Frontier Lab Commitments

OpenAI Downtrend in CoT Monitorability Astra is harder to monitor.
OpenAI To Slow Down Development When Safety Needs It In An Alien Mind, OpenAI commits to “unilaterally withhold further scaling as needed.”
Anthropic Pacing the Frontier, and Embedded Evaluators Anthropic has unilaterally committed to embedding third-party evaluators within its organization to verify any pacing commitments.
Anthropic Improving Alignment and Security Anthropic committed to pausing frontier RL training for a fixed time, hardening internal security, and improving alignment audits.
Anthropic Responsible Scaling Policy Anthropic’s Responsible Scaling Policy is its framework for categorizing different levels of risk and associated commitments as capabilities scale.
Anthropic Public Benefit Fund Anthropic commits 5% of its shares to a new public benefit fund, subject to stockholder approval.
OpenAI Independent Auditing In February 2023, OpenAI said its efforts should submit to independent audits before releasing new systems.
OpenAI 20% of Compute to Superalignment OpenAI committed to give its superintelligence alignment effort 20% of the compute it had secured to date.
OpenAI Frontier Governance Framework OpenAI released the Frontier Governance Framework, which contains useful information about how it grades its own models.