3 ms·

I can believe that they're doing some form of extrapolation to create novel solutions to posed problems You can believe it what sort of evidence are you using

I can believe that they're doing some form of extrapolation to create novel solutions to posed problems You can believe it what sort of evidence are you using for this belief? Edit: Also, the abstract of the Apple paper hardly says "corruption" (implying something tricky), it says that they changed the initial numerical values

famouswaffles 2y ago
Changing numerical values doesn't do anything to impact the performance of state of the art models (4o, o1-mini, preview) The only thing that does is the benchmark that introduces "seemingly relevant but ultimately irrelevant information"