Economist 2026-06-15 analysis flags intelligence explosion by 2028 absent policy infrastructure

The article summarizes capability trajectories from GPT-4 to successor systems and notes absence of binding international agreements or verifiable safety evals at deployment scale. Primary data cited includes repeated doublings in effective compute and benchmark saturation rates exceeding prior forecasts by 12-18 months. No national or multilateral body maintains real-time monitoring of training runs above 10^26 FLOP.

Epoch AI records show training compute for leading models increased 4x annually since 2022 while regulatory proposals remain at the discussion stage in both US and EU frameworks. This gap widens as multiple labs report internal metrics crossing human-expert thresholds on software engineering and scientific reasoning tasks simultaneously. Policy documents from 2023-2025 contain no mechanisms for pausing runs once those thresholds are passed.

Operational consequence is that deployment decisions rest solely with private entities whose stated timelines compress the window for oversight to months rather than years. Absent new instrumentation for detecting emergent agency or goal misgeneralization during training, post-deployment correction becomes the only remaining control point.

Next milestones include scheduled releases of models trained on 10^27+ FLOP by late 2027, at which point recursive self-improvement loops become measurable on public benchmarks if prior scaling holds.