researchOpenAI

OpenAI's Pachocki proposes perplexity as a shared metric for industry safety limits, publishes internal figures

TL;DR

OpenAI Chief Scientist Jakub Pachocki argues that perplexity (cross-entropy loss) should be a shared metric for comparing model developers and informing safety standards. OpenAI is publishing internal perplexity measurements, citing log-perplexity of ~0.99 for GPT-3.5 and ~0.91 for GPT-4.

3 min read
0

OpenAI Chief Scientist Jakub Pachocki has published a post arguing that perplexity, or equivalently cross-entropy loss, should be established as a metric for comparing model developers and informing shared safety standards. As a first step, OpenAI is releasing internal perplexity measurements for its pretraining progress. The post is dated November 22, 2026.

The figures

Pachocki reports log-perplexity on what he calls OpenAI's current main evaluation dataset, a held-out set the company has tracked for years. Lower perplexity means better next-token prediction. The post's opening plot shows the metric for OpenAI models starting with GPT-3.5. Figures stated in the text:

  • GPT-3.5: log-perplexity of ~0.99, which he describes as capable of fluent conversation and the basis for the original ChatGPT.
  • GPT-4: ~0.91, which he describes as proficient at using chain-of-thought reasoning productively. That motivated the targeted elicitation of the capability and the creation of modern reasoning models.
  • Below 0.8: Pachocki says OpenAI has seen the emergence of transformative capabilities in fields like cybersecurity and mathematics.
  • ~0.76: after OpenAI upgraded to a model at this level, it began producing major results on open mathematics problems the company has posed to its models for more than a year. Pachocki describes researchers watching resolutions appear "one after another in a single Slack thread."

He writes that it was not clear in advance that these capabilities would appear at those thresholds, and describes them as emergent latent capabilities from progress on this one metric.

The proposal

Pachocki expects further transformative capabilities as models enter the ~0.68-0.7 log-perplexity window, and says he does not know what they will be or whether anyone is testing for them. He also expects further degradation in currently dominant forms of alignment auditing and further increases in evaluation and training awareness.

He proposes that frontier developers jointly establish shared norms for evaluating such models and red-teaming safety cases before first use, including shared transparency standards. He also proposes common limits, based on perplexity and other measurements, that developers commit not to cross until understanding and monitoring of AI behavior reaches a concrete milestone, such as a successor to chain-of-thought (CoT) monitors. The goal, in his words, is an industry standard that specifies a first limit on a metric trackable across labs, what must be achieved before going beyond it, and a framework for successive limits.

Pachocki also raises the possibility of AI contributing to algorithmic research on pretraining. Drawing on the mathematics experience, he says that "as soon as 2027" the field could conceivably see a similar avalanche of fundamental pretraining advances. He expects alignment and monitoring progress is unlikely to be fast enough to allow the full potential acceleration in pretraining. He says publishing OpenAI's internal numbers gives competitors a clearer picture of where it stands, but that this concern is outweighed by the urgency of shared safety measures.

Stated limitations

The post lists challenges with perplexity as a shared metric:

  • Cross-lab comparison: developers use different evaluation datasets and data mixtures. Pachocki believes transferable comparison methods can be found if transparency becomes a goal.
  • Detachment from actionable evaluations: perplexity does not directly correspond to a task of interest. Its value is that advances are expected to lead to new capabilities.
  • Gains outside pretraining: chain-of-thought reasoning does not reduce a model's perplexity even though it sharply increases capability.
  • Unknown floor: Pachocki does not expect a loss of 0 to be achievable.

What this means

The post ties a single pretraining metric to specific capability thresholds in OpenAI's own models, and uses it as the basis for a proposed cross-industry limit. By Pachocki's account, benchmark-based evaluation loses value as tests saturate, as happened around GPT-4, while pretraining loss remains informative. The post's own caveats limit how directly the numbers can be used today: the figures come from one OpenAI dataset, and the post says cross-lab comparison would require agreed datasets or transferable methods. The proposed limits are a stated position from OpenAI's chief scientist, not an agreed industry standard.

Related Articles

research

Tencent Worked With International AI Safety Groups to Assess Risk of Its Hy4 Open-Weight Model

Tencent Holdings worked with an international group of AI safety organisations to assess the risk level of its latest open-weight model, Hy4. The South China Morning Post reports it is the first time a leading Chinese tech company has disclosed using independent third-party evaluators for a new model before release.

benchmark

GPT-6.2 Astra and Sol cut answer length 34 to 40 percent but fall sharply on Admits Uncertainty

An Artificial Analysis evaluation finds OpenAI's GPT-6.2 Astra and Sol produce answers 34 to 40 percent shorter than their predecessors and lead with the answer more often. Both score lower on uncertainty and transparency, and trail Anthropic's Claude Fable 5.2 and Opus 5.5 on the Transparency and Organization indexes.

changelog

OpenAI tells developers to rewrite system prompts for GPT-6.2 Astra and Sol as defaults shift

OpenAI says developers moving to GPT-6.2 Astra or Sol should review their system prompts. The company published copy-ready templates in the models' documentation, and the Responses API now discards earlier reasoning content if conversation history is edited.

model release

OpenAI releases GPT-Live-1.1 in the API, with standby mode billing idle WebRTC time at $0.02 per minute

OpenAI has released GPT-Live-1.1 in the API. The full-duplex voice model improves multilingual accents and transcription, and a new standby mode bills idle WebRTC time at $0.02 per minute instead of $0.05.

Comments

Loading...