Which is notable, because GPT-4.5 is one of the largest models ever trained. It's larger than today's production models powering GPT-5. Goes to show that "bad

Which is notable, because GPT-4.5 is one of the largest models ever trained. It's larger than today's production models powering GPT-5. Goes to show that "bad at jokes" is not a fundamental issue of LLMs, and that there are still performance gains from increasing model scale, as expected. But not exactly the same performance gains you get from reasoning or RLVR.