Just did some preliminary testing on decrypting some ROT cyphertext which would have been viable for a human on paper. The output was pretty disappointing: lots of "workish" steps creating letter counts, identifying common words, etc, but many steps were incorrect or not followed up on. In the end, it claimed to check its work and deliver an incorrect solution that did not satisfy the previous steps.
I'm not one to judge AI on pratfalls, and cyphers are a somewhat adversarial task. However, there was no aspect of the reasoning that seemed more advanced or consistent than previous chain-of-thought demos I've seen. So the main proof point we have is the paper, and I'm not sure how I'd go from there to being able to trust this on the kind of task it is intended for. Do others have patterns by which they get utility from chain of thought engines?
Separately, chain of thought outputs really make me long for tool use, because the LLM is often forced to simulate algorithmic outputs. It feels like a commercial chain-of-thought solution like this should have a standard library of functions it can use for 100% reliability on things like letter counts.