OpenAI's internal model bel said to be a generation ahead of Astra
A post claims OpenAI's internal model bel solves roughly 2.5x as many problems at similar compute on a published math test, with its lowest compute setting beating Astra's highest, and training is still ongoing.
openai’s internal model ‘bel’ will feel like a full generation ahead of astra the clearest number so far: roughly 2.5× as many problems solved at similar compute on its published math test. even the internal model’s lowest shown compute setting beats astra’s highest. openai also reports record performance across internal benchmarks, and training is still ongoing. i’d expect that to mean stronger reasoning, better coding and more complicated projects actually getting finished. the math result gives us a sense of the possible jump; how broadly it carries over is still the prediction. https://openai.com/index/navier-stokes-solution/