@PackBropagated@prajdabre Optimizers are for LLMs are not finding local minima. If you keep training, the loss will keep going down (slowly). There are too many directions to to lower the loss from any given point.
@PackBropagated@prajdabre Does local minima have anything to do with high dimensional optimization? I thought local minima are essentially practically impossible to find.
@prajdabre They are in different optimization basins, their weights are in wildly different spaces. You either need to start from a single trunk that has been somewhat trained to break the symmetries. Or you can do some intense remapping to put them in the same space. See linear mode con
@pmddomingos@andrewgwils What kinds of errors do people make when they forget NFLs? Can you explain one case where thinking about NFL would have saved someone time/effort?