1. It seems that lazy functional programming languages (like Haskell) may provide a basis for a serious improvement in more robust natural language processing. A survey paper: [http://cs.uwindsor.ca/~richard/PUBLICATIONS/NLI_LFP_SURVEY_D...]
2. Semi-Human Instinctive AI, a new dynamic, nondeterministic decision-making process, seems to be the new hotness in robotics/learning algorithms. In it, a given agent is given a set of basic behaviors ("instincts") that it hones with both open and closed learning methods in a problem space. [http://en.wikipedia.org/wiki/Semi_Human_Instinctive_Artifici...]
3. Anatoly Shalyto's Automata-based programming, using finite state machines to describe program behavior, seems to have a lot of potential. It attempts to view programs from the context of engineering control theory, which opens the door to the use of powerful techniques from dynamical systems in mathematics.
If it catches on, architecture classes might get higher precedence in curricula. Moreover, it might unify (to some extent) the theoretical background around hardware and software.
Look at this diagram: http://en.wikipedia.org/wiki/File:Linking-Open-Data-diagram_... All these datasets have already been interlinked and are available for you to use. This is the linked open data approach (http://en.wikipedia.org/wiki/Linked_Data) The opposite approach is to use data from a single already-interlinked source through an unified API, exemplified by Freebase (http://freebase.com), which is more straightforward but perhaps offers less control. I've found these resources invaluable in more than one project that I'm working on, and every hacker should at least keep abreast of what is available so that you can use it if you need to.
I tried to dig into it, looking for some data and to see what you are talking about, and I finally find a piece of RDF, real semantic web stuff: http://dbtune.org:3030/sparql/?query=describe%20%3Chttp://db...
Um, ok. Now what? This is short XML file containing links, half of which are dead. The biggest problems with SW is that no one agreed on the labels, inputs and outputs, and that there are no mechanisms for data preservation or trust.
How have those been solved now?
(edit) I'm not hating on the idea, btw. It just doesn't seem to be a technological problem. It's a social one. The second you find a way to get people to structure their data for fun & profit, the SW will blossom. And then it will be spammed. And then someone will find a way to index it and filter out the spam, and by then it will be something good, but quite different from what was intended.
I am genuinely curious to know what has changed in the last few years that academics now take SW for granted.
This is absolutely not the case. Among the Semantic Web community, the SW might be taken for granted (to some extent), but among the wider CS community, there is a lot of scepticism about whether the SW is feasible. Most of the people I know in the academic database research community dismiss the Semantic Web, for example.
http://books.nips.cc/nips21.html
There were several papers near applied areas like text classification, breaking audio captchas, and even brain machine interfacing. However, even the theoretical papers usually come with examples (e.g. image classification) that show optimistic results. If you were doing any learning task that is definitely the place to find the state of the art.
It's starting now, and it's starting the same way the web started - working poorly, very fragmented, cool but not yet practical. This will change soon.
Anyway, the BlueWaters project will be done by 2011. It will be the fastest computer then.
Distributed computing just means computers getting faster. There is no killer app for this. Yes, you may say cancer research or flight simulations or so on, but there are not the next big things - not the way the web was.
Admittedly, the concepts involved are dated since PRAM theory (http://en.wikipedia.org/wiki/Parallel_Random_Access_Machine) dates to the 70's. However, this project marks the first successful commitment of PRAM theory to silicon
