I suspect respect also needs to be topic-specific. For example. http://represent.berkeley.edu/umati/ http://represent.berkeley.edu/umati/ is a candy machine th

I suspect respect also needs to be topic-specific. For example. http://represent.berkeley.edu/umati/ http://represent.berkeley.edu/umati/ is a candy machine that bribes undergraduates to grade exams. It learns both about the exams, and about the graders. But "grader X is a generally B-ish student" is much less valuable than "grader X does well on questions on topic Q, but is unreliably confused about topic R". People's comments within their area of professional expertise are often more valuable than their comments which aren't. Also, my threshold for "insightful enough to spend time on" varies by subject. And by time of day. Just now, I'd love to be able to ask HN "just give me a couple of really nice, deeply insightful comments, on any topic". But yes. Being able to filter by cohort could finally interrupt the dynamic of quality discussion fleeing from site to site, chased by a regression-to-mean eternal September. Which suggests a need for a rich subcommunity selection language. "Show me HN default, but boost the weight of HN old-timers, plus people they highly respect, transitively with a decay factor of R". So rather than social graph, I'd focus on broader machine learning / statistical analysis and computer-human hybrid processes. For example, say a person is sometimes insightful, but sometimes crackpot. An aggregate metric isn't very helpful. And it's hard for automation to make that call. But it's easier for automation to recognize the problem is an issue for some comment, and to select a couple of humans to give moderation judgment.