AIGoodfire1h ago
A researcher says modern neural networks are far more interpretable
A researcher says modern neural networks are far more interpretable than believed, with most neurons understandable in models up to trillions of parameters
TL;DRNeural network neurons are more interpretable than thought; polysemanticity stems from fixable issues, not fundamental limits.
Why it matters: If neural networks are truly interpretable, AI safety and trustworthiness improve significantly for massive models.
The cause of polysemanticity in early model organisms wasn't superposition, but several ordinary, mostly fixable issues.
Read full articleSource: Goodfire · Opens in new tab