OpenAI Wants to Use a Single Number to Put a Ceiling on AI Development
OpenAI's chief scientist Jakub Pachocki published a post arguing that perplexity should be used as a universal metric for measuring AI intelligence, and disclosed OpenAI's internal pretraining progress data. He proposes that the industry establish shared safety limits on this basis: until humanity's ability to understand AI behavior reaches a sufficient level, models must not cross a clearly defined development red line.
OpenAI's chief scientist Jakub Pachocki published an article arguing that perplexity should be used as a universal metric for measuring AI intelligence, and he disclosed OpenAI's internal pretraining progress data. He proposes that the entire industry establish shared safety limits on this basis: before humanity's ability to understand AI behavior reaches a sufficient level, models must not cross a clearly defined development red line.
The reasoning is that modern AI is essentially a compressor, or a predictor. Models are pretrained on massive amounts of data and learn to predict the next unit of content (token). Perplexity measures this predictive ability: the lower the number, the more accurate the prediction. In practice, people usually track log-perplexity (also called cross-entropy loss). Pachocki calls the model produced by pretraining an 'alien mind': an intelligence with hard-to-predict emergent capabilities. Subsequent prompting, fine-tuning, and reinforcement learning are merely steering this mind's capabilities.
There is a clear correspondence between falling perplexity and jumps in capability. GPT-3.5 acquired fluent conversational ability at a log-perplexity of about 0.99 and became the basis for ChatGPT; GPT-4 could effectively use chain-of-thought reasoning at about 0.91. OpenAI's log-perplexity has now dropped below 0.8, and transformative capabilities are appearing in cybersecurity and mathematics.
Evolution of log-perplexity across OpenAI models, recorded starting with GPT-3.5 (Source: OpenAI)
The moment that felt most surreal to him and his team occurred in mathematics. OpenAI spent more than a year posing open math problems to models, and when the models were upgraded to a log-perplexity of about 0.76, they began producing breakthrough results. Answers to long-standing open problems started appearing one after another in Slack threads.
These capabilities cannot be predicted in advance. As the metric drops to specific points, they suddenly emerge. Pachocki believes this is exactly why perplexity is more worth tracking than downstream benchmarks: using modern benchmarks to evaluate early models is of limited value, and the benchmarks from that era had already saturated by GPT-4. By comparison, the metric of pretraining progress remains useful over the long term. He predicts that when the metric enters the 0.68–0.7 window, new and unknown capabilities will appear; what exactly they will be, there may be no corresponding test methods for them today. At the same time, existing alignment audit methods will become even less effective.
What worries him is not only the capabilities themselves, but also AI accelerating its own research. In recent models, OpenAI has already seen concrete contributions from AI in design. Pachocki judges that 'avalanche-style breakthroughs' similar to those in mathematics could first appear in pretraining research as early as 2027. AI will also accelerate alignment research, but likely not as fast as pretraining progress.
This is exactly why he proposes setting limits. Pachocki recommends that frontier model developers jointly establish 'shared limits' based on metrics such as perplexity, committing not to cross the red line until the ability to understand AI behavior reaches specific milestones, such as developing a successor to CoT monitors. He also calls for shared transparency standards and red-team safety testing protocols before new models are first used. This approach continues the idea previously put forward by Anthropic CEO Dario Amodei: 'we must pace the frontier.' Pachocki acknowledges that disclosing data will let competitors see where they stand, but the urgency of establishing shared safety standards is greater.
Perplexity is not a perfect universal metric. Different labs use different datasets, making cross-lab comparisons difficult; it does not map directly to specific task capabilities; reasoning techniques such as chain-of-thought do not change perplexity but substantially improve capabilities; and no one knows the intrinsic entropy of the data—the theoretical lower bound for perplexity. The closer it gets to that lower bound, the slower apparent progress becomes, and each small decrease may correspond to a larger jump in intelligence.
But Pachocki believes that this metric has stood the test of time and is one of the few measures that consistently delivers gains in general intelligence. OpenAI has chosen to take the first step and hopes the industry can jointly establish this scientific foundation.
发布时间: 2026-11-23 05:45