To give a sense if scale: It’s not that o3 failed to solve that red blue rectangle problem once: o3 spent thousands of gpu hours putting out text about that problem, creating by my math about a million pages of text, and did not find the answer anywhere in those pages. For other problems it did find the answer around the million page mark, as at the ~$3000 per problem spend setting the score was still slowly creeping up.
If the trajectory of the past two years is any guide, things that can be done at great compute expense now will rapidly become possible for a fraction of the cost.
