Google DeepMind's Pranav Shyam and researcher Shubhendu Trivedi argue AI interpretability research relies too heavily on novel visualizations
They expected the field to focus more on functional control.
Story Brief
They expected the field to focus more on functional control.
Source Posts on X
Cluster sources
Shubhendu Trivedi@_onionesqueTECH
It's surprising to me that interpretability research didn't move towards control (tending towards stuff like instrumentation). In 2018 or so, I used to think that that was the natural direction for the field to move towards given how large the models were. But perhaps it was too https://twitter.com/recurseparadox/status/2074684223701889077
Pranav Shyam@recurseparadox