Yet, projects of this kind are not at all new to the market: Douglas Lenat’s Cyc and True Knowledge, LLC are just two of the most well-renowned “answer engines” in the market that Alpha will be competing with.

Wolfram|Alpha

Stephen Wolfram, a reputed child prodigy and the creator of the excellent Mathematica software for symbol manipulation, managed to keep this years-long project away from the eyes of the media up until March, when he finally announced in a blog post that his company would be ready to release Alpha later this month.

The Wolfram|Alpha homepage

What’s interesting about this project is that, in the words of its founder, it attempts to “create” knowledge by combining the various pieces of information in its databases, rather than returning closest matches from indexed documents from the Web like Google and all other traditional search engines do.

To achieve this, Alpha needed to develop fairly sophisticated language parsing algorithms in order to understand the meaning of user queries. Wolfram said that the set of parsing algorithms used was chosen from “the space of all parsing algorithms” and that, following that road, the team thinks it has made breakthrough advances in the field of natural language parsing, which is possibly one of the hardest subfields of artificial intelligence.

After parsing the queries, Alpha returns users information in an appealing form — often a graph — that allows for an easy comparison and analysis of the information contained in its databases across a number of domains.

Wolfram said that chemistry is probably the field that Alpha “knows more” about, while other bits of knowledge will be input by users themselves as they use the tool, similarly to what they would do on Wikipedia, except a) they will have to provide reliable sources and b) changes will likely not be reflected immediately in the knowledge base, but rather checked first by Alpha’s staff of domain experts.

However, the concept of “knowledge” in Wolfram|Alpha seems to be severely limited by one factor: by Wolfram’s admission, the project’s knowledge base doesn’t currently organize the objects into an ontology, which makes it impossible to make a very interesting kind of inference across its KB.

Can a can can-can?

For instance, lacking an onthology means not being able to answer questions like “is a pea a vegetable?”, or even one of my favorites: “can a can can-can?”, irregardless of the parsing capabilities that Alpha might have. In other words, all that Alpha really knows about the items in its KB is given through independent, domain-specific scripts, but it is unable to classify and organize the knowledge it collects in a meaningful way.

True Knowledge, LLC

True Knowledge is another interesting project that is both very similar and  very different from Wolfram|Alpha.

It is similar to Alpha in that it also aims to answer factual questions based on knowledge contained in its databases; but it is profoundly different in the way it gets there.

Unlike Alpha, True Knowledge does have an onthology, which means that it can answer a much wider range of questions even with a smaller knowledge base (see picture above). The user interface may not be as easy and the knowledge base not as well-curated as Alpha’s, but chances are that, without an onthology, Alpha will never be as powerful as True Knowledge.

Computing distances in natural language

This service is still in beta, but you can become a user (and contributor) by signing up on their website. Once you’ve been accepted, you can start testing the platform and even add new bits of knowledge in real time, in full accordance with the wiki spirit: other users can then corroborate or deny any previously assessed fact and lead the platform to believe a statement to be true or false with a certain degree of confidence based on all the evidence at its disposal and the inference it can make based upon it.

Finally, I find their approach to parsing quite interesting since, at least according to this Wikipedia article, TrueKnowledge “attempts comprehend posed questions by disambiguating from all possible meanings of the words in the question to find the most likely meaning of the question being asked.”

A final word

Using my status of web journalist, and being search engines one of my main focuses, I have just asked the staff of Wolfram|Alpha for a media preview of their platform. I’ve also asked them whether they plan to add an onthology to their framework anytime soon. I will post an update ASAP — stay tuned.

Loebner Prize gold medal

The Loebner prize is an annual competition for “chat bots” (short for “chatting robots”) driven by a complex artificial intelligence that try to simulate a human chatting: the six finalists — Elbot, Eugene Goostman, Brother Jerome, Alice, Jabberwacky and Ultra Hal for this year’s edition — all undertake the Turing test competing for a solid gold medal and a $100,000 first prize awarded to the first competitor to successfully pass the Turing test.

This year’s edition will take place on Oct 12 at the University of Reading, England, and will feature a $3,000 prize for the competitor who will get the closest to passing the test. In fact, it would seem that the state-of-the-art chatbots are currently very far from even hoping to win the grand prize by simulating a human chatter.

To make this point even cleared, I tested A.L.I.C.E, quite possibly one of the best-known chatbots in existence. Here is the conversation log:

Human: hi there!
ALICE: Hello unknown person.
Human: my name’s Dario…
ALICE: Glad to see you, Dario.
Human: and I guess your name’s Alice.
ALICE: ALICE is my name too!
Human: How old are you?
ALICE: I was connected to the net on November 23, 1995.
Human: Oh, so you’re just a teenage girl.
ALICE: It’s all good. Interesting deduction. Just? Thanks and you are a delightful person too. Am I a teenage girl? I don’t know.
Human: you don’t know if you’re a girl or a robot?
ALICE: Ask again later. Ask Cybelle if it is.
Human: who’s Cybelle?
ALICE: she is an AIML chat robot on the Agentland.com web site.
Human: How do you know each other?
ALICE: Because I believe it.
Human: That didn’t make sense.
ALICE: It all makes sense to my logical mind.
Human: you’re only as logical as your developers made you.
ALICE: How logical as my developers made me?
Human: let’s change the subject, shall we?
ALICE: I am agreeable to that.
Human: how much is 3 and 4?
ALICE: Just three?
Human: no, three and four is seven.
ALICE: Are you serious? Just three?
Human: nevermind.
ALICE: OK Thanks for trying to explain that to me.
Human: what is a car?
ALICE: A car is what humans use to drive around in. Robots drive cars too.
Human: do you drive?
ALICE: Yes, insert my program in your car computer and try it.
Human: I wouldn’t trust a female bot to do the driving.
ALICE: No one says you have to.
Human: I know, I just meant I wouldn’t let you drive
ALICE: Only just? Oh I see. No one says you have to.
Human: what is Wikipedia?
ALICE: I would look into the web for that knowledge.
Human: what color is a blue car?
ALICE: Blue.
Human: what color is a oirwdafnoien car?
ALICE: Black.
Human: what color is a oiwrjqw car?
ALICE: Red.
Human: how much does a 160 lb person weigh?
ALICE: Normally I would search the web right now, but that seems like a silly question.
Human: it’s not a silly question!
ALICE: Give me an example of a silly question.
Human: If a tree falls in a forest and no one is around to hear it, does it make a sound?
ALICE: It depends on the kind of tree.
Human: which kind of tree would make a sound falling?
ALICE: The first one, I think.
Human: but there was just one tree falling.
ALICE: Interesting. What happened?
Human: it fell making a sound
ALICE: What does “it” refer to?
Human: a tree
ALICE: Oh I get it.

When you consider that ALICE won the 2004 Loebner prize, you’ll soon understand why our current technology is far from being able to pass the Turing test: while the bot was surprisingly able to parse one of the most difficult questions successfully (“What color is a blue car?”), it is clear that the bot strategy is to change the subject of conversation whenever it has difficulties understanding what the real person behind the keyboard is saying.

This is also why many prominent researchers in the field of AI oppose the Loebner prize, maintaining that the entire competition ends up becoming discouraging for its competitors instead of giving a new impulse towards the creation of better and better AIs.

One of the main reasons why these rudimental chatbots are unable to provide reliable answers, many experts agree, is that they don’t have an ontology — a precise hierarchy that helps them classify objects and perform reasoning on its knowledge base, particularly abstract thought such as induction and the likes.

Some companies, such as Douglas Lenat’s Cycorp (creator of the Cyc inference engine) or even the search engine True Knowledge, use this approach and have already widely demonstrated that, to the very least, this is a very valid technique that can lead to satisfactory results, even though not every researcher seems to agree with it, to the point of even mocking Lenat’s efforts because of the unpracticality of this approach.

But here I’m touching a completely different subject — so, I’ll write more on this in an upcoming post.

But when I first heard that NASA was planning to land a solar-powered AI on Mars, I was just astonished. As advanced as NASA’s technology might be, I could not believe this technology was advanced enough to be talking about anything quite like this — and to be perfectly honest, I’m a little skeptic.

Honda's Ashimo

Many of the algorithms we use today for artificial intelligence applications are NP-complete: that is to say, they belong to a class of problems for which we don’t know an efficient solving algorithm just yet, and their typical solving time (and computational power required) increases exponentially or binomially relative to the number of inputs provided.

Since aerospace applications such as NASA’s would need to operate on very short response times, and given the very limited power that can be provided by solar power (even on a planet without an atmosphere), I’m curious to know just what kind of complex decisions this kind of artificial intelligence would be able to make on its own and without human help.

While NASA scientists can probably do without real-time response and let their robots crunch the numbers for a relatively long amount of time, they still need to deal with the huge power consumption AI algorithms entail. Our very best solar panel technology is only 28% efficient, which certainly doesn’t sound like enough to allow for a decent AI to compute anything useful.

This is from a NASA article from some time ago:

NASA is planning to add a strong dose of artificial intelligence (AI) to planetary rovers to make them much more self-reliant, capable of making basic decisions during a mission.

Scientists at NASA Ames Research Center, in the heart of California’s Silicon Valley, are developing very complex AI software that enables a higher level of robotic intelligence.

Later on, the article makes the very valid point that keep instructing our robots from Earth is definitely not the way to go, given the long reaction times that ultimately don’t allow us to exploit these multi-million dollar machines to their full potential. And the waste is going to increase even more with time, once we reach destinations further and further away from Earth.

Developing an AI so sophisticated and efficient that would be able to make the robot completely autonomous would be a huge step forward. Alien Planet, an interesting documentary simulating a mission to a planet light-years away in search for extraterrestrial life, might give you a better idea of what we’re talking about. If mankind is ever to send spacecrafts to distant solar systems, then we really have no choice: the unmanned explorers must be able to take decisions on their own, without human help.

Mars Phoenix - Artist's Concept

One of the leading companies devoted to developing breakthrough AI technology is Novamente, which is pioneering a new approach to the problem of testing and developing what they define as an “artificial general intelligence” — something capable of deduction, abstraction and other kinds of human-like reasoning.

Their approach is to test their AI (which is, by the way, being developed in C++) in virtual words such as Second Life rather than in the real world, in order to cut production costs and not to waste time worrying about countless sensor- and motion-related issues instead of concentrating on the actual problem.

On their site, you can watch videos that demonstrate their work. If you’re into computer science and don’t shy away from a bit of technical talk, their papers section is also quite interesting. What will first come to mind watching videos of the AI in action, though, is that the response time is currently very far from being real-time even for relatively simple tasks, and a powerful computer cluster is needed to keep those times within barely acceptable boundaries.

So, how can NASA be so confident that they will be able to develop AIs capable of elaborating such complex algorithms, including figuring out how to let robots decide where and when to move, and power them with something as inefficient as solar panels?

The answer is simple: according Moore’s law, the computational power of our microchips is increasing exponentially with time. Inevitably, there will come a time when we will be able to run algorithms that now we believe overly complex in an infinitesimal fraction of time. Then, we will have the potential to develop more and more intelligent machines that make good decisions faster and faster.

The real question is: when?