IntelliGenesis

Feature 2.1 · Evolution timeline

How AI capability evolved

Each stage shows the technology, what it unlocked, what held it back, and a milestone — with sources for every historical claim. The final two stages are violet and badged: hypothetical, not history.

  2. Expert systems

    Unlocked
    Encoded scarce human expertise as thousands of if-then rules.
    Limitation
    Expensive to build, hard to maintain, helpless with uncertainty or novelty.
    Representative systems
    MYCIN (medicine) · XCON (computer configuration)
    Milestone
    XCON saves its manufacturer millions per year — the first big commercial AI win.

    Sources: Expert systems — overview (MYCIN, XCON)

  3. Statistical machine learning

    Unlocked
    Learned patterns from data: spam filters, recommendations, search ranking.
    Limitation
    Needed hand-engineered features; each model stayed locked to one task.
    Representative systems
    Support vector machines · Random forests
    Milestone
    Data + compute start beating hand-written rules on real-world tasks.

    Sources: Machine learning — overview

  4. Deep neural networks + GPUs

    Unlocked
    Learned its own features: breakthroughs in vision and speech.
    Limitation
    Data-hungry, opaque, and still one-model-per-task.
    Representative systems
    AlexNet · Deep speech recognition
    Milestone
    AlexNet wins ImageNet 2012 by a wide margin — deep learning takes over vision.

    Sources: AlexNet (2012) — overview

  5. Transformer architecture

    Unlocked
    Scaled sequence modeling across text; one architecture for many language tasks.
    Limitation
    Enormous training cost; fixed context windows; still pattern-based.
    Representative systems
    Transformer · BERT
    Milestone
    “Attention Is All You Need” (2017) introduces the architecture behind modern LLMs.

    Sources: Vaswani et al. — “Attention Is All You Need” (2017)

  6. Large language & diffusion models

    Unlocked
    Fluent text, image, and code generation usable by a general audience.
    Limitation
    Hallucinates, holds no persistent goals, waits for prompts.
    Representative systems
    GPT-3 · ChatGPT · Stable Diffusion
    Milestone
    Few-shot learning (2020) and mass adoption (2022) move AI into everyday life.

    Sources: Brown et al. — “Language Models are Few-Shot Learners” (GPT-3, 2020)

  7. Tool-using AI agents

    Unlocked
    Multi-step jobs with tools: browsing, running code, calling APIs.
    Limitation
    Unreliable over long horizons; still needs human supervision.
    Representative systems
    ReAct-style agents · Coding assistants
    Milestone
    Reasoning + acting loops let models use tools instead of just producing text.

    Sources: Yao et al. — “ReAct: Synergizing Reasoning and Acting” (2022)

  8. Multimodal foundation models

    Unlocked
    Joint reasoning over text, image, audio, and video in one model.
    Limitation
    Still narrow, supervised, and unable to redesign itself.
    Representative systems
    CLIP-style encoders · GPT-4o · Gemini
    Milestone
    Joint image–text pretraining (2021) matures into consumer multimodal assistants.

    Sources: Radford et al. — “Learning Transferable Visual Models” (CLIP, 2021)

  9. Human-breadth general intelligence (unachieved)

    Hypothetical
    Unlocked
    Would handle any cognitive task a skilled human could — including unfamiliar ones.
    Limitation
    A threshold nobody has verifiably crossed; definitions themselves are debated.
    Representative systems
    — no verified system —
    Milestone
    A conceptual goalpost, not a dated event. Treat any claimed date with skepticism.

    Speculative content — labeled, not sourced (see content rules).

  10. Superintelligence (speculative)

    Hypothetical
    Unlocked
    Would surpass humans in every domain and redesign its own intelligence.
    Limitation
    Entirely hypothetical — included to reason about, not observed anywhere.
    Representative systems
    — no such system exists —
    Milestone
    A thought experiment about where recursive self-improvement could lead.

    Speculative content — labeled, not sourced (see content rules).

Explore the AGI architecture →All sources on the References page