Jakub Pachocki, OpenAI’s chief scientist, published a blog post suggesting that perplexity should serve as the standard metric for regulating artificial intelligence development. This proposal follows recent warnings from researchers and executives, including Dario Amodei of Anthropic and Elon Musk, who argued that unchecked progress poses existential risks to society. Pachocki argues that while benchmarks measuring downstream capabilities become obsolete quickly, perplexity remains a stable indicator of model performance. He defines the metric as a calculation of how well a model predicts the next token, noting that lower scores indicate better prediction accuracy. His analysis shows a consistent decline in perplexity scores since the GPT 3.5 era. Pachocki writes that reaching losses below 0.8 has enabled transformative capabilities in cybersecurity and mathematics. Consequently, he proposes using this specific threshold to establish a shared benchmark for general intelligence. Under this framework, AI companies would agree not to exceed a certain perplexity level until their ability to monitor AI behaviour reaches a concrete milestone. This approach aims to create a common safety standard that limits the rate of unchecked progress across the industry.
* Pachocki defines the metric as measuring how well a model predicts the next token on average.
* He notes that reaching losses below 0.8 correlates with transformative capabilities in mathematics.
* The proposal requires companies to stop progress until monitoring abilities achieve a specific milestone.
OpenAI’s chief scientist has thoughts on how to limit AI development
Jakub Pachocki, OpenAI’s chief scientist, published a blog post suggesting that perplexity should serve as the standard metric for regulating artificial intelligence…

Source Read original →



