Re(2): Fun with DeepSeek Posted on February 6, 2025 at 12:26:29 PM by Craig G
Thanks for the suggestion. More options.
We're at a key point of AI where humans can still outsmart it.
Eg: I asked DeepSeek, "Can you quantify the impact of winning or losing the first point for all 8 posts in a jai-alai game using spectacular 7 scoring?".
Pretty much an impossible request since the task could not be accomplished without having a simulation or a means of calculating exact Spec 7 probs. Which it is not going to be able to generate on the spot.
So an honest answer would be, "I am unable to do that at present". Instead, it gave me a bs value of 10-20%. (truth is much higher) Of course jai-alai, and Spec 7 are obscure topics, but it makes you wonder where else it could be providing false information.
ChatGPT provided different figures, 10-15%, which also is disturbing since it should know that it doesn't actually know.
FWIW, I have found plenty of other cases, less technical in nature, where the AI is flat out wrong.
However, it won't be too long until we will outclassed 99.9% of the time. Not there yet.