Deb Raji
AI accountability, audits & eval. Keen on participation & practical outcomes. CS PhDing @UCBerkeley.
- Interesting - I agree that simple proxies should not be the only way to get to predictable model outcomes. Tho I worry this criteria won't be well operationalized (ie "aligned"?). My pet theory is that model predictability is likely a byproduct of training *data* properties more than anything else.
- Interviewed for this doc years ago & have yet to see the final cut. Most of what I recall is how much Daniel had truly riled himself up - the confusion & chaos of this exaggerated boogyman version of "AI" had excited a genuine emotional response, despite successfully disguising AI's real terrors.
- It made me wonder: why are the real world harms perpetrated by AI today not enough for some people to feel urgency? to take action? to care? What he was hearing from corporate execs & "doomers" was scaring him but the real world issues me & Karen brought up didn't seem to have the same effect..
- Me talking through AI definitions actually came from a longer exchange where I was trying to convince him that the real world issues we see today (& w past "AI" tech) are already worth taking action on, especially since it affects the poor, PoC, ie. vulnerable populations that don't look like him
-
View full threadAI can cause harm even if it doesn't seem to affect *you*, even if those harms aren't observable or felt in any direct way by *your* future children. Being protected from certain harms doesn't make them less real or less important to address. Educating the public requires taking this broader view.
