Announcement
Goodfire says tiny neuron teams handle arithmetic, ALL CAPS and date recall in LLMs
Goodfire says its unsupervised method finds thousands of small neuron groups that work together on specific tasks.
TLDR
Switch off the small team, change the predictions: Goodfire says disabling a neuron community substantially alters predictions on its task. But disabling everything else in those layers largely preserves predictions, the company reports, provided the community and 17 always-on neurons remain active.
Combined views
—
3 Sources, first seen ago

Comments