Skip to content

Eight AI Forces

Open the Eight AI Forces Diagram Fullscreen

About This MicroSim

This interactive diagram visualizes the eight loops from the Winner Takes All? causal-loop analysis. Four reinforcing forces (left, warm tones) try to push AI Model Capability into a runaway regime where one lab pulls permanently ahead. Four balancing forces (right, cool tones) hold the system in oligopoly. The question of whether AI tips into a single dominant player is really a question about which loops dominate, and when.

Click any of the eight icons — or the numbered markers in the panels — to read about that force, see its current (May 2026) status, and understand how it interacts with the others.

The Eight Forces

Reinforcing (try to make capability run away)

  1. R1 — Recursive Self-Improvement — better models make better engineers, who ship better models.
  2. R2 — Autonomous Research — the supercritical loop; the model designs and runs its own experiments.
  3. R3 — Capital → Compute — capital buys compute, compute buys capability. Google's $40B Anthropic commitment lives here.
  4. R4 — Data Flywheel — agent traces and enterprise deployments produce private training signal competitors cannot replicate.

Balancing (try to keep capability in check)

  1. B1 — Compute Constraint — chip supply, datacenters, gigawatts of power.
  2. B2 — Evaluation Bottleneck — you cannot self-improve faster than you can self-evaluate.
  3. B3 — Diffusion / Fast-Follow — papers leak, employees move, ideas spread.
  4. B4 — Cost-Performance Friction — production agents route to the cheapest competent model.

How to Use

  • Explore mode (default) — Click any marker or label to see that force's full description and current 2026 status.
  • Quiz mode — Toggle to Quiz and identify forces from a visual hint.
  • Edit mode — Append ?edit=true to the URL to drag markers and recalibrate positions over the background image.

Source

This MicroSim accompanies the article Winner Takes All? A Systems View of the AI Race, which walks through each of these loops one at a time as a separate causal-loop diagram before assembling them into a full system view.