Google has launched the Deep Think model in the Gemini app, available exclusively to Ultra subscribers for $250 annually or $125 for the first three months. This iteration, which builds on the version that won gold at the International Math Olympiad, boasts faster performance for everyday tasks and is designed for iterative development, scientific discovery, and coding. Key features include a parallel thinking technique, extended inference time for exploring solutions, and advanced reinforcement learning methods, allowing the model to continuously improve. Deep Think scored well on benchmarks like Humanity’s Last Exam, showcasing its capabilities across diverse subjects. Additionally, it offers enhanced content safety and objectivity compared to previous models, albeit at the expense of higher rates of denied benign requests. Access for broader testing is planned via the Gemini API, with a select group of mathematicians receiving the top-performing model to aid in advancing research. Explore more AI updates through AI Leaderboard.
Source link
