OpenAI has proposed industry-wide limits on AI development tied to safety milestones, starting with a 0.70 log-perplexity cap. The move aims to ensure alignment and monitoring research can keep pace with rapid advances.
TL;DR
- OpenAI proposes shared limits on AI development based on safety milestones, starting with a 0.70 log-perplexity cap.
- The initiative aims to ensure alignment and monitoring research can keep pace with rapid AI advances.
- OpenAI shares internal perplexity measurements to foster industry-wide collaboration on safety standards.
What happened
OpenAI Chief Scientist Jakub Pachocki released internal log-perplexity measurements and proposed shared limits on frontier AI development on November 22, 2026. The proposal suggests pausing development at specific safety milestones, starting with a 0.70 log-perplexity cap, to ensure alignment and monitoring research can keep pace with rapid advances.
Pachocki reported that GPT-3.5 measured about 0.99, supporting fluent conversation, while GPT-4 measured about 0.91, enabling chain-of-thought reasoning. New capabilities emerged at losses below 0.8, and notable results on open mathematics problems began at about 0.76.
The proposal calls for frontier developers to establish evaluation norms, red-team safety cases before model deployment, and shared transparency standards. OpenAI disclosed the figures to encourage urgent shared safety measures, despite giving competitors a clearer picture of its position.
Why it matters
This initiative could set a precedent for industry-wide collaboration on AI safety, potentially slowing the pace of development to ensure responsible advancements. For developers and startups, this means adhering to new safety standards and milestones, which could impact project timelines and resource allocation.
Investors may see this as a sign of increased focus on safety and responsible AI development, potentially influencing funding decisions. The proposal highlights the need for continued investment in alignment and monitoring research to keep pace with rapid AI advances.
The competitive angle is significant, as OpenAI is calling for industry-wide collaboration. This could lead to a more level playing field, with all developers adhering to the same safety milestones and standards. However, it also raises questions about how competitors will respond and whether they will agree to these shared limits.
Key facts
- OpenAI proposes shared limits on AI development based on safety milestones, starting with a 0.70 log-perplexity cap.
- GPT-3.5 measured about 0.99 log-perplexity, supporting fluent conversation.
- GPT-4 measured about 0.91 log-perplexity, enabling chain-of-thought reasoning.
- New capabilities emerged at losses below 0.8, and notable results on open mathematics problems began at about 0.76.
- OpenAI attributes the steep acceleration in its 2026 loss curve mostly to long-term research and compute investments.
- The proposal calls for establishing evaluation norms, red-team safety cases before model deployment, and shared transparency standards.
- OpenAI disclosed internal perplexity measurements to encourage urgent shared safety measures.
- The acceleration scenario is conditional, with rapid advances dependent on AI contributing substantially to algorithmic research.
Context
This proposal comes at a time of rapid advancements in AI capabilities, with concerns about the pace of development outstripping safety and alignment research. OpenAI's move highlights the need for industry-wide collaboration to ensure responsible AI development.
The focus on log-perplexity as a metric for setting development limits is notable, as it is a key indicator of model performance. However, OpenAI acknowledges that perplexity alone cannot carry the entire standard, and cross-lab comparisons are difficult due to different evaluation datasets and training mixtures.
The broader AI landscape is seeing increased focus on safety and ethics, with calls for regulation and responsible innovation. OpenAI's proposal aligns with this trend, emphasizing the need for concrete oversight milestones and shared transparency standards.
