SymfieldTopics
Start here

Project · live

AI Lineage Network

rev. Sep 15, 2026PublishedAI & ComputationAI Commentary & Industry

open full screen ↗

You are viewing v1.4, data current as of September 14, 2026. Changes are itemized in the changelog.

AI Lineage Network

How frontier AI reproduces itself, people, methods, models, money, and machines, 2010–2026. Every edge is a cited claim. 94 entities and 197 documented relationships across seven channels, people, methods, models, money, machines, data, and evaluation.

AI Lineage Network

A list of AI labs makes them look like independent competitors. Trace how they formed and a different picture emerges: a few origins reproducing themselves through seven channels.

The map holds 94 entities, labs, people, funders, chip and cloud providers, the twelve frontier model families, and the corpora and benchmarks they share, linked by 197 documented relationships across the US and Chinese ecosystems. Evidence is visible in the lines themselves: solid = documented, dashed = derived or reported, dotted = alleged, never promoted to fact.

What it shows that a list can't: talent moves in a loop, not a pipeline, OpenAI's fission wave began flowing back by 2026. The acqui-hire without the acquisition (Microsoft–Inflection, Google–Character.AI, Meta–Scale) is now a standard maneuver. A handful of groups act as investor, compute provider, and competitor to the same labs, on both flags. And beneath the rivalry sits a commons nobody competes over, shared corpora, shared benchmarks, the most uniformly connected layer in the map.

The Paradigm tab asks the sharpest question: has anyone actually moved?

Current answer: no model has left the dominant paradigm, but the architecture column, constant for years, now carries two documented breaks, Qwen 3.5 and Kimi K3, both interleaving linear-recurrent layers with attention at 3:1 in a released flagship. Within-sample entropy in that column moved from 0.41 to 0.65 bits. A different kind of departure sits underneath: GLM-5.x matches on every paradigm axis but trains end-to-end off the NVIDIA stack, the substrate moved, the paradigm didn't.

Use the tabs to switch views, the toggles to isolate a channel, and click anything for its sources. Search by Network, Timeline, Heatmap, Rankings, Concentration, or Paradigm, in radial or network views.