Swyx on AI.Engineer + State of SWE

Jud Rosenblat critiques safety community circular firing squads

Rosenblat called out AI safety advocates for lacking cognitive empathy toward the Trump administration and holding allies to unreasonably high standards instead of constructive engagement.

Clumsy action beats perfect inaction

Guest Lyron Shapi supported the administration's interventionist approach to AI, viewing export controls and engagement as legitimate paths toward an effective pause despite ham-handed execution.

Hosts shift from criticism to constructive engagement

After reflection, the hosts acknowledge they were initially too harsh on the administration's lack of sophistication, recognizing that early governmental engagement deserves encouragement rather than ridicule given the vast knowledge gap between AI insiders and general society.

Dean Ball joins OpenAI for strategic futures

The former Trump administration adviser who authored America's AI Action Plan will lead a new team focused on frontier AI policy rather than traditional government affairs or lobbying.

Diffusion counters nationalization risks

Ball argues broad technology diffusion is essential to prevent state monopolization of AI, creating constituencies that resist nationalization while maintaining public benefits.

Regulating invisible models presents challenges

The team faces the paradox of governing recursively self-improving systems that labs may deploy internally without public disclosure, making traditional transparency-based regulation difficult.

Rumors of secret Mythos 2 development

Unverified reports suggest Anthropic has completed a new model internally despite keeping Mythos and Fable inaccessible even to most employees, raising questions about undisclosed capabilities.

Internal deployment creates governance gaps

If labs deploy advanced models internally while restricting employee access, they create sleight-of-hand situations where powerful systems exist outside regulatory visibility and public accountability.

Recursive self-improvement accelerates unpredictably

The industry is entering weird times where models may achieve 8x coding output improvements through self-improvement loops, potentially triggering government takedown demands that could force deletion of valuable internal systems.