There was a paper or two a few weeks ago about a smaller deep learning model that interpretability researchers had found certain neurons in that behave almost l

There was a paper or two a few weeks ago about a smaller deep learning model that interpretability researchers had found certain neurons in that behave almost like an API. Change the value on those those very small number of neurons and a large cascade effect happens resulting in drastically different output. Super interesting discoveries lately.