A researcher says modern neural networks are far more interpretable than believed, with most neurons understandable in models up to trillions of parameters (Nick Cammarata/Goodfire)

Cammarata argues that ordinary, mostly fixable issues made early neural networks look harder to decipher than they were.

The Goodfire researcher says apparent polysemanticity—neurons responding to seemingly unrelated concepts—was misattributed to superposition. He credits modern agentic tools and closer inspection with changing his earlier pessimistic view. Neel Nanda sees useful evidence for investigating specific model behaviors, but disputes claims that the work amounts to major progress toward fully reverse engineering neural networks.

Techmeme's note