Zo Garden

DIA-005-loops

4 min read

DIA-005: When the Loops Meet

Snapshot

  • Diagram file: garden/content/ai-governance/part-1/diagrams/DIA-005-loops.svg
  • Theme: Multi-loop interaction around the AI capability stock
  • Primary audience takeaway: Frontier capability is shaped by a forming reinforcing loop, an inverted balancing loop, and a weak brake — sharing one stock, none catching up.

Loop nomenclature (internal — NOT displayed on the diagram)

These L# labels are working references for our conversations. The diagram renders the loops via color and arrow style only; no R/B/L badges are drawn on the canvas.

  • L1 — AI capability → training rate. Forming reinforcing loop. Drawn as red curve from the left-offset port on the stock’s bottom edge back to the training T-fitting. “AI begins to build AI” — capability now feeds back into its own training rate.
  • L2 — capability gap → investment & compute → training rate. Working balancing loop. Drawn as black info-links: stock right-offset port → capability gap node → investment & compute node → training T-fitting port. The classic Meadows-style inventory adjustment loop: gap drives investment, investment drives training, training closes the gap.
  • L3 — visible harm → investment & compute. Inverted balancing loop (paradoxical reinforcement). Drawn as deeper rust info-link sweeping from harm down across the bottom and back up to investment & compute. Section 5’s claim: alarm signals raise spending instead of reducing it. The corrective inverts.
  • L4 — visible harm → oversight & response → investment & compute. Weak balancing loop (the intended brake). Drawn as thin sage dotted info-links. The brake is real and closes the loop, but its strength is small relative to L3 and L1, so it cannot bind. Style choice: dotted/thin/low-opacity to indicate weakness — note this is a deviation from strict Meadows convention; we accept the deviation because it adds clarity without inventing new badges.

Why L4 closes to investment & compute (not capability gap)

Real-world oversight does not change the target the system aims at — it raises the cost of getting there. Liability law, compliance burden, regulatory delay, public scrutiny all act on the funding side: making investment less attractive, slower, more expensive. That is an investment-side intervention, not a goal-setting one. So oversight & response feeds back to investment & compute.

Color and style conventions in the SVG

  • Black info-links: ordinary causal influence (capability gap chain — L2)
  • Rust (#A85B49): forming reinforcing loop (L1)
  • Deep rust (#823F30): inverted balancing loop, paradoxical reinforcement (L3)
  • Sage thin dotted (#5A8270, dasharray 2,3, opacity 0.7): weak balancing loop (L4)

Structural plan (current implementation)

Main pipe (left to right): source cloud (data · chips · talent) → training T-fitting → AI CAPABILITY stock → deployment T-fitting → sink cloud (products · users).

Stock bottom info-link source ports:

  • Left-offset port (rust outline): emits L1
  • Right-offset port (black outline): emits L2’s first segment

Training T-fitting bottom port: receives the closing segments of L1 and L2.

Deployment T-fitting bottom port: emits the first segment of L3 and L4 (via visible harm).

Node naming choices

  • Clouds: named with small italic labels (data · chips · talent on left, products · users on right). Clarifies what’s outside the system boundary without cluttering the canvas.
  • Visible harm: named with the article’s term, with a small italic sub-line listing concrete modes (bias · errors · layoffs). This addresses the reader’s “what kind of harm” question while keeping the structural label short.
  • Oversight & response: positioned at y=360 (was y=395) for cleaner spacing relative to the harm node and L4 paths.

How to read it

  1. Start at the AI CAPABILITY stock between training (inflow) and deployment (outflow). Source cloud feeds raw inputs; sink cloud receives products/users.
  2. Notice the two ports on the stock’s bottom edge — separate origins for L1 and L2.
  3. Trace L1 (red, curving back from left port to training T) — AI accelerating its own training.
  4. Trace L2 (black) through capability gap → investment & compute → training T — the working balancing loop.
  5. From deployment T, follow to visible harm.
  6. From visible harm, two branches:
    • L3 (deeper rust, swept low across bottom, back to investment) — the inverted balancing loop.
    • L4 (thin sage dotted, through oversight & response, back to investment) — the intended brake, weak.

Governance relevance

Two reinforcing loops climbing share a single stock with two corrective loops that fail in different ways: L3 inverts (alarm becomes spending), L4 is too weak to bind (regulatory pressure exists but is small relative to L1 and L3). The shifting-the-burden archetype names why structural reform pressure does not accumulate — visible safety apparatus absorbs the urgency that would otherwise reach liability law, audits, hard deployment gates.

Suggested caption

A forming reinforcing loop, an inverted balancing loop, and a weak brake — sharing one stock.

Graph View