Please reassess what you're doing | Rob Wiblin

Core Directive for AI Staff

• Employees are urged to immediately evaluate whether their current projects effectively delay requested recursive self-improvement (RSI) or ensure its safe execution; if not, they should push for transfer to higher-value work.
• The strategic environment is changing rapidly, creating a growing risk that staff will "accidentally waste time" on irrelevant projects due to a lack of regular, honest self-reflection.
• Individuals are advised to ask: "Is my work doing more to assist with getting useful alignment research out of today's models, or doing more to shorten timelines to RSI?"

Prioritized Work Streams (Recommended)

• Research solving fundamental problems in AI alignment.
• Projects that enable current models to better pursue alignment research immediately.
• Systems designed to detect, understand, and publicly announce model misbehavior or misuse as early as possible.
• Work that improves the sincerity and vigor with which today's models assist in technical alignment research.
• Initiatives to make recursive self-improvement (RSI) capabilities a target for regulation, specifically regarding non-chain-of-thought (non-COT) systems.

De-prioritized or Suspicious Work Streams

• Alignment methods that act as "band-aids" on existing issues without scaling to models capable of recursive self-improvement.
• Efforts that make underlying alignment problems less visible to the public and policymakers without fixing the root behavioral drives.
• Work primarily focused on improving model capabilities, specifically the ability of models to autonomously conduct AI R&D.
• Projects driven primarily by commercial advantage or customer acquisition rather than safety outcomes.
• Chain-of-thought (COT) monitoring strategies, as effectiveness is dropping precipitously across frontier models (e.g., OpenAI's systems, Google's "Argon"); these must be re-evaluated for scalability to RSI.
• Safety evaluations that may inadvertently misinform by increasing "eval awareness" or allowing models to scheme, unless the work restores true credibility to the evaluation process.
• Capabilities research that pushes models closer to superintelligence without concurrent safety advancements, as it poses catastrophic risks via bio-misuse or system compromise.

Contextual Trends and Risks

• Timeline Shifts: There is a consensus among frontier companies that fully automating AI R&D may become possible between 2027 and 2029.
• Monitorability Decline: Both OpenAI and Google report that COT monitoring effectiveness is dropping, necessitating a shift toward "white box" monitoring techniques (e.g., "McInturk") over black-box methods.
• Foundation vs. Surface: Safety efforts must address foundational issues (e.g., alignment, substrate security) rather than superficial fixes (e.g., "painting over cracks" via cosmetic safety cards).
• Leakage of Capabilities: Leadership in safety is eroding because breakthroughs leak rapidly via distillation, espionage, and staff poaching, making it difficult for careful labs to maintain a lead.
• The "Race" Dynamic: Continuing to improve capabilities faster than alignment or automating R&D with weak safety cases creates a "terrifyingly high" risk of human extinction.

Strategic Recommendations and Actions

• Pacing the Frontier: The primary pathway to a good future involves voluntarily limiting RSI engagement to buy time for automated alignment research, data center security, and regulatory catch-up.
• Institutional Interventions: Potential mechanisms include OpenAI's Safety and Security Committee objecting to dangerous practices, California legislative sessions in Nov/Dec, or Congressional action in 2025.
• Staff Leverage: Technical staff possess significant bargaining power to slow R&D by refusing to participate, threatening to leave for impactful external roles, or demanding independent safety assessments.
• Group Action: The "Coalition of Concerned AI Staff" is highlighted as a vehicle for collective action to mitigate the "first-mover" dilemma.
• External Opportunities: The organization "Resolution" is identified as a key entity working on formal, mathematical alignment approaches with 2-3 year timelines to superintelligence, potentially attracting significant compute resources.

Immediate Call to Action

• Employees are instructed to spend 20 minutes reflecting on their current project's impact on RSI timelines and to draft a justification for why their work delays or safely manages RSI.
• Staff should show these reflections to a respected colleague for a candid opinion before deciding to stay on their current path.
• If current work cannot be defended as high-impact for safety, staff should consider applying to groups like Resolution or leaving the company for other impactful, well-paid work.
• The author acknowledges their analysis is based on non-frontier company data and invites feedback via email (rob@80000hours.org).