OpenAI’s Jakub Pachocki proposes perplexity as a measure to limit AI development
OpenAI’s chief scientist Jakub Pachocki proposed a new framework for regulating AI development in a blog post on November 22, suggesting that perplexity could serve as a key measurement for setting limits on AI progress.
The urgency to regulate AI development has intensified, particularly after a former researcher from OpenAI and Anthropic warned that the companies were ‘gambling with our lives.’ This alarming sentiment was echoed in Anthropic’s IPO filing, which cautioned that AI technology could pose ‘catastrophic or existential risks to society.’ In light of these concerns, Anthropic CEO Dario Amodei outlined a plan to ‘pace the frontier,’ a proposal that has garnered support from notable figures such as OpenAI CEO Sam Altman and SpaceX CEO Elon Musk.
Pachocki’s blog post, which Altman described as ‘the most important thing OpenAI has published in a while,’ emphasizes the necessity of a standardized measurement to gauge the speed of AI development. He argues that existing benchmarks, which focus on AI’s capabilities, quickly become outdated as the technology evolves. Instead, he advocates for perplexity, a metric that measures how well a model predicts the next token, with lower values indicating better predictions.
According to Pachocki, perplexity has remained relevant even as AI models have significantly advanced. In his analysis, Pachocki illustrated the decline in perplexity of OpenAI’s models since the release of GPT-3.5, noting that the latest model, GPT-4, achieved a perplexity of approximately 0.91. This improvement has enabled the model to excel in complex reasoning tasks, leading to advancements in fields such as cybersecurity and mathematics.
As perplexity continues to drop below 0.8, Pachocki envisions it as a shared benchmark for general intelligence, allowing AI companies to agree on limits for their development. By establishing perplexity as a standard measurement, he proposes a collaborative commitment among AI firms to refrain from exceeding certain thresholds until they reach a more comprehensive understanding of AI behavior.
This framework could foster a more responsible approach to AI development, ensuring that safety standards are met before further advancements are made. As the conversation around AI regulation continues to evolve, the adoption of such metrics could play a crucial role in shaping the future of the technology and its impact on society.
Why it matters: The proposal from OpenAI's Jakub Pachocki could significantly influence how AI companies approach development and safety standards. By establishing perplexity as a common metric, it may lead to a more unified effort in regulating AI advancements. Watch for any agreements among major AI firms regarding the adoption of this measurement.
Original reporting: TechCrunch AI