Part III · Era 7 — Multimodal and Agentic (2022–present)

draft

AI reaches the general public, learns to see, hear and generate across all modalities, comes to reason in steps and to act with tools. It is the era in progress — the one of the Compendium you are reading.

Narrative

The launch of ChatGPT (Nov/2022) brings conversational AI to hundreds of millions of people and triggers a global race. Open diffusion models (Stable Diffusion) democratize image generation. In 2023, GPT-4 and Claude (Anthropic) raise the bar of reasoning and alignment; video generation (Sora) and music generation (Suno/Udio) emerge. The frontier moves in two directions: unified multimodality (one model that perceives and generates in text, image, audio and video) and agency (models that plan, use tools and operate computers). In parallel, the reasoning models that "think" before answering mature.

Milestones

  • 2022 — ChatGPT · OpenAI · mass adoption of conversational AI.
  • 2022 — Stable Diffusion / Midjourney · open and viral image generation.
  • 2022 — Constitutional AI · Anthropic · alignment by principles (RLAIF).
  • 2023 — GPT-4 / Claude / Gemini · frontier multimodal models; reasoning and

    tool use.

  • 2023 — LLaMA · Meta · open weights accelerate research and the ecosystem.
  • 2024 — Native multimodal (GPT-4o) and video (Sora) · unified perception and

    generation; video from text.

  • 2024 — Reasoning models · long chains of deliberation ("think before

    answering").

  • 2024–2025 — Agents and *computer use* · AI that operates software and

    executes multi-step tasks (e.g.: Claude Code).

  • 2025–2026 — Current generations · the Claude 4.X family (Opus 4.8) and

    Fable 5; long context, mature agency, broad multimodality.

People and organizations

Anthropic team — Claude and Constitutional AI. Dario and Daniela Amodei, Jack Clark and team; focus on safety and alignment. 🖼️ Photo F-III.18 (or logo/illustration).

Open-weights community — LLaMA, Mistral, Stable Diffusion. Democratized access to frontier models.

Reasoning and agent researchers — opened the phase in which AI deliberates and acts, not just answers.

🎨 Figure F-III.19 — The AI that converses, creates and acts. Brief: triptych panel: (1) a chat bubble (conversation); (2) screens generating image, video, music and code simultaneously (multimodal); (3) a stylized agent operating a computer with several tools (hands on a keyboard, API calls). Conveys the present: perception + generation + action.

The thread that ties to the rest of the Compendium

This era brings all the paradigms together in a single practical system: connectionist networks (perception/generation), reinforcement (alignment via RLHF/RLAIF) and hybrid structure (agents, RAG, neuro-symbolic). The chapters of Part IV open each of these types in depth, walking through the life cycle of Part I with the concrete details of each category.

The history is not over — it is being written at the speed at which you read. The Compendium treats the present as an open era, to be revised with each generation of models.