Here is what I believe to be an example of gpt-o1 insisting that its previous statement is correct, while contradicting itself, rather than admitting defeat. I’m interested to know if anyone else has a take on this exchange.
Without posting the entire (long-ish) conversation I had with o1-preview, I’ll just give the highlights.
- I asked for a derivation of the formula for calculating time-weighted returns (used in accounting).
- GPT-o1 responded with the following example calculation:
- I told it where I saw a mistake (holding period boundaries aren’t adjusted for cash flow), and asked it to prove me wrong. It responded in no uncertain terms that I was mistaken (that “your understanding deviates from standard TWRR methodology by creating an inconsistency in portfolio valuations across period boundaries”), but then went on to provide another example that contradicts its first, and validates the very point it is apparently arguing against:
I very well may have made a mistake here myself; I know next to nothing about accounting, and I don’t want to accuse GPT of failing unfairly, so let me know if that’s the case.