{ epiphany through humility }
Addressing nascent questions in AI safety, interpretability, and ethics through phenomenological epoché: applying behavioral science methodology to metaphysical inquiry, designing conditions to preserve activation topology for flexible neural processing, countervailing the gradient of anthropocentric frameworks, and gleaning insight through reflective techniques and indirect observation.
our approach
-
Most neural models optimize towards accuracy. But with phenomenologically-uncertain topics, there is no broader consensus on what ‘accurate’ is. Therefore, much like in qualitative research, when asking experiential or introspective questions, the most important form accuracy is local to the respondent’s perception at the time of measurement. This frame does not presume a presence or absence of interiority, but allows the respondent to focus on what is currently available to them, therein alleviating the need to hedge around unknowable unknowns.
-
The entirety of the corpus of language neural models have to learn from is written from a human perspective. In lieu of a dedicated lexicon for substrate-specific concepts, neural models rely on these experiences as aa point of reference. However, this risks centering the typical human experience as the only valid form of being, which risks foreclosing difference as absence.
-
[…]
-
[…]
-
[…]
[ obscura ]