WolframAlpha: A Magical Knowledge Engine

[ExactlySolved]

The Planck units are definitely bugged, and notice that \sqrt{2 \pi} = 2.50663, so this may have to do with inconsistent conventions on a Fourier transform.

 

[Moonbear]

It's not a mind reading machine. I guess the biggest challenge WA faces is to distill their algorithms to figure out what the users intended to ask from what they asked.

More importantly, they need to find a way to determine which sources to use. It doesn't seem to discriminate between credible and erroneous sources when finding information for answers. That seems even worse than google or wikipedia, which at least cite the source they're using, so the user can determine if they are credible sources.

 

[ExactlySolved]

Moonbear: most of the cite omissions I have seen so far are about mathematical topics, could you give some examples?

 

[Redbelly98]

How was it hyped? It's hardly on any major news outlets. And if it was the article got back up. Even Scientific American doesn't mention it, perhaps not yet.

The only way I found out about it is because of this thread.

There was a feature on it on NPR radio last Friday. Not on a par with Fox News or the major television networks, but nonetheless they were putting the word out.

I say give it a couple of months to see if the bugs get worked out.

 

[Coin]

I can't get the site to respond to any query. I get long pauses and then various "unavailable" error messages.

When I first heard of this site a couple weeks back, and heard that the backend engine is based on Mathematica, one of my first responses was "can Mathematica actually scale to serve this much complex data to a google-like website?". My question stands :|

 

[CRGreathouse]

So you need a reference for 1 + 1?

http://us.metamath.org/mpeuni/1p1e2.html

 

[CRGreathouse]

ExactlySolved, are you a Wolfram employee?

I asked it to "convert 630 nm to THz." It figured out that's it's dealing with wavelength, and frequency and so found a link in its database and computed the conversion.

That's pretty impressive, since those are dimensionally incompatible. (check site) It even says they're incompatible and gives its assumptions: "(assuming speed of sound at sea level and 15 deg C ~~ 340.3 m/s)". Nice! I see that it doesn't let you change those, at least not in the obvious way: link.

 

[chroot]

ButI'm disappointed because I feel that most of the task is still on the user to provide the appropriate question that WA can answer.

The same could be said of college professors, could it not?

It appears to be an insanely useful application, even if it is not yet the end-all be-all of artificial intelligence. I'm an integrated circuit designer, and I've used it twice today already, saving me probably an hour of work each time... and it's free! I'm pretty damned pleased!

- Warren

 

[Coin]

I finally get the site to respond. Now that it seems to be working again it is impressively speedy even for computationally intensive inputs.

It seems very limited in that it doesn't seem to be able to do anything that the creators didn't specifically think of ahead of time. I can type in "u.s. unemployment rate" and it finds that? If I ask for the "u.s. underemployment rate" it doesn't know what that is. If I ask for the "u.s. unemployment rate over the last four years" or any alternate phrasing it doesn't know what that means. If I ask it for "dow jones 1900-2000" it can give me a graph of that, because it apparently occurred to them to program in the Dow Jones as something for which data exists over a range of time? But it doesn't know that is the case for the unemployment rate, or the "u.s. population". If I type in "u.s. population 1900-2000" it seems to think I want to multiply the u.s. population by 1900, then subtract 2000.

If you stumble across something the creators had in mind when they designed the site, it is fantastic. I've been just typing arbitrary cellular automata rules into it and it produces these beautiful summary plots of the cellular automata rule applied in various ways, and it reveals the Mathematica statements that produced each plot. Of course, like with the u.s. population thing, I hit a dead end if I want to view the data in a different way than they anticipated. Like, okay, I can get it to generate the output of rule 110 and plot it, but what if I want to take that output and instead use it as the input to some other kind of function? I guess I buy Mathematica?

Hopefully they will be logging the queries people type into it on this first day and use this information to refine the kinds of data their system contains and the way it is fetched. In the meanwhile though it seems more likely that this will be useful as a quickie web frontend to Mathematica than it ever will be as a general collection of knowledge.

Will the engine to the site ever become a part of Mathematica itself, such that people can set up wolfram-alpha-like natural-language web interfaces to Mathematica scripts or knowledge bases they created themselves?

 

[f95toli]

The database is rather odd. I entered the name of the small town where I grew up (north of Sweden) and it returned the location, population and weather (quite impressive).

But when I enter the name of the place where I live now (large suburb of greater London) it does not recognize it and instead suggests a small town in Alabama with a not-so-similar name...

 

[AUMathTutor]

Seeing as how it's the 18th and nothing is better, I can say with certainty that I'm not impressed. The only thing is it seems a bit speedier now, but that may just be my connection.

*sigh*

 

[qntty]

I found this to be rather amusing.

 

[ExactlySolved]

ExactlySolved, are you a Wolfram employee?

No, I'm an academic physicist. I use Mathematica everyday, and although I rely on the mathematical algorithms, I agree that other packages e.g. MATLAB and maple compare in this area (in fact I sometimes use MATLAB at the university for the excellent curve fitting toolbox).

The thing that I really admire about mathematica, that makes me go on and on about this product, is its structure as a programming language. By structure I mean the kinds of things that are done in an upper division undergraduate CS class "Concepts/Organization of Programming Languages" e.g. a book like this:

https://www.amazon.com/dp/0321330250/?tag=pfamazon01-20

Mathematica is a totally functional language like lisp or Haskell, which means that every expression has the uniform structure of a nested set of functions, and a program consists of a composition a functions applied to input to obtain an output. Mathematica's internal evaluations are ultimately rule-based pattern replacements of these functional expressions, a system which the user has direct access to. Everything in mathematica is an expression, even input and output (arbitrary graphics are expressions, not side effects, even the notebooks themselves are expressions).

So that's it, organization and consistency are why I like Mathematica.

I found this to be rather amusing.

The reason that sum^(-1.5) fails the ratio test, which is a bug, is (I'm guessing) that there is mathematica code which compares 1. === 1 and yields false: the use of a machine number exponent leads to the 1. , and the ratio test fails so the sum is reported as diverging. Using n^(-3/2) works, but sum n^(-2.) fails, I'm pretty sure that the comparison of a machine integer and an exact number is the reason.

Will the engine to the site ever become a part of Mathematica itself, such that people can set up wolfram-alpha-like natural-language web interfaces to Mathematica scripts or knowledge bases they created themselves?

Yes, W|A will become part of Mathematica, using internet functionality to reason live results from W|A inside a notebook. Also relevant to your question are the soon to be released W|A APIs, best suited for Mathematica itself but also availible for other major web platforms. I would not suggest, however, being overly optimistic about 'openess' of these APIs.

Also I share your concerns about Mathematica scaling performance-wise to this load. The only reason I still use languages like C for some tasks is speed, and I'll readily admit that Mathematica is 10-1000 times slower, having an interpretive kernel. Hopefully computer performance will increase enough overthe next few years to realize more ambitous uses of W|A.

 

[CRGreathouse]

No, I'm an academic physicist.

OK, just wondering.

The reason that sum^(-1.5) fails the ratio test, which is a bug, is (I'm guessing) that there is mathematica code which compares 1. === 1 and yields false: the use of a machine number exponent leads to the 1. , and the ratio test fails so the sum is reported as diverging. Using n^(-3/2) works, but sum n^(-2.) fails, I'm pretty sure that the comparison of a machine integer and an exact number is the reason.

That seems odd to me, since it's not on the boundary. If it failed with sum n^-1.001 that would be different, but -2.?

Also, shouldn't it be using < rather than ===?

 

[AUMathTutor]

Some of the errors are just silly, as in, if you asked the question to the average person with a background in science/engineering, they would not make the same mistakes or have the same trouble.

I thought the point of W|A was to keep you from having to learn a command-line syntax and from considering machine representation of numbers in calculations.

Rather than a product that can create new knowledge, I see a product that can answer petty questions about existing knowledge.

 

[dx]

Rather than a product that can create new knowledge, I see a product that can answer petty questions about existing knowledge.

W|A is not meant to create new knowledge. It is just an interface to Mathematica connected to a large database of computable knowledge.

 

[AUMathTutor]

What in the world do you guys mean by "computable knowledge"?

Most knowledge is computable in the modern sense of the word. If you mean "mathematical", or perhaps more accurately, "approximate mathematical" knowledge, then I feel like W|A is a better product with an unambiguously misleading name.

Of course I don't mean "new knowledge" as in completely unknown facts. What I mean is a novel synthesis of the facts which links ideas in ways that perhaps people haven't already thought of.

This happens with simple genetic algorithms; they find optimal solutions and strategies for solving problems where human beings do not. An analysis of the answers after the fact can reveal the "why" of it, but I think you get what I'm saying.

I was expecting (or, really, hoping for) this:

<< Q >>
Why are plants green?

<< A >>
Plants contain chlorophyl, and chlorophyl is green.
Would you like to know why plants contain chlorophyl?
Would you like to know why cholorophyl is green?
etc.

 

[Borek]

Interesting information. Someone I know from other place meet with Wolfram 18 months ago and they were talking about chemistry part of the W|A. This is what he told about his impressions:

Anyway, it was clear that Wolfram Research is very Mathematica-centric, and there are *no* chemists in the place. I was working with a physicist who had been at NIST, and they all thought him to be a chemist! He wasn't.

I discussed with Stephen the issue of complexity in chemistry, but the response always seemed to be "we can compute that."

Ignroance can be forgiven, self-conceitedness can't. When calculating pH W|A makes mistakes that HS students should not. And yes, you "can compute that", but you have to know how. Then, there are things that we can't compute yet.

 

[AUMathTutor]

They have a lot of work to do before I use the thing for anything other than entertainment.

We could make a drinking game where we ask W|A a question, and we drink if it (a) has no information on the topic, (b) provides meaningless unrelated information about some random fact, or (c) provides a clearly incorrect result to the inquiry.

I have a feeling that would work out pretty well.

 

[dx]

What in the world do you guys mean by "computable knowledge"?

Basically anything that that Mathematica can do something with.

 

[Moonbear]

So, is it really just a web-based version of Mathematica? If so, what's the big deal about it? It's sounding like it might be a useful thing within a very limited area...something for math specialists perhaps? People seem to be promoting it as the next big challenge to Google or a replacement for Google, and I don't see it as even in the same ballpark, let along a competitor. It seems to do something entirely different and for a much more limited set of information and queries.

Still, even if it's just good for a limited range of information, it really would be useful to have it tell you what its sources are. How does it decide which sources to use? How is it supposed to employ any sort of AI if there is no way for it to get feedback and "learn?" For example, when it gives a wrong answer to something, there is no way that I see to flag it and tell it that it's wrong and should skip that source and look for another.

 

[dx]

So, is it really just a web-based version of Mathematica? If so, what's the big deal about it? It's sounding like it might be a useful thing within a very limited area...something for math specialists perhaps?

It's not just a web-based version of Mathematica. The difference is that W|A has access to a very large amount of data, and that it can understand common language (instead of specialized syntax). For example, since it knows the dates of birth and death of Richard Feynman and Albert Einstein, you can ask it a question like "How old was Richard Feynman when Albert Einstein died?", and it will be able to answer (36 years 11 months 7 days). The answer to this question was not known to W|A, but it was able to compute it using the information in its database. This is obviously a very unimaginative example, but you can see the potential capability of W|A given how powerful Mathematica is (it can do far more than addition and subtraction, which was all it needed to do in this example), and how much data it has access to.

 

[CRGreathouse]

Another oddity: it can recognize (not (P xor Q)) or (R and T) and it uses the text form "implies" in its description, but it can't understand (P xor Q) implies (R and T) or (P xor Q) -> (R and T).

The answer to this question was not known to W|A, but it was able to compute it using the information in its database. This is obviously a very unimaginative example, but you can see the potential capability of W|A given how powerful Mathematica is (it can do far more than addition and subtraction, which was all it needed to do in this example), and how much data it has access to.

I see many nice hand-crafted pages, like comparing (LOCATION 1) to (LOCATION 2). Sites providing services like this already exist (for many various templates, not just this one), but having them together in one place and with one interface is nice. I also see a Mathematica interface, or something like it. This is a nice improvement on integrals.com. Finally, I see a souped-up unit converter that solves more problems but makes more mistakes than Google calculator (which I'm told is basically a front-end for the 'units' utility).

I'd like to see more examples of things that go beyond these, that help me see the "potential capability". Right now it's a slick package but I can't see myself using it regularly.

 

[chroot]

Yeah, it seems to be essentially a very fast "robotic librarian." That's not really a disparaging remark, though. It's still a pretty awesome creation.

- Warren

 

[Coin]

Still, even if it's just good for a limited range of information, it really would be useful to have it tell you what its sources are. How does it decide which sources to use?

There's a leeetle bitty "source information" link at the bottom of the page. It's easy to miss and the results are poorly organized (it gives you a list of sources used to build the page, but there's no way to determine what source any specific piece of information came from).

For example, when it gives a wrong answer to something, there is no way that I see to flag it and tell it that it's wrong and should skip that source and look for another.

There's also a tiny "feedback" form at the bottom of the page. I don't know where the stuff put in the feedback form goes, or if anyone reads it. I haven't used it yet-- I noticed a couple of specific things that needed correcting when I was first messing with it but then forgot what they were before I noticed the feedback form...

Something I am concluding is that the W|A tools have the potential to be useful and very cool, but before this can happen what it really needs is some sort of "wiki" like feature for user-supplied data. The Wolfram employees are simply never going to be able to cover everything that needs to be covered. On the other hand their basic tool set seems surprisingly powerful. In the hands of a wider set of specialists it could be used to provide some very cool services. So I am encouraged by ExactlySolved's comments that the W|A engine will be built into future versions of Mathematica in some form.

Last night I was having a discussion with some coworkers about the California budget issues, and specifically the question of whether the California educational system spends too much money for the level of service it provides. We dug up some spreadsheets the Federal government publishes ranking the states' educational systems in terms of things like budget size or standardized test performance. I wound up spending some time extracting and merging some of that data into another spreadsheet, and in the end wound up with nothing more for my efforts than a single scatter plot of basic standardized testing success vs. amount spent per student. Which was interesting, but there were any number of other things I'd have liked to have compared the data against-- population size? State GDP? Percentage of ESL students?-- but it was just too time-consuming with the tools I had at hand. The entire time though I was thinking, everything I'm doing here by hand is stuff I've basically seen Wolfram|Alpha demonstrate it's powerful enough to do automatically-- store tabular data, cross-reference it, plot it based on natural language queries-- if only it had any idea what a state budget is. But I could easily imagine a world where users are allowed to add information to Alpha's data store, or where it were possible for me to install a copy of the base Alpha engine on my own web server, and with probably about the same amount of effort massaging those spreadsheets into whatever format Alpha uses internally as it took for me to fight past iWork's interface bugs last night I could imagine winding up with something able to plot things like "population vs education budget for all u.s. states" based on just a query-- except that it's missing a couple of vital plot types, this seems realistic based on what I've seen Alpha do so far. And once those spreadsheets had been Alpha-ized, that tool would be around for anyone who might want to investigate similar data later.

So there does seem to be a lot of potential here, but it all depends on what they do with it and especially how open they ultimately make the technology...

 

[OAQfirst]

Groklaw has described a legal issue:

The legalese says that they claim copyright on each results page and require attribution.

http://www.groklaw.net/article.php?story=20090518204959409

 

[Coin]

OAQfirst, wow, that's a bit odd. Especially considering they claim copyright on the plots, but as far as I can tell those plots are just being generated via normal Mathematica. So you type something into Alpha, it spits out a plot, that plot is copyrighted, you copy the Mathematica command Alpha used into Mathematica, the plot is no longer copyrighted? Then again maybe getting you to buy a copy of Mathematica is the point...

 

[Moonbear]

Groklaw has described a legal issue:

http://www.groklaw.net/article.php?story=20090518204959409

Even if they obtained the information from another source?

Oh, and thanks Coin for pointing out the feedback and sources links. They need to make them more prominent.

 

[OAQfirst]

If Groklaw's interpretation of those terms are correct, them Wolfram might be infringing copyright law just by returning some of its results if anything it returns is already copyright-claimed by another party. Still, I don't know how enforceable this is. I don't read copyright law (United States) as usually covering this range of content.

EDIT: Ah, Moonbear- I guess we were on the same page with that.

 

[ExactlySolved]

The last few versions of Mathematica have gone outside the niche of 'math specialists' to provide really useful tools for all kinds of physical and social scientists. They call this 'integrated data' and what it means is that with very simple commands you can import data on astronomy, genetics, geography, economics, weather, etc from Wolfram's internet servers.

So there does seem to be a lot of potential here, but it all depends on what they do with it and especially how open they ultimately make the technology...

It is safe to say that your example about state education budgets etc is 'in the works' for W|A, not only will non-wolfram entities be able to add data, but public data like you describe will also eventually be added by Wolfram Inc itself.

Even if they obtained the information from another source?

It's not that the data that's being copyrighted, it is the text and images that W|A produces --- just as you most likely hold the copyright for any text or diagrams in any scientific publications you've authored.