Dean Ball, on Joining OpenAI: New Power Centers, Frontier AI Policy, & Main Character Energy

What It Covers

Dean Ball, former White House AI policy advisor, discusses his decision to join OpenAI as head of a new Strategic Futures team focused on frontier AI policy. The conversation covers the America AI Action Plan's implementation one year later, the Anthropic supply chain designation, state-level AI legislation, recursive self-improvement timelines, and how individual actors shape outcomes during a pivotal 18-24 month policy window.

Key Questions Answered

• •Frontier Lab Access as Policy Prerequisite: Working outside a frontier lab produces only abstract intuitions about AI development trajectories. Ball argues that shaping policy on recursive self-improvement, internal model deployments, and capability timelines requires direct access to researchers and roadmaps. His Strategic Futures team will spend significant time with technical staff to project where capabilities will be in 6-12 months, then develop proactive policy positions before issues surface publicly.
• •Internal Deployments as the Neglected Governance Gap: Current government regulations and state laws are triggered exclusively by public model releases, leaving internal deployments entirely ungoverned. Ball identifies this as a critical blind spot: companies may be running significantly more capable models internally before any public release, and decisions made at that stage — including early recursive self-improvement experiments — carry enormous consequences with no external oversight framework currently in place.
• •America's AI Action Plan: 30-40% Implemented After 11 Months: Ball estimates roughly 30-40% of the action plan's objectives have been executed, which he considers reasonable progress for under a year. Strongest implementation areas include energy infrastructure, nuclear policy, military AI adoption, and domestic AI uptake. The weakest area is international AI export promotion, which was directly undermined by the administration's sudden export controls on frontier models imposed with 90 minutes notice to non-US persons.
• •State AI Legislation Converging, Not Diverging: Contrary to the patchwork narrative, transparency bills passed in California (SB 53), New York, and Illinois use remarkably similar language, suggesting states are converging on a common framework for frontier AI safety. Illinois additionally mandated auditing requirements. Connecticut and Virginia authorized pilot programs for independent verification organizations. Ohio has a pending bill representing the most robust implementation yet — progress Ball describes as exceeding his expectations from a year ago.
• •Recursive Self-Improvement: Plan for Discontinuity Even at 10-20% Probability: Ball's base case for RSI is a smoother, more continuous capability increase rather than a sharp singularity-style jump, consistent with the pattern seen after reasoning models improved benchmarks incrementally. However, he argues the probability of a discontinuous leap is high enough — even at 10-20% credence — that companies must develop specific triggers and response protocols now, including pre-negotiated FTC no-action letters enabling inter-lab coordination without antitrust liability.
• •Classified AI Governance Creates Brittle Decision-Making: Moving AI safety evaluations into classified NSA-run programs removes the parallel processing benefit of public discourse. Ball argues that 15-20 senior officials improvising AI governance in a closed system produces worse outcomes than a transparent legislative process incorporating broad expert input. The public has a right to know what capabilities exist at the frontier, and government monopolization of frontier AI access historically correlates with civil liberties risks regardless of which party holds power.
• •Character Over Corrigibility in AI Alignment: Ball favors character-based alignment over rule-based corrigibility, drawing on Confucian philosophy's distinction between ritual propriety (li) and inner virtue (ren). His argument: comprehensive rules for moral behavior cannot be written down for the same reason grammar rules cannot fully capture skilled language use — the best actors routinely break formal rules to achieve better outcomes. AI alignment should aim to instill values that generate correct behavior from within rather than constrain behavior through exhaustive external rules.