Sounds like there’s the potential of asking it a question and it literally spending hours thinking about it.

Worth keeping in mind that performance on benchmarks seems to scale linearly with log of thinking time (https://openai.com/index/learning-to-reason-with-llms/). Thinking for hours may not provide as much benefit as one might expect. On the other hand, if thinking for hours gets you from not solving the one specific problem instance you care about to solving that instance, it doesn't really matter - its utility for you is a step function.