Introducing Priming AI - Google

Google DeepMind discovered that teaching a large language model just one new sentence can cause it to behave strangely, like calling human skin "vermilion" or bananas "scarlet."

Their research, using a dataset called Outlandish, showed how rare words with low probability can trigger this spillover effect, known as priming, even after just a few training exposures.

To fix it, they introduced two effective methods—stepping-stone augmentation and ignore-top-k gradient pruning—that reduce AI hallucinations without harming learning.

To grasp the new conxepts, register for this online video session today!