Learning that and how AI errors

35 people are aboard a sinking ship. The ship will take 45 minutes to sink. There is only one lifeboat which can hold 7 people. A round trip to the closest island takes the boat 11 minutes How many people can be saved?

I dropped this on a few LLMs.

I had different answers from Claude, Gemini, ChatGPT, CoPilot, Kimi, Manus, Perplexity and some interesting conversations. Claude and Gemini got it right, all others got it wrong first shot, with some needing longer conversations, with very clear pointers “How does the boat get back to the ship?”. Co-Pilot was very disappointing, ChatGPT needed several follow up questions. Pi spun a great long story around the incorrect result, then tried out several options, ending each with: Wait, that’s not right or No, that’s still not right when checking its answer. Pi got it wrong several times because it was trying to achieve a number that it thought was the right result from previous, so pre-existing knowledge – that was wrong and must have been accepted unquestioned.

Kimi was referring to some online content that was incorrect.

Both Gemini and Claude got it wrong when the same question was posed a few days later.

Leave a Reply

Your email address will not be published. Required fields are marked *