Researcher: neural networks are far more interpretable than believed

Researcher Nick Cammarata of Goodfire says modern neural networks are far more interpretable than assumed, with most neurons understandable in models up to trillions of parameters. He argues polysemanticity in early models stemmed not from superposition but from several ordinary, mostly fixable issues.

• Most neurons are understandable in models up to trillions of parameters
• Polysemanticity in early models was not caused by superposition
• The causes were several ordinary, mostly fixable issues