Wolfram Alpha: First-Hand Impressions

You’d think that, after my less than flattering coverage of the Wolfram Alpha pre-launch hype, they would have blocked my IP address even after the public launch. I certainly thought so.

But I was wrong. Last week, someone from their business development group contacted me to arrange a preview of the system. He spent an hour with me today, discussing the system and showing me how it worked. I couldn’t type into the search box myself, but he did let me suggest queries and then entered them for me. He also encouraged me to speak freely about what I’d seen–they’re actually concerned that the hype and its propagation through the echo chamber are setting inappropriate expectations for the product, and eager for cooler heads to prevail or at least temper the exaggeration.

While I stand by my critique of the marketing, I am persuaded that Wolfram Alpha has built something interesting. It is unfortunate that it’s invited so much comparison to Powerset and is being hyped as a potential Google killer.

As Nova Spivack and others correctly pointed out, Wolfram Alpha isn’t really competing with Google, but rather with Wikipedia and Freebase. Wolfram Alpha offers two major core competencies. The first is a vast amount of curated knowledge: they claim to have roughly 20 trillion “facts” as raw inputs, not the results of any kind of inference process. The second is the capability to relate those facts programmatically by exploiting their formal structure and to generate inferences from them.

Yes, Wolfram Alpha also has a natural language component, but that aspect of their offering is, by their own admission, its weakest point. Having a natural language interface is a necessary evil to provide a way for human beings to access this structured data. But unfortunately its the aspect that people have fixated on most–and will likely continue to do so.

In my view, a more productive way to think of Wolfram Alpha is as a highly structured content repository that offers an API in order to get at the content programatically. I saw example queries like (population of china) + (population of japan) / (population of USA). By itself, that’s a parlor trick that feels a bit like Google Calculator. But the ability to include expressions like is_a (china, country) and population (capital_city (china)) in an application (e.g., Excel) would create real value. Note: this is my fanciful syntax, not theirs.

In May, Wolfram Alpha plans to open up their web site to the public, which means that people will have access to the natural language interface to their content. While the launch will surely generate a lot of press, I suspect that most people will focus on the natural language interface, or on the inability to handle subjective information needs.

That’s a pity. The technology has challenges, but I think the questions people should be asking are if, when and how they will be able to integrate it with business applications. Wolfram Alpha may have value as a stand-alone alternative to Wikipedia for objective data, but its real potential is as a service to use in other applications.

By Daniel Tunkelang

High-Class Consultant.

15 replies on “Wolfram Alpha: First-Hand Impressions”

One point I’d like to mention is the following: I’m not sure if people will like a natural language interface. After all, years of googling taught them to reformulate their information needs as keyword-based queries…or am I simply getting you wrong and queries in this funny query language (reminds me of Prolog) is the way to ask this tool for information?


As things stand, they’ve implemented a natural language interface, so you can enter queries like average rainfall in china. They do some amount of natural language processing to handle imprecise queries. It sometimes works, but it’s brittle. I’ve never seen a general NLP query interface that isn’t brittle.

What I’m suggesting is that they shouldn’t be optimizing for human end-users, but rather for programmatic use by other applications. Think Excel rather than Google.


My tour guide from Wolfram Alpha was quite adamant that it is not. I asked him specifically to compare Wolfram Alpha to Cyc and Freebase, and he said it was somewhere in between. I don’t know enough about the internals of either, but it certainly seems to fit into that general class of tools.

According to Wikipedia, “Doug Lenat estimated the effort to complete Cyc would be 250,000 rules and 350 man-years of effort”. Freebase has been crowd-sourced, which makes it harder to know how much effort has gone into it, or how much more effort would be needed to harmonize it into a representation standardized enough to imagine computing over it.

In contrast, Wolfram Alpha seems to have been assembled from existing structured content repositories, over about 2 years (and surely with far fewer than 175 people).

Let me be clear: I’ve moved from deeply skeptical to cautiously optimistic. That’s a huge shift for me–I don’t usually like eating my words. But I still want to bang hard on it before declaring it a success. I’m on the list for early access. If you think you can convince Wolfram Alpha that they’d want you on that list, let me know, and I’ll happily introduce you to my contact there.


Immediate plan is a freely available web site supported by ads. But they’re looking for potential partners. I don’t think they’ve really thought it through. I do think their most likely road to success is charging businesses for programmatic access via an API–and told them as much–but it was hard to tell how much they’d thought about it. That said, the guy I talked to did have previous experience selling enterprise software.


David, I agree that the expectations have been all over the map, but I think the company itself shoulders some responsibility for how it managed those expectations. I’m glad they’re taking a more sensible approach now, and I think they’ll get a fairer hearing if they tone down the hype.

But I also think the primacy of their natural language interface will undermine their true value proposition. I’m not sure they agree with me on that point, and we’ll see how it plays out.

By the way, you might want to link here from your blog, rather than to the syndicated post on Smart Data Collective.



Thanks! I changed the link.

And yes, I agree Re the expectations management issue. I suspect that the initial teaser was intended to both test the way reactions might initially form themselves and also to attract the attention of interested parties and “movers and shakers” to seed business development.

I also agree that the natural language interface is only one part of the larger technology–and perhaps the smaller portion of what is valuable. There are are clearly multiple layers of value propositions.

The way most people will see this is through the natural language box on the W|A site. And there is an expectation management issue there as well: to use the abbreviated grammar of asking for information on one thing or relationships between several entities (mortgage information for example, or chemical reactions, or ….) .

People will certainly test it with all sorts of absurd inputs to see what works and what doesn’t. And this is the child scientist in all of us that tries to test things by banging on them. But I think that folks will settle down into the site’s grammar–just as they do with Google. So, eventually it will come down (for the general public) whether it fills in a need–perhaps as yet unperceived–or perhaps creates a new landscape.



Comments are closed.