I mapped the entire AI ecosystem into one interactive graph. Here's what it revealed.

There are over 80 major players shaping AI right now — models, companies, researchers, foundational concepts, seminal papers, and infrastructure. I wanted to see how they all connect. Not as a list. Not as a market map. As a graph.

So I built one. An interactive knowledge graph of the AI landscape with 80+ nodes, 100+ edges, and 6 categories — all rendered as a force-directed network you can explore in your browser.

What surprised me wasn't any individual node. It was the structure. When you lay out the entire ecosystem as a network, three patterns jump out that you can't see from reading the news.

1. Everything traces back to one paper

Click on "Transformer" in the graph and count the connections. GPT-4, Claude, Gemini, LLaMA, Mistral, DeepSeek — every single frontier model is architecturally downstream of one 2017 paper: "Attention Is All You Need" by Vaswani et al.

Eight researchers at Google Brain wrote a paper. Nine years later, it underpins a trillion-dollar industry. The Transformer node has more connections than any company or model in the graph. That's not a coincidence — it's a structural insight about where value actually originates in AI. It's not in products. It's in paradigms.

2. The real moat is infrastructure, not models

Here's what caught my eye: NVIDIA and CUDA are more central to the graph than any single model. The H100 GPU connects to GPT-4, Claude, and LLaMA. CUDA powers both PyTorch and TensorFlow. Every training run, every fine-tune, every inference call flows through this infrastructure layer.

Hugging Face tells a similar story. It hosts LLaMA, Stable Diffusion, Mistral, and thousands of fine-tunes. It's the connective tissue of open-source AI.

If you're thinking about where durable value lives in the AI stack — look at the nodes with the most edges, not the ones with the most hype. The picks-and-shovels thesis isn't just a cliché. It's visible in the topology of the graph.

3. The open-source cluster is now as dense as the closed one

Zoom out and you'll see two distinct clusters forming. One is the closed ecosystem: OpenAI → Microsoft → GPT-4 → Codex. The other is the open ecosystem: Meta AI → LLaMA → PyTorch → Hugging Face → Mistral → Stable Diffusion.

What's remarkable is how densely interconnected the open cluster has become. LLaMA spawned thousands of derivatives. Mistral's Mixture-of-Experts models are served through vLLM. Everything runs on PyTorch. The open graph isn't sparse anymore — it's a self-reinforcing network with its own gravity.

DeepSeek sits at the intersection: a Chinese lab that used open techniques (MoE, chain-of-thought) to match frontier performance at a fraction of the cost. In the graph, it connects to both clusters. That structural position tells you something about where AI competition is heading.

How I built it

The tech is simple. The data curation was the hard part.

Stack: A single HTML file. D3.js for the force-directed layout. No build step, no framework, no backend. The entire graph — data, rendering, and interactions — lives in ~650 lines of vanilla JavaScript.

Data model: Each node has an id, label, type (model/company/researcher/concept/paper/infrastructure), description, year, and an importance score (1-10) that controls its visual size. Edges carry a relationship label: "built-by", "evolved-into", "technique", "architecture", and so on.

The real work was deciding what to include and how things connect. I curated every node and edge by hand — reading papers, tracing lineages, mapping which researchers moved between which labs. The graph is opinionated. That's the point. A knowledge graph isn't a database dump. It's an argument about what matters and how things relate.

Interactions: Click any node to highlight its neighbourhood. Search by name or concept. Filter by category. Navigate between connected nodes through the detail panel. Press / to search, Esc to reset.

What's next

This is v1. I'm planning to expand it with a timeline slider (watch the graph evolve from 2017 to 2025), pull in citation data from OpenAlex to weight edges by influence, and add a "shortest path" feature — click two nodes and see the chain of connections between them.

If you're building in AI, I'd encourage you to try mapping your own landscape. The act of deciding what connects to what forces a level of clarity about the ecosystem that reading articles never gives you.