Dario Amodei published an essay this month calling on the AI industry to pace itself: more auditing, collective restraint among labs, and coordination among allied governments. Venture investor Tomasz Tunguz, writing at tomtunguz.com, argues the proposal has a hole at its center. Nobody, including Amodei, has said what “pace” actually means in days, months, or FLOPS.
Tunguz sorts the public response into five camps, each pricing the cost of slowing down without naming a rate. That gap matters because “pace” without a number is a request for trust, not a policy, and trust is exactly what a Sacks-versus-Khan fight over regulation cannot supply.
The interpretability camp wants time to understand the systems being built. Anthropic alignment science lead Evan Hubinger puts the chance of a catastrophic AI failure within the next ten years at more than 10 percent, according to CNBC, and says the field has no plan for that outcome. The labor camp wants protection for workers rather than researchers: Bernie Sanders has proposed halting the construction of new AI data centers nationwide until safeguards are in place.
The economic camp inverts the argument entirely. Treasury Secretary Scott Bessent told the G20 in early September that AI-driven productivity growth is how the United States manages its debt load, predicting at the briefing that the benefits would show up within six months, the only dated claim any camp made. The geopolitical camp, represented by President Trump’s remarks in Ireland this month that the United States is “leading China in AI,” treats any slowdown as ceding ground to a rival. The fifth camp, which Tunguz calls regulatory capture, opposes new rules altogether: David Sacks, who chairs the President’s Council of Advisors on Science and Technology, told Amodei and Sam Altman directly to stop seeking permission from Washington, while former FTC chair Lina Khan countered that existing law already applies and needs no carve-out for AI companies.
Tunguz’s sharpest point is historical, not rhetorical. The one tool that could have produced an actual number, a training-compute threshold, was already tried. In 2023 an executive order made labs report any training run past 10^26 FLOPS; the order was revoked in January 2025 before any model had crossed that line. The first model to clear it, Grok-3, shipped weeks later, and Tunguz notes that roughly ten models are on pace to exceed the threshold this year, with frontier compute still growing about fivefold annually. A static ceiling stops mattering once the floor rises to meet it within a single product cycle.
That history is the piece’s real argument against “pace” as a workable policy lever: any fixed number becomes obsolete on the same timescale it takes a lab to ship its next model. For operators, the practical read is not to wait on Washington or on the labs to define a speed limit. Compute budgets, procurement contracts, and model-selection roadmaps should assume the current growth rate holds through 2027, because the only mechanism that could have capped it has already failed once and shows no sign of being reintroduced with better enforcement.
Tomasz Tunguz, “What Does Pacing Mean?,” tomtunguz.com, September 2026.
