Version 2

The AI safety community's governance debate has expanded on three fronts. Plan A's US-China cooperative pause framework has drawn substantive reactions from Vitalik Buterin, who defends it as the only approach that takes superintelligence's power-concentration risks seriously, and Ryan Greenblatt, who calls it the best plan anyone has written up [1]. Nathan Lambert reports White House discussions of an executive order that could ban frontier-capability open-weight models within six months, and accuses Anthropic's campaign against Chinese model distillation of being regulatory capture [3]. A separate Alignment Forum post argues that political will—not research—is now the main AI safety bottleneck, citing a 3.6:1 researcher-to-advocate ratio and evidence that AI companies secured seven times as many European Commission meetings as civil society in 2023 [2].

If Lambert's report of imminent executive action is accurate, the open-weight debate moves from policy discussion to binding regulation within months—and the regulatory-capture accusation against Anthropic, if credible, would mean safety arguments are being deployed for commercial advantage. The political-will framing suggests technically sound safety proposals go nowhere without advocacy infrastructure the AI safety field currently lacks by roughly two orders of magnitude compared to climate.

The AI safety community is split between a cooperative international approach and domestic competitive controls. The AI Futures Project's Plan A proposes that the US and China jointly control chip supply, audit data centers, and share research information to slow the development of superintelligent AI [1]. Zvi Mowshowitz, after examining Plan A and the reactions it generated, argues the proposal deserves serious engagement rather than reflexive dismissal; his central framing is that the key dividing question is whether superintelligence arrives soon enough to justify Plan A's costs—those who think it won't should oppose Plan A, while those who think it will should engage with it seriously [1]. Vitalik Buterin defends Plan A against critics who call it naive, arguing they apply coordination skepticism to the pause but not to the alternative: an AI transition that 'just goes well by default,' where humanity's hard power drops to zero if AIs can perform every task better [1]. Ryan Greenblatt calls Plan A the best plan anyone has written up while acknowledging it is unlikely to happen; his probability estimates—cited separately—suggest that strong political will could reduce conditional AI takeover risk from approximately 45% to 7% [2][1].

The domestic-controls side has moved toward potentially more aggressive regulatory action. Export controls on model weights took effect in January 2026, and Nathan Lambert reports White House discussions of an executive order that could ban or indefinitely delay open-weight models above the capability level of current frontier closed models, potentially within six months [3]. Lambert treats this as the most serious regulatory threat open-source AI has faced and accuses Anthropic's campaign against Chinese model distillation of constituting regulatory capture: Anthropic 'would gain substantial economic security in its products if the Chinese model makers they accused were banned' [3]. He disputes the stated security rationale by noting that Anthropic's Mythos model was accessed through unauthorized Discord channels even during private beta, suggesting model APIs are not meaningfully more secure than open weights [3]. Anthropic has publicly called for a coordinated, verifiable pause on frontier AI development [4][5], a position Lambert's critics say sits uncomfortably alongside a lobbying campaign that would primarily benefit Anthropic commercially.

A parallel argument holds that the entire governance debate is being fought on the wrong terrain. Charbel-Raphaël, writing on the Alignment Forum, argues that most top global policymakers have never had a serious conversation about catastrophic AI risk, and this absence—not a shortage of research—is why safety measures are not being implemented [2]. He cites a 3.6:1 researcher-to-advocate ratio in the US AI safety field that he argues should be inverted, notes that AI companies secured seven times as many European Commission meetings as civil society in 2023, and estimates the entire AI safety governance field is roughly two orders of magnitude smaller than the climate advocacy ecosystem [2]. Buck Shlegeris is quoted framing the situation as a list of roughly 40 non-hard interventions that would probably prevent major harm—none of which AI companies have the time or appetite to implement [2].

The Trump administration has structured AI policy around competitiveness rather than safety, preempting state AI regulations, declining to build a formal licensing regime, and leaving governance to executive discretion [6][7][8]. Anthropic ended contract negotiations with the Pentagon after the Defense Department demanded blanket 'anything lawful' usage rights that Dario Amodei said left no room for Anthropic's redlines on mass surveillance and autonomous weapons targeting [6]. Legal analysts have concluded a coordinated lab pause may be the most valuable safety intervention available but that antitrust law complicates direct lab-to-lab agreements without regulatory authorization [9][10].