AI is not a product of the last five years.
It is the accumulated product of eight decades of mathematics, military funding, corporate ambition, academic inquiry, and hardware revolutions — most of it invisible to the public until the moment it wasn’t.
Overview
When presenters say “AI has been around for a long time,” they are correct — but the statement barely scratches the surface. The field of artificial intelligence as a formal discipline is almost 70 years old, and its intellectual roots run back to the 1940s. What has changed is not the idea of AI, but the scale of compute, the availability of data, and — crucially — who can access it. For most of its history, AI was the exclusive property of defence agencies, elite research universities, and large corporations. The democratisation of AI — making it available to anyone with a browser — is the genuinely new development, and it dates to November 2022.
This article traces that full arc: the foundations, the booms, the busts, who held access at each stage, and how the current frontier model landscape maps globally.
Part 1: The Foundations (1943–1955)
The First Artificial Neuron
The story of AI does not begin with a computer or a lab; it begins with a mathematical idea. In 1943, Warren McCulloch and Walter Pitts published a paper modelling how biological neurons work using formal logic. Their artificial neuron — a simple binary unit that fires when its inputs exceed a threshold — is still recognisable in the architecture of modern neural networks. This was a theoretical contribution with no practical machine to run it, but it planted the conceptual seed.[^1][^2]
Who had access? Nobody in any applied sense. This was pure academic mathematics, circulated among a small number of logicians, mathematicians, and neurologists.
Alan Turing and the Question of Machine Intelligence
In 1950, British mathematician Alan Turing asked a deceptively simple question in his paper “Computing Machinery and Intelligence”: Can machines think?. Rather than debating philosophy, he proposed a practical test: if a human interrogator cannot reliably distinguish between a machine’s responses and a human’s responses in written conversation, the machine should be considered intelligent. This became known as the Turing Test (formally, the Imitation Game). Turing’s framing was radical — it shifted the question from “what is thinking?” to “what does thinking look like from the outside?”[^3][^4][^5]
Turing was writing at a time when computers were enormous, rare machines — used by universities for calculation, by the military for code-breaking, and by governments for infrastructure projects. His 1950 paper was a thought experiment, not a product roadmap.[^6]
Part 2: The Golden Era (1956–1973)
The Dartmouth Conference — AI Gets a Name
The formal birth of AI as an academic discipline occurred in the summer of 1956. John McCarthy, then at Dartmouth College, organised an eight-week workshop bringing together key researchers in mathematics, cognitive science, and engineering. McCarthy chose the name “artificial intelligence” for the project — the first recorded use of the term. The Dartmouth conference was funded in part by the U.S. Air Force and Navy, a pattern of military sponsorship that would define the field for the next two decades.[^7][^8][^9]
The founding attendees included McCarthy, Marvin Minsky, Claude Shannon, and Nathaniel Rochester — names that would define the field for generations. Their shared premise: “every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it”.[^8][^5]
DARPA and the Military’s Central Role
From the early 1960s, the U.S. Department of Defense — primarily through its Advanced Research Projects Agency (ARPA, later renamed DARPA) — became the dominant funder of AI research. This was not an accident; it was Cold War strategy. The Soviet launch of Sputnik in 1957 had shocked Washington, and DARPA was created in 1958 specifically to ensure technological superiority.[^7][^10]
DARPA’s Information Processing Techniques Office, run by J.C.R. Licklider, funded projects across the AI spectrum: natural language processing, chess-playing programs, robotics, and theorem proving. Concretely, DARPA invested $2.2 million in MIT’s AI lab and funded the Stanford University AI project established by McCarthy in 1963, as well as Carnegie Mellon University and SRI International.[^11][^12]
Who had access to AI in this period?
- U.S. military and defence agencies — primary funders and operators
- Elite research universities — MIT, Stanford, Carnegie Mellon, Edinburgh
- Government research labs — SRI International, Lawrence Livermore
- No commercial or public access whatsoever
This exclusivity mattered. The breakthroughs of the 1960s — early natural language programs, the Logic Theorist, chess-playing systems — were academic achievements, not products. The public had no interaction with any of it.
Early Milestones: ELIZA, Shakey, and the First Programs
Several landmark systems emerged in this “Golden Era” that are worth noting as evidence of how early practical AI experiments began:
| System | Year | Developer | What It Did |
|---|---|---|---|
| Logic Theorist | 1956 | Carnegie Mellon (Newell, Simon) | Proved 38 of 52 theorems from Principia Mathematica |
| LISP Programming Language | 1958 | John McCarthy | First dedicated AI programming language, still in use[^9] |
| Perceptron (Mark I) | 1958 | Frank Rosenblatt, Cornell | First machine that learned from data[^2][^13] |
| ELIZA | 1964–1966 | MIT (Weizenbaum) | First chatbot; simulated a psychotherapist |
| DENDRAL | 1965 | Stanford | Identified chemical molecules — first expert system[^14] |
| Shakey the Robot | 1966–1972 | SRI International | First mobile robot that could perceive, reason, plan, and communicate in basic English[^15] |
The optimism of this era was enormous — and dangerously unrealistic. In 1970, Marvin Minsky predicted publicly that “within three to eight years we will have a machine with the general intelligence of an average human”. These predictions would trigger the field’s first catastrophic collapse.[^16]
Part 3: The First AI Winter (1974–1980)
The Lighthill Report and the Funding Drought
By the early 1970s, the promises had not been kept. Machine translation — one of the earliest AI goals — had proved intractable. Natural language understanding remained primitive. Neural network research stalled after Minsky and Papert’s 1969 book Perceptrons demonstrated fundamental limitations of single-layer networks.[^17][^2]
The decisive blow in the UK came from mathematician James Lighthill. In 1973, commissioned by the British Science Research Council, he published what became known as the Lighthill Report — a comprehensive critique arguing that AI had failed to deliver on any of its major application areas: robotics, language processing, or general problem-solving. The report concluded that the field had not produced significant breakthroughs and had chronically overstated its prospects. The UK immediately ceased government funding for AI research.[^18][^16][^17]
In the United States, two parallel mechanisms caused the same result. The Mansfield Amendment (1969) had already restricted DARPA to funding only research with a direct military application, squeezing out the basic AI research that had flourished in the 1960s. When the Speech Understanding Research program at Carnegie Mellon — a flagship DARPA investment — failed to meet its goals, funders lost confidence across the board. By 1974, AI funding had fallen dramatically, and the field entered what became known as the First AI Winter (1974–1980).[^19][^20]
What Caused the Winter: The Real Lessons
The First AI Winter was not caused by AI being wrong. It was caused by a mismatch between what was promised and what was computationally possible. The hardware of the 1960s and early 1970s simply could not run the algorithms researchers had imagined. Available memory was measured in kilobytes. Training even simple models was prohibitively slow. The “combinatorial explosion” problem — where the number of possible states a program must evaluate grows exponentially — made general intelligence tasks computationally impossible with 1970s hardware.[^19][^21]
Part 4: The Expert Systems Boom (1980–1987)
AI Gets Commercial — But Only for Corporations
The first AI winter ended not because the fundamental problems were solved, but because researchers found a different approach: rather than trying to build general intelligence, build narrow systems that encode the expertise of human specialists. These were called expert systems.
The model was simple: interview domain experts, extract their decision rules, encode those rules in a knowledge base, and build an inference engine to apply them. The resulting systems could not learn; they could only match patterns against rules. But for well-defined domains, they worked remarkably well.
MYCIN (Stanford, developed from the late 1960s, deployed 1970s–1980s) was the landmark medical expert system, designed to diagnose bacterial infections and recommend antibiotics. It contained approximately 450 rules derived from expert interviews and was demonstrated to perform “better than junior doctors” on standardised cases. Critically, MYCIN was never deployed clinically — its developers had concerns about legal liability and the ethics of autonomous medical decisions. It remained a research system, but it proved the concept.[^22]
XCON (formerly R1, Carnegie Mellon / Digital Equipment Corporation) was the breakthrough commercial application. DEC, a major computer manufacturer in the 1980s, had a problem: configuring custom orders for its VAX computers was enormously complex, requiring expensive specialist engineers. XCON automated this entirely, containing over 10,000 rules covering DEC’s full product line. By 1986, it was saving DEC an estimated $40 million per year.[^23][^22]
XCON’s success triggered a frenzy of corporate adoption:[^24][^23]
| Industry | Expert System Application |
|---|---|
| Banking | Loan application evaluation and credit scoring |
| Insurance | Risk assessment and underwriting |
| Healthcare | Diagnostic support (MYCIN, CADUCEUS) |
| Geology | Oil exploration analysis (PROSPECTOR) |
| Manufacturing | Quality control and maintenance scheduling |
| Computer configuration | XCON at DEC; similar systems at other hardware makers |
Who had access in this period? Large corporations with the budget to commission bespoke development, and the government and military. Expert systems of the 1980s cost millions of dollars to build and required specialist knowledge engineers to maintain. The idea that a small business, let alone an individual, could access AI remained entirely science fiction.
Japan’s Fifth Generation Challenge
In 1982, Japan’s Ministry of International Trade and Industry (MITI) launched the Fifth Generation Computer Systems (FGCS) project — a $400 million, ten-year national programme designed to leapfrog Western AI capabilities. The goal was to create computers capable of performing AI-style inference from massive knowledge bases and communicating in natural language. The project alarmed Western governments and triggered competitive government AI spending in the US and UK.[^25][^26]
It ultimately failed to meet its commercial objectives, ending in 1992–1993 without the promised breakthrough systems. But it demonstrated that AI was becoming a matter of national strategic interest — not just an academic curiosity.[^27]
Part 5: The Second AI Winter (1987–1993)
The Collapse of Expert Systems
Expert systems had two fatal weaknesses. First, they could not learn — every new rule had to be manually encoded by a human knowledge engineer. Second, they were brittle: they worked only within their defined domain, and any question outside their rule set produced failure. As businesses discovered these limitations through painful experience, investment collapsed.
The immediate trigger was the hardware market. LISP machines — specialist computers optimised for the LISP language that most expert systems used — had become a significant commercial market. When cheaper general-purpose workstations (principally from Sun Microsystems) proved fast enough to run the same code, the LISP machine market evaporated overnight, destroying hundreds of millions in invested capital.[^17]
Simultaneously, Japan’s Fifth Generation project wound down, DARPA cancelled its Strategic Computing Initiative, and the corporate expert systems market contracted sharply. The Second AI Winter (1987–1993) was underway.[^19][^27]
The lesson encoded into the field from two winters: overpromising kills funding. This lesson shaped how serious AI researchers communicated about their work for the next decade.
Part 6: The Statistical and Machine Learning Renaissance (1993–2012)
A Quieter Revolution
The third phase of AI development was less dramatic than its predecessors but arguably more consequential. Researchers shifted from trying to encode human knowledge into machines to letting machines learnstatistical patterns from data. The pivot was philosophical as well as technical: rather than asking “what rules does an expert follow?” researchers began asking “what patterns exist in large datasets?”
Key milestones in this transition:
| Year | Development | Significance |
|---|---|---|
| 1990 | Support Vector Machines (Vapnik) | New approach to classification; dominated until deep learning[^28] |
| 1995 | Random forests and ensemble methods | Improved accuracy across diverse tasks |
| 1997 | IBM Deep Blue defeats Garry Kasparov | First machine to beat a world chess champion in tournament conditions[^29] |
| 1997 | LSTM networks (Hochreiter & Schmidhuber) | Solved the “vanishing gradient” problem; enabled sequence learning[^28] |
| 1998 | Convolutional Neural Networks (LeCun) | Foundation of modern computer vision[^30] |
| 2006 | Deep Belief Networks (Hinton) | Showed deep networks could learn useful representations[^31] |
| 2009 | ImageNet dataset created (Fei-Fei Li) | 14 million labelled images — the fuel for the deep learning revolution[^32][^30] |
| 2011 | IBM Watson wins Jeopardy! | Demonstrated natural language understanding at scale[^33] |
| 2012 | AlexNet wins ImageNet Challenge | CNN reduced classification error dramatically; launched the deep learning era[^31] |
IBM Deep Blue and Watson: Public Demonstrations of AI Power
Two IBM systems defined public perception of AI in this period. Deep Blue (1997) used brute-force computation — 256 custom processors evaluating 200 million chess positions per second — to defeat world champion Garry Kasparov. It was not general intelligence; it was purpose-built for chess. But it shattered the assumption that the board game most associated with human strategic brilliance was beyond machines.[^29][^34]
Watson (2011) was more sophisticated. It demonstrated natural language understanding at a level sufficient to win Jeopardy! against the show’s greatest champions, fielding questions full of wordplay, metaphor, and cultural reference. IBM immediately pivoted Watson toward commercial healthcare applications, signing a research agreement with Nuance Communications to apply Watson’s capabilities to medical diagnosis. This represented one of the first genuine attempts to bring AI into clinical practice — though the promises again exceeded what could be delivered commercially at scale.[^33][^35][^36]
Who had access in this period?
By the late 1990s and 2000s, AI had begun percolating into consumer products without being labelled as such: spam filters, recommendation engines, credit scoring, fraud detection. These were invisible AI systems, running in the background of daily life. The researchers who built them worked at Google, Amazon, Yahoo, and major financial institutions. Access was still institutional — but the institutions were now commercial technology companies as well as governments.
The ImageNet Moment
Fei-Fei Li’s creation of the ImageNet dataset deserves particular emphasis. Training machine learning algorithms requires data — huge quantities of correctly labelled examples. Before ImageNet (2009), no such resource existed at meaningful scale for computer vision. ImageNet provided 14 million hand-labelled photographs across 21,000 categories. When Alex Krizhevsky, supervised by Geoffrey Hinton, trained a deep convolutional neural network (AlexNet) on ImageNet in 2012, it reduced the classification error rate so dramatically that every subsequent entry in the competition used deep learning. The deep learning era had formally begun.[^31][^30]
Part 7: The Deep Learning and Transformer Revolution (2012–2022)
Compute + Data + Algorithms
The period 2012–2022 produced the architecture that underlies every frontier AI model today. Three factors converged: the availability of massive datasets (ImageNet and its successors), the repurposing of GPUs(graphics processors originally designed for gaming) for neural network training, and a series of architectural breakthroughs that made deep networks trainable at previously impossible scale.[^37]
The most consequential paper in AI history — arguably in the history of technology — was published by a team of eight Google researchers in June 2017. Titled “Attention Is All You Need,” it introduced the Transformer architecture. The Transformer discarded the recurrent neural networks that had previously dominated language modelling and replaced them with a “self-attention” mechanism that allowed every word in a sequence to be processed simultaneously, not sequentially. This made training massively parallelisable on GPUs and enabled models of previously unimaginable scale.[^38][^39]
The impact was immediate and total:
- BERT (Google, 2018): Used the Transformer encoder to achieve state-of-the-art on 11 language benchmarks simultaneously[^40]
- GPT-1 (OpenAI, 2018): Used the Transformer decoder for text generation
- GPT-3 (OpenAI, June 2020): 175 billion parameters — the first model that demonstrated few-shot learning at scale[^41]
- AlphaGo (DeepMind, 2016): Defeated world Go champion Lee Sedol — Go had been considered decades away from being solved[^42]
- AlphaFold (DeepMind, 2020): Predicted 3D protein structures with near-experimental accuracy — transforming biology[^43]
Who had access in 2012–2022? Primarily large technology companies — Google, Meta, Microsoft, Amazon — and well-funded AI startups. Access to frontier models required enormous GPU clusters worth hundreds of millions of dollars. Academic researchers could study the papers but rarely replicate the training runs. For the general public, AI was still something that happened to you invisibly — in your social media feed’s recommendations, your phone’s face unlock, your bank’s fraud detection — not something you could interact with directly.
Part 8: The ChatGPT Era and the Frontier Model Landscape (2022–Present)
The Democratisation Moment
On 30 November 2022, OpenAI released ChatGPT — a conversational interface built on GPT-3.5 — to the general public, free of charge. It reached one million users in five days. 100 million users in two months. No technology in history had been adopted at that speed. For the first time in 80 years of AI history, anyone with internet access could interact directly with a frontier AI system. The exclusivity that had defined the field since 1956 was over.[^41]
The Frontier Model Landscape: A Global Survey
The following tables map every significant frontier model by organisation, country, and launch date.
🇺🇸 United States — Dominant
The United States hosts the majority of the world’s frontier models, produced by a combination of purpose-built AI labs and technology conglomerates.
| Organisation | Model Series | Key Releases | Notable Characteristics |
|---|---|---|---|
| OpenAI | GPT | GPT-1 (2018), GPT-2 (2019), GPT-3 (2020), GPT-3.5/ChatGPT (Nov 2022), GPT-4 (Mar 2023)[^44], GPT-4o (May 2024), o1 (Sep 2024), GPT-5 (Aug 2025)[^45][^46] | Most widely used consumer AI globally |
| Anthropic | Claude | Claude 1 (Mar 2023), Claude 2 (Jul 2023)[^47], Claude 3 family (Mar 2024)[^48], Claude 3.5 Sonnet (Jun 2024)[^49], Claude Opus 4.6 (Feb 2026)[^49] | Safety-focused; Constitutional AI; strong reasoning |
| Google DeepMind | Gemini / PaLM | LaMDA (2021), PaLM (2022)[^50], Bard / Gemini 1.0 (Dec 2023)[^51], Gemini 1.5 (Feb 2024)[^51], Gemini 2.0 (Dec 2024)[^51], Gemini 2.5 (Mar 2025)[^51] | Multimodal from the ground up; long context windows |
| Meta | Llama | Llama 1 (Feb 2023)[^52], Llama 2 (Jul 2023)[^52], Llama 3 (Apr 2024)[^53], Llama 3.1 405B (Jul 2024)[^54], Llama 3.2 (Sep 2024), Llama 4 (Apr 2025)[^55] | Open-weight; democratised model access globally |
| xAI (Elon Musk) | Grok | Grok 1 (Nov 2023)[^56], Grok 1.5 (Mar 2024)[^56], Grok 2 (Aug 2024)[^57], Grok 3 (Feb 2025)[^58], Grok 4 (Jul 2025)[^56] | Real-time access to X/Twitter data; open-sourced Grok 1 |
| Cohere | Command | API launch (2021), Command model (2022)[^59], Command R (2024), Command A (Mar 2025)[^60] | Enterprise-focused; Canadian-founded, US-based |
🇨🇳 China — Rising Challenger
China’s AI development accelerated dramatically after ChatGPT’s launch. By 2023, over 200 Chinese companies had announced LLM projects.[^61]
| Organisation | Model Series | Key Releases | Notable Characteristics |
|---|---|---|---|
| DeepSeek | DeepSeek / R-series | DeepSeek Coder (2023)[^62], DeepSeek-V2 (May 2024)[^63], DeepSeek-V3 (Dec 2024)[^63][^64], DeepSeek-R1 (Jan 2025)[^63], V3.2 / R1-0528 (Sep 2025)[^62] | Trained for a fraction of Western costs; shocked global markets on release |
| Alibaba | Qwen | Qwen 7B/14B (2022)[^61], Qwen2 (May 2024), Qwen3 235B (Mar 2025)[^61] | Open-sourced; multilingual across 201+ languages; competitive with frontier models |
| Baidu | ERNIE | ERNIE 1.0 (2019), ERNIE Bot (Mar 2023)[^61], ERNIE 4.5 / X1 (2025)[^65] | First major Chinese public ChatGPT competitor; China’s largest search company |
| Beijing Academy of AI | Wu Dao | Wu Dao 2.0 (Jun 2021)[^61] | 1.75 trillion parameters at launch — briefly the world’s largest model |
| Zhipu AI | GLM | GLM-130B (late 2021)[^61], GLM-5V-Turbo (2026)[^66] | Tsinghua University spin-off; bilingual English/Chinese |
| Moonshot AI | Kimi | Kimi K2 (May 2026)[^61] | 1 trillion parameter MoE; led Chatbot Arena rankings on launch |
| Tencent | Hunyuan | Hunyuan family (2022+)[^61] | Multimodal; embedded across Tencent’s product ecosystem |
🇫🇷 France — Europe’s Champion
| Organisation | Model Series | Key Releases | Notable Characteristics |
|---|---|---|---|
| Mistral AI | Mistral / Mixtral / Le Chat | Founded Apr 2023[^67]; Mistral 7B (Sep 2023); Mixtral 8x7B (Dec 2023); Mistral Large (Feb 2024)[^68]; Mistral Large 3 (Nov 2025)[^69]; Mistral Medium 3.5 (Apr 2026)[^70] | Open-weight philosophy; Europe’s only credible frontier model competitor; valued at €12B as of Sep 2025[^71] |
Mistral was founded by three researchers who had worked at Meta and Google DeepMind. Its €105 million seed round in June 2023 was, at the time, the largest seed round in European history. The company’s open-weight approach — releasing model weights for free use — positioned it as the European answer to Meta’s Llama, with the additional claim to sovereignty: trained in Europe, on European infrastructure, for European use cases.[^71][^67]
🇬🇧 United Kingdom — Sovereign AI in Development
The UK does not yet have a frontier commercial LLM competing globally, but has significant strategic activity:
| Initiative | Description | Status |
|---|---|---|
| BritLLM / UK-LLM(UCL + Bangor + NVIDIA) | Building an open-source LLM trained on English and Welsh language data, running on the Isambard-AI supercomputer in Bristol[^72][^73] | In development (2023–present) |
| Sovereign AI Unit | £500 million government fund to build and scale AI capabilities on British shores, chaired by VC James Wise[^74][^75] | Launched Nov 2025 |
| AI Pathfinder(Government) | £150 million+ to provide secure AI infrastructure for the public sector and national security[^74] | Active 2025–26 |
DeepMind — founded in London in 2010 and acquired by Google in 2014 — remains the UK’s most consequential AI contribution. AlphaGo (2016) and AlphaFold (2020) are arguably the two most important scientific AI achievements of the past decade, though both are products of a US-owned entity.
🇦🇪 United Arab Emirates — Sovereign AI Strategy
| Organisation | Model | Launch | Characteristics |
|---|---|---|---|
| Technology Innovation Institute (TII) | Falcon | Falcon-40B (May 2023)[^76], Falcon 2 (May 2024)[^77], Falcon 3 family (2024–25)[^78], Falcon-H1 (2025)[^78] | Open-sourced under Apache 2.0; government-funded by Abu Dhabi’s ATRC; positioned as “made-in-UAE AI” |
The UAE has arguably the world’s most ambitious sovereign AI strategy of any middle-power nation, including a dedicated Ministry of AI — possibly the first such ministry globally. Falcon has consistently ranked in the top tier on the Hugging Face Open LLM Leaderboard among non-US/non-China models.[^76][^77]
🇩🇪 Germany — Enterprise Sovereignty
| Organisation | Model | Launch | Characteristics |
|---|---|---|---|
| Aleph Alpha | Luminous / Pharia | Luminous (2022)[^79]; Pharia-1 7B (2025)[^79] | Sovereign LLMs; focuses on multilingualism, explainability, EU regulatory compliance; German, French, Spanish, Italian |
| Black Forest Labs | FLUX (image generation) | 2024 | Image generation models; used by xAI’s Grok for image capabilities[^43] |
🇰🇷 South Korea
| Organisation | Model | Launch | Characteristics |
|---|---|---|---|
| LG AI Research | EXAONE | EXAONE 1.0 (Dec 2021)[^80], EXAONE 2.0 (Jul 2023)[^80], EXAONE 3.0 open-source (Aug 2024)[^80], EXAONE 4.0 (Jul 2025)[^81], K-EXAONE 236B (Jan 2026)[^82] | South Korea’s flagship LLM; strong Korean language capability; LG conglomerate-backed |
| Samsung | Gauss | Samsung Gauss 1 (Nov 2023)[^83], Gauss 2 (Nov 2024)[^84] | Integrated into Samsung devices and Galaxy products |
🇮🇱 Israel
| Organisation | Model | Launch | Characteristics |
|---|---|---|---|
| AI21 Labs | Jurassic / Jamba | Jurassic-1 (Aug 2021)[^85][^86], Jurassic-2 (Mar 2023)[^87], Jamba (Mar 2024)[^85], Jamba 1.5 (Sep 2024)[^85], Maestro (Mar 2025)[^85] | Founded 2017; pioneered commercially available large language models; enterprise focus |
🇨🇦 Canada
| Organisation | Model | Launch | Characteristics |
|---|---|---|---|
| Cohere | Command / Aya | API (2021), Command (2022)[^59], Command R (2024)[^88], Command A (Mar 2025)[^60], Command A+ (2026) | Founded by Transformer paper co-author Aidan Gomez; enterprise and multilingual focus; $5.5B valuation[^59] |
Other Notable Global Contributions
| Country | Organisation | Model | Notes |
|---|---|---|---|
| 🇸🇬 Singapore | Menlo Research | Jan | Open-source; Southeast Asian language focus[^89] |
| 🇨🇭 Switzerland | Swiss AI Initiative | Apertus | Fully open (data, training code, weights); ~Llama 3 level performance[^89] |
| 🇮🇹 Italy | Sapienza NLP / CINECA | Minerva | Trained on 500B+ Italian words; national language AI[^79] |
| EU (multi-country) | EuroLLM Consortium | EuroLLM-9B | Covers all 24 official EU languages plus 11 more; 4T+ token training[^79] |
Part 9: What Held AI Back — A Structural Analysis
For anyone presenting on AI history, it is useful to characterise the brakes on progress structurally, not just chronologically.
The Three Recurring Constraints
1. Compute Every AI winter was, at its root, a compute problem. The algorithms researchers imagined in the 1950s and 1960s required hardware that would not exist for 30–50 years. Even the expert systems of the 1980s required expensive, specialised LISP machines. The GPU revolution of the 2000s — originally a gaming technology — was the single most important enabling factor in modern AI, making it economical to train large neural networks.[^37]
2. Data Statistical machine learning requires large, labelled datasets. These did not exist in usable form until the internet created massive passive data generation, and until projects like ImageNet (2009) created structured labelled corpora. Modern foundation models are trained on essentially the entire digitised output of human civilisation — something that was physically impossible at any earlier point.[^32]
3. Funding and Institutional Access AI research from 1956 to roughly 2010 was accessible only to those with institutional backing: military agencies, elite universities, and large corporations. The academic research community could study and publish, but training cutting-edge models required resources that only DARPA, IBM, or Google could provide. This is why the democratisation of AI via open-weight models (Llama, Mistral, Falcon) and consumer interfaces (ChatGPT) is genuinely historically significant — it ended 65 years of structural exclusivity.
The Hype Cycle Problem
AI has a well-documented tendency to overpromise. Minsky’s 1970 prediction of human-level AI “within 3–8 years” is merely the most famous example. Each wave of excitement has attracted funding, which attracted promises, which attracted disappointed funders, which caused winters. The pattern is visible across:[^16]
- 1960s optimism → First AI Winter (1974)
- Expert systems hype → Second AI Winter (1987)
- 1990s neural network enthusiasm → narrow and slow progress through the early 2000s
- Current LLM excitement → concerns about a potential third cooling period
Understanding this cycle is essential for any professional presenting on AI’s trajectory.
Conclusion: A Field 80 Years in the Making
The AI that became visible to the world in late 2022 was not invented in 2022. It was the product of:
- 1943: The first mathematical model of a neuron (McCulloch & Pitts)
- 1950: The first formal framework for machine intelligence (Turing)
- 1956: The naming and formalisation of the field (Dartmouth)
- 1963–1974: $100M+ in US military funding building the research foundations
- 1980s: The first commercial AI deployment at scale in corporate settings
- 1986: Backpropagation making deep networks trainable (Hinton, Rumelhart, Williams)
- 1997: Deep Blue proving machines could outperform humans in bounded tasks
- 2009: ImageNet providing the data infrastructure for deep learning
- 2012: AlexNet demonstrating the power of deep convolutional networks
- 2017: The Transformer architecture making large language models possible
- 2020: GPT-3 demonstrating emergent few-shot learning at scale
- 2022: ChatGPT making 80 years of research accessible to anyone with internet
The geography of frontier AI is not accidental. American dominance reflects decades of DARPA investment, world-class universities, and the concentration of talent and capital in Silicon Valley. China’s rapid rise reflects determined national strategy, enormous domestic data, and the efficiency-driven innovation forced by US chip export restrictions. France’s Mistral reflects the strength of European mathematical traditions and a deliberate open-weight strategy. The UK’s BritLLM and sovereign AI unit reflect the recognition that dependence on US or Chinese AI infrastructure carries geopolitical risk. The UAE’s Falcon reflects the most explicit governmental AI sovereignty project outside the major powers.[^61]
AI has always been a geopolitical technology. That has not changed. What has changed is that, for the first time in its 80-year history, it is also yours.
Sources and data accurate as of June 2026.
References
- ECS170 : Artificial Intelligence: Neural Networks – As mentioned in the research history McCulloch and Pitts (1943) produced the first neural network, w…
- Chapter 20: From McCulloch-Pitts to Backpropagation – The …
- The History of Artificial Intelligence: A Timeline from Turing to Today – The history of artificial intelligence. From Alan Turing and early AI research to modern generative …
- The birth of Artificial Intelligence (AI) research | Science and Technology – Artificial intelligence research begins at Lawrence Livermore.
- Complete History of AI – SAP LeanIX – Artificial intelligence (AI) isn’t new. Learn the history of AI from early concepts to modern advanc…
- A brief history of AI – Discover AI fundamentals and career opportunities. Learn AI techniques and ethical considerations. I…
- The us army has been innovating in AI before Silicon … – Yes, the U.S. Army (and the broader U.S. military, often through the Defense Advanced Research Proje…
- A history of artificial intelligence in 10 objects – V&A – Here we present 10 objects from the V&A collection that help tell the story of artificial intelligen…
- What is the history of artificial intelligence (AI)? – Tableau – Understanding the history of AI is the key to understanding where we are now, and where we can go. L…
- A Cautionary Tale on Ambitious Feats of AI: The Strategic … – Machine intelligence has been a military research goal for decades, but is it even worth it? Artific…
- DARPA and the Exploration of Artificial Intelligence – An article analyzing the Defense Advanced Research Projects Agency’s (DARPA) work on advancing resea…
- [PDF] AI Watch Historical Evolution of Artificial Intelligence
- The Perceptron – Foundation of Modern Neural Networks – Interactive – In 1958, Frank Rosenblatt created the perceptron at Cornell Aeronautical Laboratory, the first artif…
- What Were AI’s First Successful Expert Systems? – AI and Machine Learning Explained – What Were AI’s First Successful Expert Systems? Have you ever wondered how early AI systems helped s…
- Artificial Intelligence Timeline – Military Embedded Systems
- The History of Artificial Intelligence – IBM – As AI improves, its influence will continue to grow. To understand the directions the technology can…
- The AI Winter: Why AI Almost Died (Twice) – Explore the two major AI winters — the periods when funding collapsed, research stalled, and artific…
- What is AI Winter? Definition, History and Timeline – TechTarget – AI winters are periods of stagnation in interest in funding for AI. Learn the history of AI winters,…
- AI Hype Cycles: Lessons from the Past to Sustain Progress – NJII – The recent wave of achievements in artificial intelligence has understandably generated tremendous e…
- The First AI Winter (1974–1980) — Making Things Think – Holloway – It’s difficult to be rigorous about whether a machine really ‘knows’, ‘thinks’, etc., because we’re …
- Why Did AI Winters Stall AI’s Progress? – AI and Machine Learning Explained – Why Did AI Winters Stall AI’s Progress? Have you ever wondered why progress in artificial intelligen…
- [PDF] Expert Systems Notes
- The History of AI – The History of AI, Article 6 (The Comeback of Knowledge): Expert Systems The First AI Winter had bee…
- Expert Systems: When AI Went Corporate | Artificial Heights – From Labs to BoardroomsBy the 1980s, artificial intelligence had made the leap from academic curiosi…
- Fifth Generation Computer Systems – Wikipedia – The Fifth Generation Computer Systems was a 10-year initiative launched in 1982 by Japan’s Ministry …
- The Japanese Create a Stir (Chapter 22) – The Quest for Artificial Intelligence – The Quest for Artificial Intelligence – October 2009
- 1995: All Institutional… – This document describes the fifth generation of computers, including the fifth generation project la…
- 𝐃𝐞𝐞𝐩 𝐋𝐞𝐚𝐫𝐧𝐢𝐧𝐠: 𝐀 𝐓𝐢𝐦𝐞𝐥𝐢𝐧𝐞 𝐨𝐟 𝐊𝐞𝐲 … – Deep learning has evolved over several decades through continuous advances in neural network…
- Deep Blue | IBM Supercomputer, Artificial Intelligence & Machine … – Deep Blue, computer chess-playing system designed by IBM in the early 1990s. As the successor to Chi…
- Timeline of Deep Learning’s Evolution – 1990s: Yann LeCun pioneers Convolutional Neural Networks (CNNs), laying the groundwork for modern co…
- CHAPTER 4
- Timeline of machine learning – Wikipedia – This page is a timeline of machine learning. Major discoveries, achievements, milestones and other m…
- From checkers to chess: A brief history of IBM AI – IBM has been involved with AI for over 70 years and partners with companies across various industrie…
- Deep Blue: The Chess Supercomputer That Changed AI and IBM … – In 1997, IBM’s supercomputer Deep Blue achieved a landmark victory over Garry Kasparov, the reigning…
- Watson, l’ordinateur d’IBM qui vient de gagner Jeopardy! conquiert le monde de la santé – Paris – 21 févr. 2011: Watson, l’ordinateur développé par IBM au cours des quatre dernières années, …
- IBM To Roll Out Watson, M.D. – In the wake of Jeopardy! success, IBM plans to roll out medical applications
- 2000s AI Milestones: Big Data, GPUs, and Deep Learning – 🧠 The 2000s — when Big Data, GPUs, and new neural breakthroughs powered the dawn of deep learning an…
- Attention Is All You Need – Wikipedia
- Attention Is All You Need — the Transformer — History of AI – Google Brain publishes the Transformer architecture, which replaces all recurrent networks and forms…
- The Evolution of Transformers: From ’Attention Is All You … – Introduction
- ChatGPT timeline: Is OpenAI’s pursuit for speed costing them … – Digit – There was a time when a new OpenAI model felt like an event. GPT-3 arrived in 2020 after nearly two …
- Google DeepMind — AI Wiki – Google’s unified AI research division, formed by merging DeepMind and Google Brain in 2023. Behind G…
- Top 9 AI models from Europe – TechBehemoths – All sectors in the world, including finance, biology, and enterprises, will experience the integrati…
- BrutusAI Result: List The Dates For Each Public Release Of Openai’s Gpt | Brutus AI – BrutusAI Answer for: List The Dates For Each Public Release Of Openai’s Gpt
- Gpt-4 (march 2023) – The complete story of OpenAI’s GPT series – from 117 million parameters to PhD-level intelligence. E…
- ChatGPT version history: Evolution timeline – nexos.ai – Explore ChatGPT evolution, how GPT-5 changed the future of LLMs, and what comes next. Read more abou…
- Claude (language model) – AI Wiki – Claude is a family of large language models (LLMs) developed by Anthropic, an American artificial in…
- Claude Version History (2023–2026): Complete Timeline – Techiefied – Read the full Claude version history | Claude Opus – Claude Sonnet | Claude Cowork – Claude Code – C…
- Anthropic – Wikipedia
- What is Google Gemini? – IBM – Gemini is Google’s large language model (LLM). More broadly, it’s a family of multimodal AI models d…
- Google DeepMind – Wikipedia
- Llama 3.1’s Timeline — Full Story History | Shapes AI – Trace Llama 3.1’s development from Meta’s early Llama releases to its 2024 launch and 2026 updates.
- AI Timeline — Key Events in Artificial Intelligence (2020-2026) – Interactive timeline tracking every major AI milestone from GPT-3 to GPT-5 and beyond.
- The future of AI: Built with Llama – Meta AI – As we close out 2024, Meta is leading the industry forward in AI product and technology experiences …
- The Llama series of models from Meta – Meta’s most popular LLM series is Llama. Llama stands for Large Language Model Meta AI. They are ope…
- Grok (chatbot) – Wikipedia
- Decoding Musk Timelines – DEV Community – In the twelve months between February 2024 and February 2025, Elon Musk’s xAI released three major…..
- Grok, the AI Chatbot From Elon Musk’s XAI, Explained – Elon Musk’s xAI launched its chatbot in 2023 as an alternative to other “woke” bots. Here’s what you…
- Cohere Statistics Statistics: Market Data Report 2026 – Gitnux – Cohere has grown from a 2019 startup into a platform with over 1 million API developers and more tha…
- Cohere | nextomoro – AI Research Lab Intelligence – Cohere is a Canadian artificial intelligence company founded in 2019, developer of the Command, Embe…
- ai, deepseek, machinelearning – title: The Rise of China’s LLMs: A Complete History from 2017 to 2026 published: ture description:…..
- What is DeepSeek AI? China’s Top AI Chatbot Explained – Beebom – DeepSeek AI is an AI chatbot similar to ChatGPT, and it has been developed by a Chinese company. It’…
- DeepSeek | Rise, Technologies, Impact, & Global Response – Building on this momentum, DeepSeek released DeepSeek-V3 in December 2024, followed by the DeepSeek-…
- China’s DeepSeek Launches New AI Model—Here’s What To Know – The Chinese AI company said its latest model demonstrated ‘significant improvements’ in benchmark pe…
- Meet 5 of the Chinese AI models upending the market – One of the latest contenders is Manus, a Chinese AI agent being hailed as the next potential “DeepSe…
- China AI Bulletin 3 – by Emmie Hine – Developments from 8/4/26-29/4/26
- Mistral AI – Wikipedia – Mistral AI SAS is a French artificial intelligence (AI) company, headquartered in Paris. Founded in …
- Frontier AI Models – Frontier AI models represent the cutting edge of artificial intelligence technology, pushing the bou…
- Mistral releases new AI models, rivals US with multiple languages – “Mistral Large 3 was trained on a wide variety of languages, making advanced AI useful for billions …
- LLM Large Language Model Directory – DocsBot – Comprehensive directory of large language models (LLMs) and their capabilities. Compare details, pri…
- What is Mistral AI? Everything to know about the OpenAI competitor – Mistral AI, which offers some open source AI models, has raised significant funding since its creati…
- UCL chosen as UK partner to help develop sovereign AI … – Leading chip designer NVIDIA will partner with UCL to help optimise the UK’s national artificial int…
- UK unveils Sovereign AI collaboration designed for its languages and … – A collaboration led by UK-LLM to develop a large language model enabling AI reasoning for the UK’s l…
- AI to power national renewal as government announces billions of … – A major package of new reforms and investment will put AI at heart of government’s mission to drive …
- Industrial Strategy 2025 – What Does it Mean for AI? – techUK – UK’s AI Sector Government has allocated £500 million to establish the new Sovereign AI unit, as anno…
- Falcon LLM open-source large model · UAE Benchmark – Falcon LLM open-source large model is one of the key initiatives in UAE’s AI strategy.
- UAE gets serious about dominating in GenAI with launch of model … – UAE has formed a Ministry of AI, which may be the first of its type globally TII of the UAE Governme…
- TII’s Falcon AI models to join Amazon Bedrock Marketplace – New AWS collaboration brings scalable access to TII and AI71 AI products
- Europe’s Top AI Models of 2025: Multilingual, Open, and Enterprise … – Explore Europe’s top AI models of 2025—multilingual, open-source, and enterprise-ready. Latest bench…
- LG debuts South Korea’s first open-source AI model – that launched in December 2021. Exaone 2.0, the second generation of the AI language model, was unve…
- South Korea’s LG unveils AI model in quickening global race – LG AI Research announced the release of the Exaone 4.0 on Tuesday. in 2021. It plans to showcase the…
- LG Rolls outs ‘K-EXAONE’: South Korea Joins the Global Frontier AI … – Jan. 12, 2026 announced the public release of its flagship AI model, ’K-EXAONE. Launched in December…
- Samsung joins the AI race with Samsung Gauss – Samsung’s new AI model could debut on the Galaxy S24
- Samsung introduced a new AI model. It will improve the functions … – Samsung has introduced a new model of generative artificial intelligence called Samsung Gauss 2. It …
- AI21 Labs – Wikipedia – AI21 Labs is an Israeli company specializing in Natural Language Processing (NLP), which develops AI…
- Announcing AI21 Studio and Jurassic-1 language models – AI21 Labs’ new developer platform offers instant access to our 178B-parameter language model, to hel…
- AI21 Labs Introduces Jurassic-2, the World’s Most Advanced … – Jurassic-2 offers best-in-class LLMs, bringing flexibility, top-tier quality, and high performance t…
- Cohere: Command A’s Timeline — Full Development History – Trace Cohere: Command A’s journey from Cohere’s founding to its 2025 release and ongoing advancement…
- What are some non US and Chinese AI models – how do they perform? – What are some non US and Chinese AI models – how do they perform?
Generated by BBEdit