OpenAI's chief scientist proposes establishing unified industry-wide AI safety standards; major breakthroughs in pre-training technology are expected by 2027.
On November 22 local time, OpenAI Chief Scientist Jakub Pachocki published a post proposing that perplexity or cross-entropy loss should be used as a unified metric for measuring progress in AI model pretraining, and that the industry should be pushed to establish common safety evaluation standards and limits on model development. He stated that as models' ability to predict data improves, breakthrough capabilities have already emerged in fields such as cybersecurity and mathematics, and he expects that more transformative capabilities will emerge when models' log perplexity drops to the 0.68 to 0.7 range. Pachocki believes that AI could bring major breakthroughs in pretraining algorithm research as early as 2027. He suggested that frontier model developers should first establish shared safety evaluations, transparency standards, and corresponding technical safeguard mechanisms before further enhancing model capabilities.
Latest
4 m ago

