I didn’t quite get Wolfram|Alpha when Stephen Wolfram wrote about it. Nova Spivack has a long article about Wolfram|Alpha following a demo that sheds a lot more light.

In a nutshell, Wolfram|Alpha is not a search engine, nor a knowledge base, but a system that computes the responses to natural language questions. As I understand it, there are basic building blocks which will use used to compute responses. It would also appear that the goal is to answer questions around formal knowledge, in other words there is probably a lot of pre-computing that will have to happen behind the scenes. Now, you, me and everyone else on the planet is skeptical. Wolfram is mad smart, but NKS, which forms part of the basis for Wolfram|Alpha, wasn’t exactly a resounding success, but this particular titbit was quite interesting.

But even more amazing than the product itself, is that he’s been able to keep this project secret for so long. I say this because it is a monumental effort (and achievement) and almost absurdly ambitious. The project involves more than a hundred people working in stealth to create a vast system of reusable, computable knowledge, from terabytes of raw data, statistics, algorithms, data feeds, and expertise. But he appears to have done it.

So it’s not really natural language processing. It actually reminds me of protein structure prediction, where we use existing knowledge about protein fold space and local protein structure to try and predict the structure of new sequences. The NLP part seems to be the front end to a set of objective functions which need to be satisfied to answer the question asked, essentially transforming natural language into machine language. As Nova writes

This is almost more of an engineering accomplishment than a scientific one — Wolfram has essentially broken down the set of factual questions we might ask, and the computations and data necessary for answering them, into basic building blocks — a kind of language if you will. Then, with these building blocks in hand his system is able to compute with them — to break down questions into the basic building blocks and computations necessary to answer them, and then to actually compute the answers.

Of course, like protein structure prediction, we are limited by current knowledge. I don’t believe Wolfram|Alpha can possibly do pure ab initio computation to give us an answer to novel questions. If it can do so generally, then we can all sit and marvel at the greatest achievement of our times.

Given the ambition and past efforts in this space which have failed, I remain skeptical, but also hopeful. If this works, it’ll be amazing. I want it to work. Don’t think it’s going to replace search engines (not all knowledge is formal) and it’s not going to retrieve useful documents, but it will become part and parcel of our knowledge. Looking forward to public launch, and then there will either be joy or supreme disappointment