United StatesA 26,000-word post reopened AI interpretability.
Goodfire researcher Nick Cammarata says language models may be easier to monitor and control than assumed.
Cammarata says Google’s 2014 Inception v1 has a car-and-cat neuron that tracks a V shape, not two unrelated things.
Jacob Hilton of the Alignment Research Center says researchers understand less than 0.01% of top models’ inner workings.
His team has not yet published technical results on modern language models but plans to soon.
How each outlet framed it
- MIT Technology Review
- tests Cammarata's 26,000-word claim against skeptics, framing it as a possible safety breakthrough that still needs verification
Sources: MIT Technology Review