●Instrument · Live Projection

The Collapse Clock

Companion instrument to Shorting Human Origin. No fixed baseline — the projected onset of model collapse is computed entirely from the live state of seven sourced markers. Move any of them and watch the date shift.

●[ Collapse Countdown — Live ]
03years
04months
25days
15hours
09min
22sec
Projected onset
March 6, 2030
System severity
Critical56%
Model horizon
30 days (terminal) ↔ 20 years (nominal)
Legacy baseline · Jul 1, 2030 anchorNet shift -221 days
Baseline date
July 1, 2030
Net marker shift
-221 days
Baseline-adjusted onset
November 22, 2029

Earlier versions of this instrument anchored to a fixed July 1, 2030 armageddon date and reported how each marker pushed it forward or pulled it earlier. That view is preserved here as a secondary readout — per-marker day shifts are shown beneath each slider below.

Top drivers · what's pulling the clock forward
  1. 01
    Open-weights parity gap to frontier
    24%
  2. 02
    U.S. interconnection-queue headroom
    21%
  3. 03
    Synthetic share of new web content
    20%
[ Scenarios ]

Markers

Each marker is calibrated against a primary source, with a safe endpoint (contributes nothing to collapse) and a catastrophic endpoint (saturates the severity index). Drag the sliders to model external shocks.

57%

Percentage of newly published long-form web content that is AI-generated. Drives recursive training contamination.

Safe 15% → Catastrophic 90%56% severity · 20% of total
vs. Jul 1, 2030 baseline-45 days
Source: Originality.AI, 2024
42% of estimated stock

Cumulative share of the ~300T-token public human-text corpus already used in training runs.

Safe 20% of estimated stock → Catastrophic 95% of estimated stock29% severity · 10% of total
vs. Jul 1, 2030 baseline+145 days
Source: Epoch AI, 2024
35%

Share of the top 1,000 websites that disallow GPTBot / CCBot / Anthropic-AI in robots.txt. Reduces fresh clean training data.

Safe 5% → Catastrophic 80%40% severity · 9% of total
vs. Jul 1, 2030 baseline+49 days
Source: Originality.AI Top-1000 study, 2024
320$B / year

Combined MSFT + GOOG + META + AMZN annualized capex. Higher rates accelerate sunk-cost lock-in.

Safe 150$B / year → Catastrophic 550$B / year43% severity · 8% of total
vs. Jul 1, 2030 baseline+32 days
Source: Reuters, 2025-02-07
22GW available to new DCs

Gigawatts the U.S. grid can deliver to new data-center load within 24 months. Lower = power wall hits sooner.

Safe 80GW available to new DCs → Catastrophic 5GW available to new DCs77% severity · 21% of total
vs. Jul 1, 2030 baseline-162 days
Source: ERCOT 2024 LTLF; LBNL DC Energy Report 2024
1months

Months by which the best open-weights model trails the best closed model on Chatbot Arena. Smaller gap = faster price-umbrella collapse.

Safe 18months → Catastrophic 0months94% severity · 24% of total
vs. Jul 1, 2030 baseline-240 days
Source: LMSYS Chatbot Arena, 2025
60–10

Composite of EU AI Act enforcement, BIS export-control stringency, and US compute-reporting rules. Higher = faster fragmentation.

Safe 20–10 → Catastrophic 100–1050% severity · 9% of total
vs. Jul 1, 2030 baseline±0 days
Source: EU Reg. 2024/1689; BIS 90 FR 4544; EO 14179
[ Method ]

No fixed armageddon date is assumed. Each marker contributes a severity score ∈ [0, 1] computed as its position between a safe endpoint and a catastrophic endpoint, both chosen against the primary sources cited beside the marker.

Aggregate system severity S is the weight-averaged mean of those scores. The projected onset is then T(S) = 30d + (20y − 30d) × (1 − S)^2.2 — meaning if every marker sits at its safe value, the horizon relaxes to ~20 years; if every marker sits at catastrophic, the window contracts to ~30 days.

This is a projection instrument, not a prophecy. The point is to make the levers visible: the direction and magnitude in which the clock moves when you drag a marker is the signal worth watching.

See Ch 02 — The Sunk Cost Trap & the Poisoned Well · Ch 03 — Model Collapse · Ch 07 — The Power Grid Reckoning