Mark Zuckerberg has rejected the case for a coordinated slowdown in frontier AI development.[1] He argues that every lab already has both the responsibility and the incentive to train models safely, and can slow its own work when necessary.[1]
His position is not that safety can be ignored.[1] Zuckerberg says trust and alignment are becoming competitive advantages because people will reject agents that do not follow their instructions.[1] He also points to legal liability, independent evaluation and Meta’s decision to delay its Muse agent for several months while safety and security work continued.[1]
That differs from Anthropic chief executive Dario Amodei’s proposal to “pace the frontier”.[2] Amodei wants embedded third-party evaluators, coordination between frontier companies in democratic countries, and an attempt at global coordination.[2] He argues that slower capability gains could give alignment and security work time to catch up.[2]
Amodei also acknowledges the geopolitical problem.[2] His essay says a full international pacing agreement is unlikely soon because secretly defecting could radically alter the balance of global power.[2] Zuckerberg’s refusal to join a coordinated slowdown sharpens that problem: a plan built around common restraint becomes weaker when a major lab chooses unilateral responsibility instead.[1][2]
The distinction is therefore about mechanism rather than whether safety matters.[1][2] Amodei argues that competitive pressure requires shared limits.[2] Zuckerberg argues that competition, liability and user demand already push labs towards safer models.[1] CNBC reported that this places him closer to Nvidia chief executive Jensen Huang, while President Trump has also criticised slowing AI development.[3]
Supporting sources:
[1] Mark Zuckerberg on X
[2] Dario Amodei: We Must Pace the Frontier
[3] CNBC
Source: Mark Zuckerberg on X