The goal is to obtain (at least some) human curation with the lowest possible friction.

What is learning? How do we test it? Just memorization? In/ex-cludes building of abstractions (see above)? How important is learning, or is just following some obscure rules,with no understanding (what's that) good enough?

Or for the CS challenged, build a massive parallel system without locks, you will need that for “all”. For a little more advancement let it do it's own decomposition, which means it should be able to LEARN math. If you think we are born with some magical math algorithm, please explain one-two-many cultures.

After that we can talk about Information and Intelligence.

For example, in my blog I use an extension called Zemanta, which offers various kinds of automatic linking and tagging. I accept less than %10 of what it offers me, but those %10 add more structure to my blog posts.

To rephrase: I hope that CS steers as far and clear of today's social sciences as possible. The idea that my categorizations carry some universal meaning (or any meaning for that matter) is entirely false. You might say, we'll mod what you say by your reputation (determined by pagerank / number of tweets / whatever). But this doesn't solve anything apart from identifying whether my arbitrary categorization is shared by (or is amusing to) a wide variety of folks. Complex scenarios that will inevitably arise (cross-referencing, context, circularity, etc etc) shall strip it of any remaining meaning.

Ex-ante, in this classification, I believe, is the only promising path. Unless you have a full, reliable, and conclusive model of “me”, you will run into falsehoods that will make your analysis an easy target for the first buddying statistician.

Disclosure 1: I'm a social scientist, with an interest and some background in CS.
Disclosure 2: <Partially-a-lie> Disclosure 1 </Partially-a-lie>

Along with Clay Shirky and Duncan Watts – he is a pioneer in this field. The Wealth of Networks is the closest thing I have to a bible.

http://www.amazon.com/Wealth-Networks-Production-Transforms-Markets/dp/0300110561

(Which is why, incidentally, racial profiling at airports is a bad idea.)

There is a saying among intelligence operatives that the dissemination of information is the _square_ of the number of people who know that piece of operation. The United States Intelligence Community includes 16 agencies, who together employ tens (hundreds?) of thousands of people. Let's guesstimate that out of these 5 to 10 thousand have a high security clearance.

Even if this Intelligence Facebook is open only to them, even assuming total imperviousness to “hard” hacking attacks and zero malfeasant users, it will be a matter of weeks until a social engineering hack or just simple negligence ends up with most of its contents posted on Wikileaks or worse (the Al Qaeda Yammer?).

The worst part is, I am certain that such a collective intelligence system would, if it were feasible to keep it safe, reap tremendous benefits in fighting terrorism. Until it is compromised with disastrous consequences.

Sometimes — sometimes — closed is better than open.

As a response to David Semeria, I would frame the collective knowledge system, not as two separate spaces, but as an interaction between user generated data (collective) and a technical algorithm (knowledge) to structure that data in a meaningful way. For example, as Benkler notes, Google is so useful because it has a collective (millions of web designers creating pages with tags and links) and knowledge (a kick-ass ranking algorithm). Wikipedia can be thought of the same way – a collective (millions of people entering words) and knowledge (a kick-ass algorithm – the brain – that can structure words in a meaningful way).

That was an interesting comment about how computer scientists should model themselves after social scientists such as economists and sociologists. As a professor in the social sciences (and an entrepreneur who is trying to leverage the open source framework), I would replace economists with biologists. I think much of the work you're talking is being done by a variety of disciplines (but not by economists), with the main intellectual anchor being networks (for example you mention Duncan Watts).

As a side note, I was speaking with Robert Frank, the Cornell economist, and he has an interesting prediction about who economists in 100 years will say was the most influential. It won't be Smith, Keynes, Milton, etc. but Darwin. So maybe economists will embrace biologists but not soon.

Again, a fascinating post.

PS I'm not sure if you knew, but the CIA has been using an internal Wiki to share knowledge and foster collaboration – http://bit.ly/6nMe4z. From what I've heard, it's been pretty successful. I'm not sure if they're taking the next step and harnessing (or mining) that information via some algorithm to find patterns, connections, etc.

I'm struck by how many mid- to large-sized companies barely make use of what's already available. It seems like only a small fraction of companies I'm familiar with make effective use of search, forums, social equity systems, or expert location tools. There are notable exceptions – typically the same oft-cited ones in books on social media – but there remains a huge opportunity for us to greatly improve productivity implementing what we already have.

Debates on machine learning always tend to veer off into philosophical abstractions – what does mean mean? etc – and so it's always important to stay practical (I'm not referring to you here, just in general).

I really liked Caterina Fake's post on a similar theme from a few days back. I think we both agree that, at least in the short term, progress will come from the innovative application of existing and not particularly ground-breaking techniques.

In other words, advances in UI & UX design which reduce the friction of adding semantic sugar on input. That kind thing: simple tech, but still hard to get right.

Resources will funnel towards increased research of machine learning without always requiring supervision, as long as there's a stepping stone or ladder that gives intermediate results (payoffs/benefits). I just don't know if there are rungs between where we're at now, and creating self learning systems.

I'm research/designing one of these for opinion and debate, it's also going to be a social game.

Ex-post techniques try to create order from disorder (Google). Ex ante systems (at a stretch Wikipedia) try to provide order beforehand (in reality, Wikipedia pages are semantically messy, contain large blocks of untagged text, and lack consistency between themselves).

Ex-ante comes with a significant user burden – who wants to semantically tag everything they write/do? – but, like all human curation methods, is good at reducing ambiguity. Ex-post is generally painless for the user, but shifts all the semantic burden onto machines (something they're not very good at, and perhaps never will be).

In summary, I would bet on ex-ante systems showing the most future promise…

In any case, I would say that ex-post is actually still the wave of the future…users just don’t want to ‘learn’ something for others benefits…to me the key is in the input systems (so that you can get ex-ante without the user nec. knowing that’s what they are doing).

As an aside, I think Twitter is a perfect example of ex-post right now…thanks to their API tons of ‘collective intelligence’ systems are being attempted on top of the massive (and fairly unstructured) data…(disclaimer: my own project http://wow.ly is slowly, and methodically, building towards being a collective knowledge system for social data)

As we funnel resources into semantic processing businesses we can expect to see improvements and algorithmic refinements based on machine learning of our languages (I’m having fun doing work with Zemanta and Alchemy’s API).

Hope not to offend any religious or anti-religious folks but I did a good bit of reading up on complex adaptive systems when I made an effort to understand the information storage capacity of nucleic DNA last week. The better our computing systems can model these systems, the more capable they will become of learning without our careful guidance, or create skynet .