Agentic AI Could Shift Interpretability Research Toward External Control
Commentary on X
Highest ranked
Shubhendu Trivedi@_onionesqueTECH
Mech interp also had an auxochrome-like intensifying effect on that tendency. The name sounds great, but it (MI) tends to favour sharp (often misleading) narratives that border on exegesis. A better ethos for the field's claimed goals would be IMO more akin to control engineering
Shubhendu Trivedi@_onionesqueTECH
premature back then, when externally legible artifacts were primarily about "inside the model" rather than "around the model." With the agentic paradigm, it might finally start to coalesce. But the stranglehold of the "chart inside the machine" is still too strong.