370
Productivity & Workflow357
Software Development252
Automation & Workflow230
Marketing & Growth212
AI Infrastructure & MLOps196
Photography & Imaging170
Writing & Content Creation196
Design & Creative176
Data & Analytics162
Customer Support161
Education & Learning150
Sales & Outreach131
Voice & Speech131
Operations & Admin117
OpenAI’s Jakub Pachocki says AI companies may need a shared score called perplexity to set practical limits on how fast AI improves.
In short: OpenAI’s chief scientist says AI companies could use a shared score called perplexity to help set limits on how quickly AI systems improve.
OpenAI chief scientist Jakub Pachocki published a blog post about ways to slow or limit AI development. The post focuses on a measurement called “perplexity,” which tracks how well an AI model predicts the next piece of text.
Perplexity is a technical term, but the basic idea is simple. It is like a spelling and grammar “predictability” score for text, where lower numbers mean the model is better at guessing what comes next. Pachocki argues that this kind of score stays useful even when other popular tests become outdated.
Pachocki’s post comes as more people debate whether AI companies should agree to “speed limits” for AI progress. Anthropic CEO Dario Amodei has also called for common safety standards and limits on unchecked AI progress, and the wider debate has been fueled by warnings from former researchers and risk disclosures in public filings.
Pachocki included a chart showing perplexity falling in OpenAI models since GPT-3.5. He wrote that GPT-4, around 0.91, was linked to stronger step by step reasoning, and that below 0.8 OpenAI has seen new capabilities appear in areas like cybersecurity and mathematics.
A “speed limit” only works if everyone agrees on the speedometer. If major AI companies used a shared measurement like perplexity, they could, in theory, agree not to push past a certain level until they can better understand and monitor what these systems do. That could affect how quickly new AI features reach the public, and how safety checks are handled along the way.
Source: TechCrunch AI