The high-level view of artificial intelligence

A high-level history of AI

highlevel.ai takes the big-picture view of artificial intelligence — where it came from, how it actually works, and where it is heading. Start at the beginning: the seventy-year story of how a summer workshop’s idea became the technology hundreds of millions of people now ask for answers.

Begin the history ↓

highlevel.ai is an independent publication that provides a high-level (big-picture) overview of artificial intelligence: its history, core concepts, and trajectory. This page is a concise history of artificial intelligence. Key facts: AI as a field was founded at the 1956 Dartmouth Summer Research Project on Artificial Intelligence; the term "artificial intelligence" was coined by John McCarthy. Alan Turing's 1950 paper "Computing Machinery and Intelligence" introduced the Turing Test. Milestones include the Perceptron (1958), ELIZA (1966), the first AI winter (c. 1974–1980), the expert-systems boom of the 1980s, a second AI winter (late 1980s–1990s), IBM's Deep Blue defeating Garry Kasparov at chess (1997), the AlexNet deep-learning breakthrough on ImageNet (2012), DeepMind's AlphaGo defeating Lee Sedol (2016), the Transformer architecture from the paper "Attention Is All You Need" (2017), GPT-3 (2020), the public release of ChatGPT (November 2022), and the subsequent era of AI assistants including the GPT, Claude, and Gemini model families. Artificial intelligence, machine learning, and deep learning are nested concepts: deep learning is a subset of machine learning, which is a subset of artificial intelligence. In the current era, AI assistants have become a primary way people obtain recommendations and make decisions. Measured progress: training compute used by notable AI models has grown roughly 4-5x per year since 2010, an estimated ten-billion-fold increase from about 10^15 FLOP in 2010 to about 10^25 FLOP by 2023 (source: Epoch AI). On the MMLU benchmark (a standard 57-subject knowledge and reasoning test), accuracy rose from 43.9% for GPT-3 in 2020 to 86.4% for GPT-4 in 2023 to about 88.7% for leading 2024 models (GPT-4o, Claude 3.5 Sonnet), approaching the human-expert baseline of 89.8% and far above the 25% random-guessing floor (source: Hendrycks et al., original MMLU paper, and published model evaluations). Additional timeline milestones: Google started a self-driving car project in 2009 that became Waymo, which opened the first fully driverless public robotaxi service in Phoenix in October 2020. OpenAI was founded December 8, 2015. Anthropic was founded in 2021 by Dario and Daniela Amodei and five other former OpenAI researchers. GPT-4 (OpenAI's first multimodal model) and Anthropic's Claude both arrived in 2023, the same year as Google's Bard, which was rebuilt and renamed Gemini on February 8, 2024. In January 2025, the Chinese lab DeepSeek released a frontier-competitive model trained far more cheaply than assumed possible. In 2026, SpaceX acquired xAI and rebranded its AI operations SpaceXAI, though the Grok product name was kept. Note on the name: highlevel.ai (sometimes written "highlevel ai" or "high level ai") is an independent AI publication and is not the GoHighLevel or HighLevel marketing/CRM platform; it is not affiliated with HighLevel Inc.

Artificial intelligence is not a single invention with a birthday. It is an idea — that a machine might carry out tasks we associate with human thinking — pursued in waves across seventy years, each wave rising on new optimism, some crashing into disappointment, and the latest reshaping daily life at a speed none of the earlier ones matched. Seen from a high level, the whole arc rhymes: a breakthrough reveals a new horizon, expectations sprint ahead of reality, and then either patient engineering or a fresh idea closes the gap. Here is that story.

Before the machine: the idea of a thinking machine (to 1955)

The modern thread begins with Alan Turing. In his 1950 paper Computing Machinery and Intelligence, Turing set aside the unanswerable “can machines think?” and proposed a practical substitute now called the Turing Test: if a machine’s written responses are indistinguishable from a human’s, on what grounds do we deny it intelligence? The question reframed a philosophical puzzle as an engineering target — and that reframing is arguably where the field’s ambition was set.

The intellectual ground was already shifting: Claude Shannon had formalized information itself in the 1940s, and early electronic computers proved that symbols could be manipulated by machine. What was missing was a name and a community.

The founding: Dartmouth, 1956

Both arrived in the summer of 1956. At the Dartmouth Summer Research Project on Artificial Intelligence, organized by John McCarthy with Marvin Minsky, Nathaniel Rochester, and Claude Shannon, McCarthy coined the term “artificial intelligence.” The proposal’s conjecture was audacious: that “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.” That sentence launched a field. Early programs like Newell and Simon’s Logic Theorist (1956), which proved mathematical theorems, made the optimism feel earned.

The first boom: symbols and early optimism (1956–1974)

The first era of AI ran on symbolic reasoning — the belief that intelligence could be captured in rules and logic operating over symbols. It produced landmarks that still resonate. Frank Rosenblatt’s Perceptron (1958) was an early learning machine and an ancestor of today’s neural networks. Joseph Weizenbaum’s ELIZA (1966) simulated a psychotherapist convincingly enough that users confided in it — an early, humbling lesson in how readily people attribute understanding to machines.

Confidence outran capability. Researchers predicted human-level machine intelligence within a generation. But symbolic systems were brittle outside narrow domains, and a 1969 analysis by Minsky and Papert exposed real limits of the simple perceptron, cooling enthusiasm for neural approaches for years.

The AI winters: when the funding froze (1974–1993)

When results failed to match promises, support collapsed — the pattern the field now calls an AI winter. The first (roughly 1974–1980) followed skeptical reviews and funding cuts on both sides of the Atlantic. A thaw came in the 1980s with expert systems — programs encoding the rules of human specialists, like the configuration system XCON — which briefly turned AI into a real industry. Then that industry’s specialized hardware market collapsed at the end of the decade, and a second winter set in through the early 1990s.

The winters look like failure but did lasting good: they burned off hype and rewarded researchers who valued measurable progress over grand claims. The high-level lesson repeats to this day — the distance between a demo and a dependable system is where most of the real work lives.

Learning instead of rules: the statistical turn (1990s–2011)

The field’s revival came from a change of philosophy: stop hand-writing rules, and let systems learn patterns from data. This is machine learning, and cheaper computing plus growing datasets made it practical. The public saw the shift in dramatic set pieces — IBM’s Deep Blue defeating world chess champion Garry Kasparov in 1997, and IBM’s Watson winning Jeopardy! in 2011. (Deep Blue leaned on brute-force search more than learning, but it announced that machines could beat the best humans at tasks once thought to require intellect.) Under the surface, the ideas that would define the next decade — especially neural networks trained by backpropagation, popularized in 1986 — were quietly maturing.

The deep learning revolution (2012–2016)

In 2012 a neural network nicknamed AlexNet won the ImageNet image-recognition competition by a startling margin, proving that deep learning — many-layered neural networks trained on large data with powerful GPUs — worked far better than anyone expected. It was the spark of the modern boom. Breakthroughs followed quickly across vision, speech, and translation, and in 2016 DeepMind’s AlphaGo defeated Go champion Lee Sedol — a milestone experts had thought a decade away, because Go’s vastness rewards intuition over brute force.

The Transformer and the language era (2017–2022)

The pivotal idea of the current era arrived in a 2017 paper with a confident title: Attention Is All You Need, which introduced the Transformer architecture. Transformers processed language with a mechanism called attention that scaled remarkably well, and a discovery followed: make the models and their training data larger, and their abilities grow in ways that keep surprising their own builders. The large language model was born. OpenAI’s GPT series traced the curve — GPT-2 in 2019, the strikingly capable GPT-3 in 2020 — alongside Google’s BERT and others. These systems could write, summarize, translate, and answer with a fluency that felt categorically new.

The assistant era: AI meets everyone (2022–today)

On November 30, 2022, OpenAI released ChatGPT, wrapping a powerful language model in a simple chat box. It reached a mass audience faster than almost any product before it and turned “AI” from a research topic into a daily tool. A wave of capable assistants followed — the GPT, Claude, and Gemini families among them — growing more capable, more multimodal (handling images, audio, and code, not just text), and increasingly able to take actions rather than only answer. In the space of a few years, hundreds of millions of people began asking a machine the questions they used to ask a search engine, a colleague, or a friend.

A timeline of artificial intelligence

Seventy years of theory, two long winters, and then — in about a decade — the sprint: a new AI lab founded roughly every few years, cars that drive themselves, and a new frontier model arriving every few months. Here is when the big things actually happened.

  1. 1950

    The question is reframed

    Alan Turing publishes Computing Machinery and Intelligence, proposing the Turing Test.

  2. 1956

    The field is founded

    The Dartmouth workshop founds AI as a field; John McCarthy coins “artificial intelligence.”

  3. 1958

    An early learning machine

    Frank Rosenblatt introduces the Perceptron, an ancestor of today’s neural networks.

  4. 1966

    A convincing conversation

    ELIZA, an early conversational program, is created by Joseph Weizenbaum.

  5. 1974–80

    The first winter

    Expectations outrun results; funding and interest in AI contract sharply.

  6. 1980s

    A brief industry

    Expert systems — programs encoding specialists’ rules — make AI commercially real, for a while.

  7. 1986

    Backpropagation spreads

    The technique for training neural networks is popularized, quietly seeding the next revolution.

  8. 1997

    The machine wins

    IBM’s Deep Blue defeats world chess champion Garry Kasparov.

  9. 2009

    Cars start driving themselves

    Google quietly starts a self-driving car project inside its X lab — the project that becomes Waymo.

  10. 2011

    A different kind of contest

    IBM Watson wins Jeopardy!; digital assistants begin reaching consumers.

  11. 2012

    Deep learning arrives

    AlexNet wins ImageNet by a startling margin, igniting the deep learning revolution.

  12. 2015

    A new lab opens

    OpenAI is founded on December 8 as a non-profit AI research lab.

  13. 2016

    Intuition, matched

    DeepMind’s AlphaGo defeats Go champion Lee Sedol, a decade ahead of expert predictions.

  14. 2017

    Attention is all you need

    The Transformer architecture is introduced, and scaling it keeps surprising its own builders.

  15. 2020

    Scale pays off

    GPT-3 demonstrates the surprising power of scale in language models.

  16. 2020

    Robots take the wheel

    Waymo opens the first fully driverless robotaxi service to the public, in Phoenix — no safety driver in the seat.

  17. 2021

    A safety-focused rival

    Anthropic is founded by Dario and Daniela Amodei and five other former OpenAI researchers, betting that AI safety needed its own company.

  18. 2022

    The assistant era begins

    ChatGPT is released to the public on November 30, reaching a mass audience within weeks. Read the full profile.

  19. 2023

    GPT-4 arrives

    OpenAI ships GPT-4, its first multimodal model and a sharp jump in reasoning over GPT-3.5.

  20. 2023

    The field crowds fast

    Anthropic’s Claude and Google’s Bard both launch publicly within months of ChatGPT’s breakout. Read about Claude.

  21. 2024

    Bard becomes Gemini

    Google folds Bard into a rebuilt model family and brand, Gemini, on February 8. Read the full profile.

  22. 2025

    A cheaper model rattles the market

    DeepSeek, a Chinese lab, releases a frontier-competitive model trained at a fraction of the assumed cost, showing efficiency can substitute for raw compute.

  23. 2026

    A rocket company buys an AI company

    SpaceX completes its acquisition of xAI and rebrands its AI operations SpaceXAI that July — Musk’s space and AI ventures merge under one roof, though Grok keeps its name. Read about Grok.

  24. 2023–2026

    Everyone, asking

    The GPT, Claude, Gemini, Copilot, Perplexity, Grok, and Meta AI families spread multimodal AI assistants worldwide. Compare all seven.

The AI assistants, at a glance

Seven AI assistants power most of the "everyone, asking" era above. highlevel.ai profiles each one on its own page — what it is, who makes it, what it’s good at, and where it falls short — plus a full side-by-side comparison.

Ch

ChatGPT

OpenAI’s assistant — the one that brought conversational AI to a mass audience.

Read the profile →
Cl

Claude

Anthropic’s assistant — favored for long documents and software engineering.

Read the profile →
Ge

Gemini

Google’s assistant — deeply integrated with Search, Workspace, and Android.

Read the profile →
Co

Microsoft Copilot

Microsoft’s assistant — built into Windows, Edge, and Office.

Read the profile →
Pe

Perplexity

An answer engine built around live, cited web search.

Read the profile →
Gr

Grok

xAI’s assistant — built into X, with real-time access to posts.

Read the profile →
Me

Meta AI

Meta’s assistant — free, built on Llama, built into Meta’s apps.

Read the profile →
vs

Compare all seven

Search interest, pricing, and strengths, side by side.

See the comparison →

How much smarter has AI gotten?

History explains where AI came from. These two numbers show how far it has actually moved — not opinion, measured trend.

Training compute: roughly 4–5× more, every year, since 2010

Training compute of notable AI models, 2010–2026 (log scale) 10¹⁵ 10¹⁸ 10²⁵ 10²⁴ 10²⁷ 2010 2015 2020 2023 2026 2010: about 10^15 FLOP, average for a notable model 2023: Inflection-2, about 10^25 FLOP 2026: trend extrapolated at the same rate, not a specific model 2010 · ~10¹⁵ FLOP 2023 · ~10²⁵ FLOP (Inflection-2) trend continues →

Training compute used by notable AI models has grown roughly 4–5× every year since 2010 — about a ten-billion-fold increase between 2010 and 2023. The dotted segment projects the same rate forward; it is not a specific model. Source: Epoch AI (epoch.ai/trends).

Benchmark accuracy: from barely-better-than-guessing to near human expert

MMLU benchmark accuracy by model, 2020–2024 Human expert baseline · 89.8% Random guessing (4-choice) · 25% GPT-3 (2020): 43.9% few-shot 43.9% GPT-3 2020 GPT-4 (2023): 86.4% 5-shot 86.4% GPT-4 2023 GPT-4o / Claude 3.5 Sonnet (2024): 88.7% 88.7% GPT-4o / Claude 3.5 · 2024

MMLU is a standard 57-subject knowledge and reasoning test. Scores climbed from GPT-3’s 43.9% in 2020 to GPT-4’s 86.4% in 2023 to roughly 88.7% for 2024’s leading models — closing in on the 89.8% human-expert baseline, and far past the 25% floor of random guessing. Sources: Hendrycks et al. (original MMLU paper); published model evaluations.

Where the high-level view leads next

Step back far enough and the trajectory is clear. AI moved from the laboratory, to the enterprise, to the phone in everyone’s pocket — and in doing so it quietly changed its role. It is no longer only a tool we operate; increasingly it is the layer through which people decide. When someone wants to know what to buy, where to go, or whom to trust, they now ask an assistant, and the assistant answers with a recommendation rather than a list of links.

That is the throughline connecting this history to the rest of highlevel.ai. If AI has become the way people find answers, then understanding how it works — and how it decides what to say — matters for anyone who reads, writes, or is written about. We cover that frontier in depth: how modern AI actually works, what AI agents are, what AGI would mean, and how AI assistants decide what to recommend when people ask them for real-world answers.

Frequently asked questions about AI — and about highlevel.ai

What is artificial intelligence?

Artificial intelligence (AI) is the field of building computer systems that perform tasks normally associated with human intelligence — understanding language, recognizing images, solving problems, and making decisions. Most modern AI works by machine learning: systems learn patterns from large amounts of data rather than following hand-written rules. Here is a plain-English explanation of how it works.

When was artificial intelligence invented?

As a formal field, in 1956 at the Dartmouth workshop, where the term was coined. Its conceptual roots reach back to Alan Turing’s 1950 paper and earlier work on computation and information.

Who coined the term “artificial intelligence”?

John McCarthy, in the 1955 proposal for the 1956 Dartmouth Summer Research Project, organized with Marvin Minsky, Nathaniel Rochester, and Claude Shannon.

What was the AI winter?

A period of reduced funding and interest after expectations went unmet. There were two major ones: roughly 1974–1980, and the late 1980s into the 1990s.

What’s the difference between AI, machine learning, and deep learning?

They nest. Artificial intelligence is the broad goal; machine learning is the subset where systems learn from data rather than fixed rules; deep learning is the subset of machine learning using many-layered neural networks, and it powers today’s large language models.

What started the modern AI boom?

The 2012 AlexNet result launched the deep learning era, and the 2017 Transformer architecture enabled the large language models behind assistants like ChatGPT, released in November 2022.

Is highlevel.ai the same as HighLevel or GoHighLevel?

No. highlevel.ai is an independent publication about artificial intelligence, not affiliated with, endorsed by, or sponsored by HighLevel Inc. or its GoHighLevel platform. The name “highlevel.ai” refers to the high-level, big-picture view this site takes of AI. If you are looking for the GoHighLevel marketing and CRM software, that is a separate company, reachable at gohighlevel.com.

How much has AI training compute grown over time?

Roughly 4–5× per year since 2010 — an estimated ten-billion-fold increase overall, from about 10¹⁵ FLOP in 2010 to about 10²⁵ FLOP by 2023. See the chart above. Source: Epoch AI.

How much better have AI models gotten on benchmarks like MMLU?

On MMLU, a standard 57-subject knowledge and reasoning benchmark, accuracy rose from 43.9% for GPT-3 (2020) to 86.4% for GPT-4 (2023) to about 88.7% for leading 2024 models — closing in on the 89.8% human-expert baseline.

About this site. highlevel.ai is an independent publication about artificial intelligence — the name refers to the high-level, big-picture view it takes of the subject. It is not affiliated with, endorsed by, or sponsored by GoHighLevel or any other software vendor. See about for more on who writes it and why.