4 articles
OpenAI introduces Random Network Distillation (RND), a prediction-based reinforcement learning method that encourages exploration through curiosity. RND exceeds average human performance on Montezuma's Revenge for the first time.
OpenAI proposes iterated amplification, an AI safety technique enabling specification of complex behaviors by decomposing tasks into simpler sub-tasks, without labeled data or reward functions. Experiments remain limited to simple algorithmic domains.
OpenAI opens applications for its second Scholars cohort: 6–10 stipends and mentorship for underrepresented individuals to study deep learning full-time for 3 months and open-source a project.
OpenAI introduces FFJORD, a reversible generative model using free-form continuous dynamics to learn complex distributions. The method reduces computational complexity and improves scalability compared to prior approaches.