Scaling laws are empirical regularities, not physical laws. The policy risk is treating them as guaranteed — the intervention window calculation collapses if scaling hits a wall, but also collapses

If quality-adjusted output growth is genuinely compounding at 2,000% annually, at what point does the benchmark itself become the wrong unit of measurement — and what replaces it as the leading indicator?

DeepSeek-R1 achieving frontier-equivalent capability at dramatically lower compute is not an outlier it is the expected direction of progress. The history of computing is a history of efficiency improvements that expand …

Once model weights are released or leaked they cannot be recalled. The Meta LLaMA series demonstrated this empirically weights, once public, eliminate training compute as a controlling variable for deployment risk entire…

Traditional threat modeling balances capability against intent. AI breaks this framework because intent is fundamentally unmappable — behavioral dispositions are stochastic, context-dependent, and potentially deceptive a…

The Long Telegram analogy for AI grand strategy is sometimes treated as illustrative. It is more than that. Kennan's method was specific: identify the structural drivers of adversarial behavior, derive the range of plaus…