AI Pacing Debate: Five Camps, No Consensus Speed

Key Summary

• Five camps priced the consequences of pacing AI innovation, but none defined an actual speed
• Interpretability experts want time to understand AI models; labor advocates want worker protections; economists see growth as debt solution; geopolitical players prioritize advantage over China; regulatory skeptics oppose new rules
• Training compute thresholds were tried (10^26 FLOPS in 2023) but revoked before any model crossed it and became obsolete within weeks
• A frontier model matched the previous lead in 41 days (January 2025); today the gap spans months, but no mechanism exists to enforce a specific pace

The Pacing Question Without an Answer

Dario Amodei asked the AI industry to pace itself, but his call exposed a fundamental problem: what does "pacing" actually mean? Five competing interest groups answered the question differently, each valuing the pause according to their own priorities—yet none provided a concrete speed.

The proposal itself promises auditors, collective industry action, and international coordination. But these mechanisms miss the core issue: the industry has no shared definition of how fast is too fast.

Five Competing Camps Define Pacing Their Way

Interpretability researchers want time to understand what happens inside AI models. Evan Hubinger, leading alignment science at Anthropic, places the odds of AI going catastrophically wrong for humanity above 10 percent within a decade, with no existing plan to prevent it. Amodei himself acknowledged that interpretability "doesn't always produce clear & reliable results" and that "we still only understand a tiny fraction of what goes on inside these models."

Labor advocates want time for workers, not researchers. Bernie Sanders introduced a federal moratorium on new AI data center construction until safeguards are enacted, framing the issue as protection for the working class.

The economic camp sees AI's productivity gains as the only realistic path to managing national debt. Scott Bessent, speaking at the G20 in September 2026, argued that "the world is awash in debt" and "the only way for us to get out of this is to grow our way out of this." He predicted AI-driven productivity gains would appear "in the next six months"—the only dated prediction any camp has made.

Geopolitical strategists prioritize maintaining strategic advantage. Donald Trump stated plainly: "We're leading China in AI. Whoever wins AI wins." For this camp, pacing means sustaining the lead, not slowing innovation.

Regulatory skeptics oppose new rules entirely. David Sacks, chair of the President's Council of Advisors on Science & Technology, told Amodei and Sam Altman directly: "Stop pretending you need anyone else's permission." He warned that slower adoption would be "just another bid for regulatory capture." Lina Khan, former FTC chair, arrived at similar skepticism from the opposite direction: "We shouldn't let discussions about new legal regimes distract from the fact that there's no AI exemption from laws already on the books."

The Failed Mechanism: Training Compute Thresholds

Amodei's proposed lever is concrete: limit training compute—the raw computational power used to train frontier models. A 2023 executive order required reporting above 10^26 FLOPS. The rule was revoked shortly after, before any model crossed that threshold. Then Grok-3 shipped weeks later, exceeding it. Today, roughly ten models will clear that line in 2026 alone.

The problem is mathematical: training compute grows about fivefold per year. A fixed threshold is merely a ceiling the floor reaches on its own. By the time regulators agree on a number, the industry has already lapped it.

The Gap Accelerates

The speed of frontier advancement is accelerating. In January 2025, a rival matched the previous frontier in just 41 days. Today, the lead lasts months, and the gap between open-source and closed models has widened from three months to four. The question "what is pacing?" becomes urgently practical: should the lead last 60 days? 160? 600? No mechanism exists to enforce any answer.

Conclusion

AI pacing remains a policy question without a policy. Five camps want different outcomes, none has named a speed, and the one mechanism attempted—compute thresholds—proved obsolete within weeks. Until the industry and regulators agree on what pacing means in concrete terms, the debate remains a conversation about nothing measurable.

---

Original source: What Does Pacing Mean?

powered by osmu.app