learning revival; in 2012 the AlexNet designed by his two graduate students Sutskever and Krizhevsky won the ImageNet competition Fei-Fei Li created by a margin far above its rivals, an overwhelming victory that triggered the collective turn of the whole AI field toward deep learning, and Hinton, LeCun, and Bengio thereafter shared the 2018 Turing Award.
The Late-Night Inspiration of the GAN
After a bar gathering Goodfellow conceived late one night the "generative adversarial network" of two networks contesting each other, letting a generator and a discriminator advance together in continual contest until able to produce image content indistinguishable from real, a conception thereafter acknowledged as one of the earliest core breakthroughs in the field of generative AI.
The Early Drivers of China’s AI Industry
The Baidu that Robin Li led invested systematically in AI research relatively early, hiring Andrew Ng as chief scientist in 2014; Kai-Fu Lee, after years shuttling between Silicon Valley and China, founded Sinovation Ventures to keep backing homegrown AI startups, and the two are widely seen as the most important early drivers of China’s AI industry.
Ⅴ · AlphaGo, the Transformer, and the Race to Scale
This stage witnessed two breakthroughs that thereafter deeply reshaped the landscape of the AI industry: DeepMind’s AlphaGo cracked Go, a field seen as the last stronghold of human intuition, and the Transformer architecture the Google team proposed thereafter became the common underlying bedrock of almost every large language model. Meanwhile OpenAI was founded and went through its founders’ falling-out at this time, AI-safety thought began to enter public view, and the dispute over the technical path of self-driving quietly opened too.
AlphaGo Beats Lee Sedol
In 2016 the AlphaGo Silver led the design of beat top Korean player Lee Sedol four to one, proving the combination of deep learning and reinforcement learning enough to crack Go, a field previously seen as the last stronghold of human intuition; Silver’s reinforcement-learning theory was deeply shaped by his doctoral adviser Sutton, and this victory was also a powerful empirical vindication of Sutton’s own years of theoretical research.
"Attention Is All You Need" and the Twin Advance of BERT and GPT
In 2017 eight researchers of the Google Brain team published "Attention Is All You Need," proposing the Transformer architecture based wholly on the self-attention mechanism; thereafter Devlin at Google led the design of the BERT model, proving the power of bidirectional pre-training, and Radford at OpenAI designed along the decoder line the earliest version of the GPT series, the two technical paths thereafter together shaping the landscape of large language models.
The Founding of OpenAI and Musk’s Exit
In 2015 Altman with Musk and others together founded OpenAI, at first as a non-profit; in 2018 Musk left the board over disagreement with management on development path and control, and thereafter the two sides’ relation gradually turned from cooperation to open commercial and philosophical dispute, while Brockman and Schulman always stayed in the company, continuing to drive the engineering of the GPT series.
The Rise of AI-Safety Thought
The book Superintelligence Bostrom published in 2014 provoked wide discussion among the upper ranks of the tech industry, Russell thereafter proposed the "human-compatible" theoretical framework, and the Machine Intelligence Research Institute and the LessWrong community Yudkowsky founded had from a still earlier time kept studying and discussing the existential risk superintelligent AI might bring, these voices together forming the intellectual soil for the AI-safety topic to thereafter enter mainstream view.
The Self-Driving Vision Dispute
In 2016 Hotz briefly headed Tesla’s self-driving vision team, leaving after only a few months over a disagreement with Musk on technical path, turning to continue exploring open-source self-driving through his own comma.ai; Karpathy, who succeeded to his post, led the development of Tesla’s pure-vision self-driving system, and this path choice of discarding lidar sensors was long disputed within the industry.
Ⅵ · Algorithmic Bias and the Awakening of AI Ethics
As AI systems were deployed at scale into the real world, the bias and misuse risks they might bring gradually surfaced. Buolamwini’s "Gender Shades" research revealed the marked accuracy gap of mainstream face-recognition systems across different groups, and the "stochastic parrot" concept Gebru and her co-authors proposed warned that large language models might amplify the social bias in their training data. Meanwhile the early alignment-research method Christiano proposed and the open-source large-model community Leahy founded also began to face the core tension, running through the whole industry thereafter, of "how ever more powerful technical capability ought to be responsibly treated."
"Gender Shades" and "Stochastic Parrots"
Buolamwini launched at MIT the "Gender Shades" research project, systematically testing and revealing the markedly higher recognition error rate of mainstream commercial face-recognition systems for darker-skinned women, prompting several tech companies to improve the products concerned; Gebru and her co-authors published a paper in 2020 proposing the metaphor of the "stochastic parrot," warning that large language models might amplify the social bias in training data, and the publication process of this paper also led to her own departure from Google, provoking wide industry discussion.
The Early Research of RLHF
The early method Christiano proposed in 2017 of training a model from human preference ranking was thereafter further developed by Schulman and others into the RLHF training paradigm widely applied in products such as ChatGPT, becoming one of the key theoretical sources behind large language models’ ability to "understand human speech."
The Rise and the Dilemma of the Open-Source Large-Model Community
In 2020 Leahy co-founded EleutherAI, opening to the public the large-language-model training technology previously held by only a few tech giants, but as model capability kept leaping he himself thereafter gradually turned to advocating stricter regulation of frontier model development, a shift of stance that also reflects the core tension, running through the whole industry thereafter, between open-source ideal and safety caution.
Ⅶ · The Generative-AI Explosion and the Industry Contest
This is the segment of the history this site covers that is closest to today and still unfolding: ChatGPT pushed large-language-model technology into the mass market for the first time in a product form anyone could use, and OpenAI at once went through a board crisis that shook the whole industry; Amodei left to found Anthropic, Huang’s decade-long GPU bet at last paid off, and Liang Wenfeng’s DeepSeek gave this compute narrative an unexpected shock; Zuckerberg chose open source, Mostaque made a top image model open directly, and the great tech giants completed their internal consolidations one after another in this race. This history is not yet finished, and this site will keep updating as it unfolds.
ChatGPT Past a Hundred Million Users in Two Months, a CEO Fired and Reinstated in Five Days
The OpenAI Altman led released ChatGPT in November 2022, past a hundred million users within two months; in November 2023 board members including Sutskever for a time suddenly removed Altman as CEO, Murati briefly served as interim CEO, Brockman resigned at once in protest, and five days later, under an employee petition, Altman was reinstated, a governance crisis that is a classic case for examining the internal path disputes of this wave of the AI industry.
The Split of Anthropic and "Constitutional AI"
In 2021 Amodei with several other researchers likewise out of OpenAI together founded Anthropic, arguing that the safety of large language models ought to be treated with greater caution; in 2024 Leike, after resigning as head of OpenAI’s "superalignment" team, at once joined Anthropic, a flow of talent widely seen as an important case for examining the internal industry dispute between safety and commercialization paths.
Nvidia’s Decade-Long Bet and the Shock of DeepSeek
The CUDA platform Huang led Nvidia to bet on a decade earlier at last met explosive return after the rise of deep learning, Nvidia thereby becoming the most core supplier of AI compute infrastructure behind this AI wave; in early 2025 the DeepSeek Liang Wenfeng founded released a high-performance open-source model at a training cost far below industry expectation, for a time provoking a market rethink of the assumption that "training a top large model must rely on piling up vast compute."
Open-Source Large Models and the Surge of Generative Images
Zuckerberg led Meta to release the Llama series of large models in relatively open weight form, in sharp contrast to the closed-source commercial path of its main rivals; the Stability AI Mostaque founded made the Stable Diffusion image-generation model open directly at almost the same period, the two decisions together greatly lowering the threshold for developers worldwide to reach frontier generative-AI technology.
The Consolidation and the Race of the Tech Giants
Pichai led the formal merger of Google Brain and DeepMind into a unified "Google DeepMind," the several rounds of strategic investment Nadella-led Microsoft made in OpenAI deeply reshaped the company’s product lines, Suleyman turned to become CEO of Microsoft AI after his Inflection AI core team was "acqui-hired" by Microsoft, and Hoffman, with his long-active venture-capital network, kept driving the capital positioning of this race.
The Voices of Critique
Marcus long publicly questioned that the capability boundary of deep learning and large language models was over-hyped by the industry, waging a public debate lasting years with front-line researchers such as LeCun; the long-form interview podcast Fridman hosts gave the public an important window into the views of the various sides behind this industry race.