AI History — From Expert Systems to Foundation Models

# AI History — Key Milestones

timeline = {
    # ── SYMBOLIC AI ERA ──────────────────────────────────────────
    1950: "Alan Turing proposes 'Computing Machinery and Intelligence' — Turing Test",
    1956: "Dartmouth Conference — term 'Artificial Intelligence' coined by John McCarthy",
    1966: "ELIZA chatbot — first natural language processor (just pattern matching)",
    1980: "Expert Systems boom — rule-based AI for medical diagnosis, chess",

    # ── AI WINTER 1 & 2 ─────────────────────────────────────────
    1987: "AI Winter — expert systems too brittle, too expensive to maintain",
    1997: "Deep Blue beats Garry Kasparov — rule-based AI, NOT machine learning",

    # ── ML REVOLUTION ───────────────────────────────────────────
    2001: "Random Forests, SVMs become dominant ML algorithms",
    2006: "Geoffrey Hinton: Deep Belief Networks — neural networks revival begins",

    # ── DEEP LEARNING BREAKTHROUGH ──────────────────────────────
    2012: "AlexNet wins ImageNet by ~10% margin — Deep Learning era begins",
          # Alex Krizhevsky, Ilya Sutskever, Geoffrey Hinton
          # 8-layer CNN on 2x GTX 580 GPUs — proved GPUs change everything

    2014: "GANs invented by Ian Goodfellow — generative AI foundations",
    2015: "ResNet — 152-layer network, residual connections solve vanishing gradient",
    2016: "AlphaGo beats Lee Sedol — RL + deep learning at superhuman level",

    # ── TRANSFORMER ERA ──────────────────────────────────────────
    2017: "Attention Is All You Need — Transformer architecture (Vaswani et al.)",
          # This single paper changed EVERYTHING. The entire LLM era traces here.
    "2018a": "BERT (Google) — bidirectional Transformers for NLP, SOTA on 11 benchmarks",
    "2018b": "GPT-1 (OpenAI) — first large generative pre-trained Transformer",
    2019: "GPT-2: too dangerous to release? 1.5B params. Refused. Then released.",
    2020: "GPT-3: 175B parameters. Few-shot learning. In-context learning discovered.",

    # ── FOUNDATION MODEL ERA ─────────────────────────────────────
    2021: "DALL-E 1, CLIP — text-to-image generation with language-image alignment",
    2022: "ChatGPT launch — 100M users in 2 months. Fastest product ever.",
          "Stable Diffusion — open-source image generation democratized",
          "RLHF (Reinforcement Learning from Human Feedback) — InstructGPT",
    2023: "GPT-4 multimodal, Claude 2, Llama 2 open-source, Gemini Pro",
          "LoRA, QLoRA — fine-tuning LLMs on consumer hardware",
    2024: "GPT-4o, Claude 3, Gemini 1.5 Pro (1M context), Llama 3",
          "AI Agents, multimodal reasoning, real-time voice models",
    2025: "GPT-4o1 reasoning models, Claude 3.5/4, Gemini 2, Llama 4",
          "Agentic AI workflows, multi-modal reasoning at scale",
    2026: "Open-weight models rival closed APIs, on-device AI, AI coding agents",
}

# WHY DID DEEP LEARNING WIN IN 2012?
three_factors = {
    "Data":    "ImageNet — 1.2M labeled images. Scale of data never seen before.",
    "Compute": "GPUs (CUDA) — 100x speedup over CPU for matrix operations",
    "Algorithm": "ReLU activations, Dropout regularization, better initialization",
}

# All three came together at the same time. That's why 2012, not 2002.
print("The 3 factors that unlocked Deep Learning:")
for k, v in three_factors.items():
    print(f"  {k}: {v}")