One card today. The prior six days captured nearly everything that cleared the bar from September 14 onward. This piece from Tunguz published on September 14 had not yet been collected.
Compute thresholds as pacing policy
Tom Tunguz published "What Does Pacing Mean?" on September 14, 2026, the same week Dario Amodei's public call to "pace" AI development circulated widely. Tunguz's observation: nobody defined the word.
Fixed compute thresholds fail as AI pacing policy because training compute grows roughly five-fold annually, making any static ceiling self-defeating maps the five camps that answered the question. The interpretability camp wants time to understand what models are doing before deploying them further. The labor camp wants displacement protection for workers. The economic camp sees AI growth as the only viable mechanism for servicing the CapEx debt already committed. The geopolitical camp cannot accept any slowdown that China would not match. The regulatory-capture camp opposes new rules entirely. Each camp priced the consequences of a pause differently. None named a speed.
The only mechanism that could produce a measurable number was a training compute threshold. The 2023 U.S. executive order tried this at 10^26 FLOPS. That order was revoked before any frontier model crossed it. Grok-3 shipped shortly after, at a scale that would have crossed the threshold. Training compute for frontier models grows roughly five-fold annually, so any fixed ceiling becomes the frontier within 12 months. Regulatory rulemaking cycles run 18 to 24 months. The math makes static thresholds structurally obsolete before they take effect.
Tunguz notes that the interpretability camp's counterproposal, capability-based thresholds rather than compute-based ones, avoids the scaling-obsolescence problem. What models can do changes more slowly and more visibly than how much compute trained them.
- Tom Tunguz, Fixed compute thresholds fail as AI pacing policy because training compute grows roughly five-fold annually, making any static ceiling self-defeating. Fixed compute thresholds fail as AI pacing policy because training compute grows roughly five-fold annually, making any static ceiling self-defeating before enforcement begins.