Jakub Pachocki, OpenAI’s chief scientist, says rapid advances in deep learning could produce major AI pretraining breakthroughs as early as 2027. In a post titled “Perplexity and Intelligence,” published on November 22, 2026, he argues for using perplexity, measured as cross-entropy loss per token, as a common metric for comparing models developed by different laboratories. He highlights that a log-perplexity of approximately 0.76 has previously been associated with sudden improvements in solving complex mathematics problems. This finding, derived from internal OpenAI work, suggests that similar thresholds could trigger advances in other capabilities as early as 2027, although the timeline is not guaranteed. Pachocki also calls for limits on further scaling until AI systems can be adequately monitored, while Sam Altman supports the transparency needed to establish common evaluation and safety standards.
Source: Read the original article


