Specialised Human Information Processing: The Case of Scoble and the FeedReader

I can't remember what I was researching, but I came across a post on Tim Ferriss' blog (fourhourworkweek) with an interview with Scoble last May titled How Scoble Reads 622 RSS Feeds Each Morning. What I found interesting was the way Scoble described the cognition involved. It seemed obvious and familiar.

I cannot remember the specifics, although generally he described the process in terms of iterative scans of an item to extract information of interest. A variety of different filters were used such as title key words, image in the content, and author reputation. The iterations were used as discriminators for interesting pieces, where attention and subsequent knowledge extraction and acquisition are treated as a commodity (naturally, 622 feeds is a lot). The first pass is raw pattern recognition: does the piece look interesting? Importantly, discrimination and acquisition are intertwined, such that as the granularity of the discrimination process progresses so does the granularity of extracted and acquired information, such that if a piece is truly interesting it is extracted and acquired in its entirely (I presume).

I'm not sure about the level of recall available after such a process, although Scoble does point out the multiple objective outcomes. Specifically, in tracking memes (for example technology he is interested in), tracking the personal details of high-reputation individuals (CEO's), and ultimately tracking broader industry trends. Scoble points out that the application of the acquired information is seeming ad hoc and passive, remaining latent until called upon in a specific circumstance (How's the kid? What do you think of X?). Active application occurs at the time of acquisition, where he in turn produces a feed of human-discriminated pieces of interest.

There is nothing special about the Scoble case other than my use as it as a catalyst for considering the general case. I proceed with a similar algorithm daily and have done for many years, therefore I'm sure many perform such a ritual. Further, I use a specialised version of the case when intently researching a technical report or paper, applying levels of discriminators toward productively educating myself on a space and achieving my goal (bi-objective at a minimum). In this latter case, the algorithm provides a ruthless top-down information acquisition process which I have come to rely on.

The process as captured in systems occurs along similar lines, using various indicators (signals) and indicators in aggregate until sufficient information is acquired to make a decision. What I have not seen, is an effective equivalent for the human-first pass "does it look interesting?". This is something humans are very good at although poorly understand, specifically the process of holistic estimation/approximation toward snap decision making.

How would you automate such a first-pass to the process? I presume a vision-based templating system, accumulated over time from experience reading interesting and non-interesting articles (inductive). I also presume that it exists in scales between context specific and non-specific decision making (specific-feed to all/any feed). How is it different from the aggregate effect of per-indicator recrimination? I'm not sure, but my gut tells me that it is (scientific method, bah). It is approximate, learned (not instinct), and typically wrong, but right enough that it is given support (you trust your own judgment). In this case, thinking over Scoble's comments makes me think of things like visual composition (hence the templating approach).

I think abstracting and modeling this process would be fun. The layering of complexity reminds me of Brook's subsumption architecture, and the emergent complexity would be exciting (if it did something useful).