Magic Tricks Created Using Artificial Intelligence For the First Time

An anonymous reader writes Researchers working on artificial intelligence at Queen Mary University of London have taught a computer to create magic tricks. The researchers gave a computer program the outline of how a magic jigsaw puzzle and a mind reading card trick work, as well the results of experiments into how humans understand magic tricks, and the system created completely new variants on those tricks which can be delivered by a magician.

Re:AI? They taught it.

[mugetsu37]

Who taught you that? :p

Re:AI? They taught it.

[cogeek]

Are you not intelligent then or did your parents, teachers, society not teach you anything, you just popped out knowing it all? Intelligence doesn't mean knowing everything, it's the ability to learn and to expand on concepts.

Re:

[jones_supa]

If we look at the article, the computer was taught only the specific algorithms to create a jigsaw puzzle arrangement or shuffle a deck of cards. Then the program just ran the data through it to create various optimal results. It didn't have capabilities to expand the concept of the trick, for example.

Re:

[AK Marc]

Is single-task optimization learning? It's not learning "new" things, but answering different paths to the same solution. It's just that the path becomes magic trick, which makes it seem smarter.

Re:

[Anonymous Coward]

I see what you're saying: if it's learned, the intelligence is bog-standard intelligence, not really artificial. If it's not learned, it's just an algorithm, thus not really intelligence. Bravo.

Re:

[i kan reed]

Right, because your ability to read and write English comes from zero active training. It's allllllll magic.

Re:

[ArcadeMan]

[The] ability to read and write English comes from zero active training.

Looking at the way kids write these days, I'd have to agree.

Re:

[The Grim Reefer]

[The] ability to read and write English comes from zero active training.

Looking at the way kids write these days, I'd have to agree.

tl;dr

Re:

[cogeek]

Nope, Skynet

Next trick

[sinij]

Next trick - Earth disappearing into a black hole at the bottom of the magician's hat.

Maybe we shouldn't ask AI things like these?

Re:

[i kan reed]

"Ask" is a bit generous. Almost every AI thing like this is the same:
1. Devise a language to describe the broad kind of thing you want(in this case board elements, and instructions to the tricked person).
2. Show the AI some working examples
3. Show the AI some non-working examples
4. Let the inferred characteristics come up.

Every time AI comes up on slashdot, this kind of magical thinking comes up, where an AI becomes capable of the extraordinary after accomplishing something specific and ordinary.

All th

Re: Next trick

[bombman]

Indeed, AI is a positively abused term. I wonder if they will manage to construct an AI that
can determine if some research is really AI or just statistics on some specific subject.

Re:Next trick

[phantomfive]

After the AI winter, the field has really gone downhill.

Since professors still need to publish, they created a distinction between 'strong AI' and 'weak AI.' For some people, this was fine and yielded useful algorithms (but not AI), but largely it's a way to get published without doing anything substantial. Like this study, for example.

Re:

[phantomfive]

"Strong AI" for all intents and purposes, is a silly concept.

It's the only AI that's actually AI. "Weak AI" is a way of saying, "we know we're not actually creating intelligent machines, but we like the name so we'll keep it." If people don't understand the algorithm, you can tell them it's AI and they'll believe it.

Re:

[i kan reed]

That's bullshit though. Complex pattern recognition, especially when scaled up to useful levels, is intelligence. There's more to humanity than intelligence, and every theory that alleges AI needs to be human to be intelligent just ignores how much "living thing" and "social animal" and "great ape" and "tool user" and "visual thinker" there is to our innate character that has little to do with intelligence.

Those characteristics have given us lots of useful attributes that aren't explicitly about our abstr

Re:

[phantomfive]

Complex pattern recognition, especially when scaled up to useful levels, is intelligence.

How are you defining intelligence?

Re:

[i kan reed]

The same way IQ tests and most GI tests do. Is that too damn crazy?

Re:

[phantomfive]

The same way IQ tests and most GI tests do. Is that too damn crazy?

ok, so how well does your image recognition algorithm do on an IQ test? Do you think it would help to add more graphics cards?

Re:

[i kan reed]

Honestly? I haven't tried.

If we took the visual items from an IQ test, and fed them into an image recognition algorithm, it might actually be able to pull the right results more often than chance.

Re:

[phantomfive]

Honestly? I haven't tried.

I know. :)

An image matching algorithm would fail hard on an IQ test. Especially when the proctor begins reading word problems to you. Even Watson, being a giant search engine, would have trouble on an IQ test.

It would be interesting to try to build an AI that could pass an IQ test, though. Your suggestion of using such a test to measure the intelligence of a computer is a good one. Your defense of weak AI sucked, though. :)

Re:

[i kan reed]

Oh, look, let's unmatch our inputs from our system's inputs!

Hey, look, phantomfive is really really really stupid when I read him his IQ test in pitches only dogs can hear.

Re:

[phantomfive]

Changing the input isn't going to help the image recognition algorithm be smart. Can you post something that shows you understand that, or are you in over your head on this topic?

Re:

[i kan reed]

No, but spacial pattern matching is exactly what this sort of thing [gstatic.com] is testing. And if you fed each of the parent segments into an image recognition algorithm, I bet it would pull out the right final result for at least some similar questions.

This is a hypothesis that requires testing, but it's still the case I'm making.

Re:

[phantomfive]

IQ tests are more involved than that, though it could be part of an IQ test

Re:

[i kan reed]

I specifically said to extract the visual components way back. But I can understand that we were talking past each other until this point.

Re:

[phantomfive]

But can you accept that you were wrong? I suspect that's a rare occurrence for you, sir.

Re:

[i kan reed]

So let me get this straight:

You want me to ignore the context of the things I was saying, and then acknowledge I was wrong in that lack of context?

No.

Fuck you asshole.

Re:

[phantomfive]

That's what I thought.

No, what I want you to do is stop pretending you understand things when actually you don't.

Because that's when learning and growth happens.

Re:

[i kan reed]

No, I'm quite ready to admit I'm wrong, when it's outside the context of dipshits willfully misunderstanding things.

You on the other hand can be completely and totally ignorant of even the conversation you're currently having and demand others admit you're right.

Fuck off and grow up.

Re:

[narcc]

Let me guess: You're a singularity nut?

Re:

[i kan reed]

Absolutely not.

Re:

[thedonger]

Are you an artificial intelligence unleashed on /. by researchers for the purpose of saying positive things about artificial intelligence researchers on /.?

Re:

[AK Marc]

So all the programs that beat captchas are AI?

Re:

[BarbaraHudson]

How about we start by defining non-artificial intelligence. Is that the same as natural intelligence? Can we even define intelligence? To do so, we'd also have to define free will (since decision-making strictly by algorithms has no free will, and no innate intelligence).

Re:

[i kan reed]

It's a bit like defining any other relatively common, complex idea. Any definition you come up with will miss some important details, until it's ballooned in complexity to be as complex as the field you're trying to "help" in the first place.

For a simpler case, imagine defining "car", and getting involved in an intense debate about the relevance of the internal combustion engine's history before you're done.

Re:

[AK Marc]

Can we even define intelligence?

Yes, to anyone but those who would ask such a question. One could presume they'd reject any and all answers that could be given here.

Re:

[AK Marc]

AI is any computer program that can write a smarter version of itself. At least, that's the requirement in most sci-fi.

Jetson!!!

[paiute]

Why am I picturing Uniblab saying "Pick a card. Any card."

Middle School

[i kan reed]

I remember doing this middle school. I was always baffled when other students couldn't grasp what was happening the moment I came to say "Now subtract your original number" or something similarly back-referential and how it stole their choices from them.

They weren't hard to come up with, especially once your framework was established.

Re:

[TheCarp]

Actually a website that does this tricked someone I know recently. I was actually engaged in a card game when they came up to me exclaiming this website could do math with the numbers in her head, and it worked every time.

It took me about 20 seconds to figure out what was going on, and even despite suggesting "why don't you try again, write out each step" and then "try it again with X for your number, and write out each step", still more than 20 minutes to get them to see what was going on.

Re:

[AK Marc]

AI seems more intelligent in the presence of natural stupidity.

OOOOHHH GAAAAWWWDD.....

[Thud457]

Now skynet can BORE us to death!
Thanks Science.

more like theorem proving

[Champaklal]

it's a variant of theorem proving and relation finding exercise in prolog / horn's clause

Computers are easy to program to do math

[Russ1642]

It's not especially difficult to get a computer to do some math. Get a computer to shuffle a deck of cards and I'll be a thousand times more impressed.

Re:

[smooth wombat]

You mean like Shuffle Master [wikipedia.org]?

Not tricks, illusions.

[SirMasterboy]

A trick is something a whore does for money...

Great plot

[Thanshin]

AI is made to invent magic tricks.
AI starts creating more and more complex magic tricks.
Magician stops understanding the tricks but keeps following the given steps and is as surprised as the audience about the result.
After a while, the AI starts giving really strange steps and it becomes clear that there is no explanation in current science that justifies the results of the tricks.
Humanity has meddled with incomprehensible forces, awakening He who was never dead.

Re:Great plot

[David_Hart]

AI is made to invent magic tricks.
AI starts creating more and more complex magic tricks.
Magician stops understanding the tricks but keeps following the given steps and is as surprised as the audience about the result.
After a while, the AI starts giving really strange steps and it becomes clear that there is no explanation in current science that justifies the results of the tricks.
Humanity has meddled with incomprehensible forces, awakening He who was never dead.

When the "AI" can invent magic tricks outside of the basic programming, then I'll be scared.

Basically, they programmed in one trick and then programmed it to compute more variations of the trick. Not much different than programming a computer to fill out a matrix based on the calculations for a single square.

We'll know that we have a true AI when it can go from calculating new card tricks to counting cards in Vegas.....

Re:

[tekrat]

Counting cards in Vegas is already terribly simple. Robots would never be allowed in a casino.

Re:

[AthanasiusKircher]

Counting cards in Vegas is already terribly simple. Robots would never be allowed in a casino.

Yes, but the point is when you have something truly intelligent, it should be able to learn completely new things. Hence, when we have an AI which was only programmed with unrelated knowledge (like card tricks), but it can figure out the necessary -- even simple -- ideas to count cards in a casino simply due to its supposedly "intelligent" algorithms, it would have demonstrated true adaptability, I.e. intelligence.

Genetic algorithms already there..

[xtal]

http://www.damninteresting.com... [damninteresting.com]

There's some magic tricks outside of the basic programming for you. Paper is linked in there or easily searchable, and is quite interesting.

And yes, there might be reason to start to get concerned. Disruptive changes happen quickly. Eventually, we will be the ones disrupted..

Worrying

[kooky45]

I just watched Colossus: The Forbin Project, and lets not teach AI anything more please!

Re:

[WrongMonkey]

Why? If an AI can be created that can do a better job running the world than human governments do, why shouldn't we let it? And honestly, doing a better job than human run governments isn't setting that high of a bar.

Re: Worrying

[Tobe Nemo]

What is something went wrong and you wanted a human the wheels again, and the AI had different ideas? You already gave away all control.

Gno Wonder!

[some old guy]

When Bristlebane hears what the gnomes have been up to he's gonna be pissed. Clockwork magicians indeed!

Whose idea IS this?

[argStyopa]

The first thing you want to teach an AI is "how to trick humans"?

Is that really smart?

To automate J. Gruber

[Tablizer]

Politicians want to use it to trick us more often than just Iraq and ACA.

Re:

[wonkey_monkey]

It's a completely new definition of the word "variant."

Re:

[BarbaraHudson]

To much hyperbole. A variant is, by definition, not something "completely new".

It's a completely new definition of the word "variant."

It's a variant definition of the word "variant".

Which brings up the question - when was the last time you saw something "completely new"?

Careful, one step closer to the AI singularity

[Tablizer]

"Damn! The pod bay doors have been open the whole time. Very clever Hal, but I'm still gonna yank your chips......hey, where did the chips go?"

teach machines to trick humans

[Tobe Nemo]

I can see many things wrong with the idea. Are you a machine? No. I believe you. Are you going to go all self aware and kill us? No. I believe you. At least teach them to be bad liars please!