Skip to content

P1-C1 · A Brief History of 70 Years of AI

Key Insight

4 winters, 4 revivals — "This time is different" requires proof, not assumption.

AI Industry Knowledge — History → Technology → Value Chain → Business → Applications → Geopolitics

P1-C1 (Part 1 has 10 chapters, Chapter 1). After this chapter, you can explain in 5 minutes: AI's 70 years across 5 eras + 4 winters + why today's industry leaders are these companies.

1. The Problem: You Think AI Started in 2022

Many people first heard of AI after ChatGPT (2022/11), so they naturally think, "It's new, no ceiling."

But AI is 70 years old, with 4 instances of "this time is different" — all 4 crashed:

Era Bold Claims Then Outcome
1956 Dartmouth "Machines will surpass human brains in 20 years" 1st Winter 1974-80
1980s Expert Systems "AI will replace doctors / configurators / lawyers" 2nd Winter 1987-93
1997 Deep Blue "General intelligence is next" AI renamed "machine learning" to avoid stigma (90s-2000s)
2016 AlphaGo "Self-driving cars in 5 years" Waymo still stacking miles in 2026

You must be able to articulate why this time (LLM) is different — otherwise, you're going all-in before the 5th winter.

2. The Solution: 5 Eras + 4 Winters Timeline

Each of the 5 eras represents a technological paradigm shift. All 4 winters follow the same pattern: expectations exceeded the compute / data / algorithm capabilities of the time.

graph LR
    A[1956 Dartmouth<br/>Genesis] --> B[1974-80<br/>❄️ 1st Winter]
    B --> C[1980s<br/>Expert Systems]
    C --> D[1987-93<br/>❄️ 2nd Winter]
    D --> E[1997 Deep Blue<br/>Statistical Learning]
    E --> F[2006 CUDA<br/>2012 AlexNet]
    F --> G[2017 Transformer ★★<br/>2020 GPT-3]
    G --> H[2022 ChatGPT<br/>1M users in 5 days]
    H --> I[2026 Now<br/>Claude 4.7 / Stargate $500B]
    style B fill:#fee2e2,stroke:#dc2626
    style D fill:#fee2e2,stroke:#dc2626
    style G fill:#fef3c7,stroke:#f59e0b
    style H fill:#dbeafe,stroke:#2563eb
    style I fill:#dcfce7,stroke:#16a34a

3. How It Works: 5 Eras in Detail (1 Leader + 1 Winter Trigger per Era)

3.1 Era 1: Genesis (1956-1974)

  • 1956 Dartmouth Conference — McCarthy coins "AI". Bold claim: "Will surpass human brains in 20 years"
  • 1957 Rosenblatt Perceptron — The first neural network capable of learning
  • 1969 Winter Trigger: Minsky & Papert prove single-layer perceptrons cannot solve XOR → Neural network research stagnates for 15 years

→ Historical Lesson: After a breakthrough, there must be sustained, explainable progress. "Black box works" isn't enough — academia and capital need to see "why it works + how much further it can go."

3.2 Era 2: Expert Systems (1980-1987)

  • 1980s expert systems boom — XCON (DEC) / MYCIN (medical). Half of Fortune 500 deployed them
  • Japan's Fifth Generation project cost $850 million
  • 1986 Backpropagation (Rumelhart-Hinton-Williams) — Hope for neural network revival, but compute power was insufficient
  • 1987 Winter Trigger: Expert system maintenance costs explode (number of rules → brittle). LISP machine market collapses, DEC disbands AI lab

→ Historical Lesson: Commercialization must scale — 1 expert system working doesn't mean 100 will. AI is a long-tail problem.

3.3 Era 3: Statistical Learning (1993-2012)

  • 90s-2000s AI renamed "machine learning" / "data science" to avoid stigma (many ML companies didn't mention AI in their IPOs)
  • 1997 Deep Blue defeats Kasparov — but it's brute-force search + hand-coded chess rules, not true intelligence
  • 2006 Hinton "Deep Belief Networks" — The 3rd revival of neural networks
  • Key Shift: Explosion of internet data + emergence of GPUs (NVIDIA launches CUDA in 2006) — setting the stage for the next era

3.4 Era 4: Deep Learning Revolution (2012-2017)

  • 2012 ★ AlexNet (Krizhevsky/Sutskever/Hinton trained on NVIDIA GPUs) wins ImageNet, error rate drops from 26% to 15%
  • This is the true starting point of NVDA's dominance — Academia discovers GPUs train neural networks 50x faster than CPUs
  • 2014 GAN (Goodfellow) · 2016 AlphaGo · Early Tesla Autopilot
  • But AI is still a "perception tool," not general intelligence — self-driving was overhyped, leading to a chill later

→ **NVDA's Key Moment: Jensen Huang decides in 2014 to go all-in on data center GPUs (shifting from gaming cards). Intel didn't believe it. History doesn't pick leaders by luck; it's about bets made 10 years earlier.**

3.5 Era 5: LLM Explosion (2017-Present)

  • 2017 ★★ Transformer "Attention is All You Need" (Google Brain — Vaswani et al.) — paradigm shift
  • 2018 BERT (Google) + GPT-1 (OpenAI)
  • 2020 GPT-3 (175B parameters, impressive zero-shot performance)
  • 2022/11 ChatGPT — 1M users in 5 days, AI's first mass adoption
  • 2024 o1 reasoning model + Claude 3.5 Sonnet
  • 2025 DeepSeek V3 (China breakthrough) + Stargate $500B (epic US capex)
  • 2026 Now — Claude 4.7 / Sonnet 4.7 1M context

→ Why isn't OpenAI Google? — The 2017 paper was written by Google, but Google's internal AI safety / brand risk slowed product rollout. OpenAI shipped ChatGPT directly in 2022, capturing an 18-month window. This is a classic startup vs. incumbent case.

4. vs. The AI Story You've Heard Before

If you've only followed media since 2022, you think AI = ChatGPT + NVDA. In reality:

What You Thought Historical Truth
AI started with ChatGPT 70 years of accumulation; this is Era 5
NVDA suddenly exploded 2006 CUDA → 2012 AlexNet → 2014 data center pivot, 20-year strategy
OpenAI is the tech leader The real breakthrough paper was by Google; OpenAI is the product leader
"This time is different" because of compute People in the 1980s/1990s also said "exponential compute growth is enough" — 4 times it wasn't
AI winter won't come again History said "won't come again" 4 times; it came 4 times

History doesn't tell you AI will definitely have another winter, but it tells you "this time is different" requires explicit proof — not a default assumption.

5. Try It: Pick 1 AI Company from a Winter Era, Read for 5 Minutes

Task: Choose one company below, spend 5 minutes on Wikipedia, and think: "If you invested in this company in 1985, how would you have died?"

  • Thinking Machines Corporation (CM-5 supercomputer, 1983-1994 bankruptcy)
  • Symbolics (LISP machine leader, 1980-1996 bankruptcy)
  • Cyc / Cycorp (Lenat's knowledge base, 1984-present, 40 years without commercialization)
  • Teknowledge (expert systems, 1981-2007 value destruction)

Self-check (3 items checked → proceed to P1-C2):

  • You can articulate what "this time is different" meant for that company
  • You can explain why it failed (technology / business / timing)
  • You can draw an analogy to a current AI company (CRWV / OAI / Anthropic / Cohere / Stability)

3 items all yes → You've built a historical anchor; you can proceed to P1-C2.

6. What's Next

You now know the 5 eras. But why was the 2017 Transformer paper a paradigm shift rather than another "AI breakthrough hype"? Why did Scaling Laws (training data + parameters + compute = capability) actually deliver?

→ P1-C2 · Transformer Revolution + Scaling Laws explains how that 8-page paper from 2017 reshaped a trillion-dollar industry.

7. Deep Dive (optional): The Common Pattern of 4 Winters + The Proof Checklist for This Time

Click to see the 4-step AI Winter Cycle + 5 Items

All 4 winters followed the same 4 steps:

  1. Technology Breakthrough (Perceptron / Expert system / Deep Blue / AlphaGo)
  2. Academic Hype ("Will surpass human brains in 20 years" / "AI will replace doctors" / "Self-driving in 5 years")
  3. Capital Influx (DARPA / Japan's Fifth Generation / VC / Now hyperscaler capex)
  4. Expectation-Reality Gap → Capital retreats → Academics rebrand → Consolidation → Next cycle

Proof Checklist for "This Time is Different" (LLM) (Your thesis must answer these):

  • Real User Base — ChatGPT >900M WAU + >50M subscribers (per OpenAI official 2026/03/31 — R6 Fix 3); Claude consumer MAU not officially disclosed (third-party estimate ~30M MAU, range 18M-220M; Anthropic does not publish consolidated MAU/WAU), but enterprise: Amazon Bedrock 100,000+ customers run Claude per AWS announcement + Anthropic >300,000 business customers per Series F, not a demo
  • Real Revenue Growth — Microsoft AI business run-rate >$10B by FY25 Q2 → >$37B by FY26 Q3 (official earnings call 2026/04/29; Copilot-only ARR not separately disclosed), not PR
  • Real Capex ROI Signal — Hyperscalers consistently raising capex guidance (counter-signal: if it flattens/drops = warning)
  • Real Technology Scaling — GPT-3 → GPT-4 → o1 → o3 capabilities still improving exponentially
  • Real Application Layer Diversification — Not just ChatGPT 1 product, but verticals (Cursor / Harvey / Clay)

If even 1 of the 5 items doesn't hold → 5th Winter Warning. Right now, all 5 hold — this is the real basis for "this time is different," not faith.

Key Warning: Historical winters weren't because the technology was wrong; they were about the time lag between expectations and delivery. The biggest risk now isn't that LLMs won't work; it's the mismatch between 2026 hyperscaler combined capex $600-725B (Big 4, ~75% AI-related, Yahoo/CreditSights 2025-12) and the speed of ROI realization.

8. Further Reading (this chapter — 70 years of AI + 4 winters)

All free sources, aligned with P5 0-paid policy

Classic papers / primary sources:

Wikipedia (3-10 min, full timelines + primary citations):

Videos / public lectures (~1-2 hr each):

Podcasts (1-3 hr each):

Books (library):

  • Cade Metz "Genius Makers" (2021) — Hinton / LeCun / Bengio trio + DeepMind / OpenAI origins
  • Stuart Russell "Human Compatible" (2019) — UC Berkeley professor, AI safety perspective
  • Kai-Fu Lee "AI Superpowers" (2018) — US-China AI competition view (slightly China-optimistic, good historical context)

Pair with this chapter's self-check:

After reading any 3-4 of the sources above, you should answer the self-check more confidently — especially "why did it not succeed (technology / business / timing)" and "analogy to 1 AI company today."