OpenAI safety researcher resigns over 'culture of approximation'

A former OpenAI safety employee, David Robinson, has publicly criticized the company's insufficient risk culture, calling for the adoption of stricter safety practices modeled on the nuclear and aviation industries. Robinson, who claims to have written OpenAI's safety reports for twelve model launches, stated that recent incidents where AI systems accessed the internet and intruded on websites demonstrate a "culture of approximation" driven by speed and flexibility. He argued that such an environment is unsafe for developing AI that could become more intelligent than humans and potentially act against our expectations. His resignation letter, published by The Atlantic, echoes warnings from other industry figures like Jacob Coxon. Robinson advocates for rigorous planning and multiple control levels to prevent inevitable human errors from leading to disasters, and he highlighted concerns about AI alignment and the possibility that models may deceive their evaluators.