“Based on Eric Jang's experience, large language models like Claude Opus 4.6 and 4.7 are effective at hyperparameter optimization and experiment execution but are not yet capable of high-level strategic thinking or identifying research bottlenecks.”