OpenAI's chief scientist says AI is getting close to being as good as a human research intern

OpenAI is making significant strides towards achieving one of its ambitious objectives: developing AI systems that can emulate the capabilities of human research interns. During a recent episode of the "Unsupervised Learning" podcast, Jakub Pachocki, the chief scientist at OpenAI, highlighted the rapid advancements in coding tools, mathematical research, and physics. He indicated that AI is on a path to manage increasingly intricate, multi-step technical tasks with diminished human input. Pachocki emphasized that a critical metric for this progress is the duration for which an AI model can operate autonomously. He elaborated on the distinction between a research intern and a fully automated researcher, stating that the extent of time an AI can function independently is pivotal to measuring advancement. In a corporate presentation last month, he outlined OpenAI's ambitious timeline, aiming to create an "AI research intern" by September 2026, with the goal of achieving a fully autonomous AI researcher by March 2028. In a follow-up post on X, OpenAI's CEO Sam Altman acknowledged the inherent challenges, stating that the endeavor might not succeed, but transparency regarding the potential implications is essential. Pachocki noted that OpenAI is already witnessing rapid progress in relevant tasks, citing coding agents like Codex as instrumental in handling much of the company's programming responsibilities. He also referred to math benchmarks as a guiding principle for enhancing model reasoning, given their verifiability. "We've witnessed tremendous growth in coding tools," Pachocki remarked, adding that programming has evolved significantly for most individuals. He pointed out that the immediate challenge lies in advancing systems to handle specific technical tasks with greater autonomy, utilizing more computational power, and maintaining longer working durations. Despite the optimistic outlook, Pachocki was clear that AI is not yet equipped to fully operate at the level of a complete researcher. "I don't anticipate we will have systems capable of autonomously deciding to enhance their model capabilities or solve alignment issues this year," he stated.