Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[nomoetoe]

I got home from a trip and I saw that a massive tree branch and fallen over my porch and by my car...it had hit the car.

not even one scratch to the car. xD

[Xantalos]

While we're waiting for wAIfunet to happen, I'm going to write a fanfic where the Adeptus Mechanicus unwittingly discover what they assume to be an ultraheretical AI STC, only to discover it is something far more terrifying and awe inspiring to behold: DF V1.0

Doitdoitdoit

[Darvi]

Read the proposals for SCP-001. It is surprisingly cool. One of them is fucking hilarious at the end.

Somebody wrote up follow-up tales to those proposals. I suggest you start with cygnus.

[KingofstarrySkies]

Read the proposals for SCP-001. It is surprisingly cool. One of them is fucking hilarious at the end.

Which one?

Database.

[Sergarr]

excuse me but DF2014 has been released

[alway]

http://en.wikipedia.org/wiki/Arimaa

For those who don't want read, its a game specifically designed so that a 4 year old could understand the rules, but a computer would have trouble winning.  There have been yearly human versus computer championships with a $10,000 for the programmer of the first computer to win, for 9 years and counting, and the prize has never been claimed.

Computers are stupid.  Believe it or not, I already knew about self driving cars and Watson, and I am still not impressed.

In particular the self-driving cars use a combination of pre-built, extremely detailed maps of the area and sensors which track distance to other cars.  They have no concept of the space around them, they ARE following a script, which is "follow this line, adjust speed based on what your sensors say".  The problem with this as an AI is that it has no concept of what any of what it does actually means.  It brings to mind a magazine article I read on one of the earlier attempts to make self-driving cars, which was based solely on cameras for sensory data.  This attempt involved having the computer record camera data from a human driver driving around, then having the computer take over.  It drove flawlessly until it reached a bridge and then it swerved off to the side and forced the human driver to take control and prevent an accident.

The reason?  It believed that keeping the color green (the grass) to the immediate right of the car was the most important thing.

To go further with it, a human looks at a room and sees walls and floors and tables.  This is because a human divides the world into objects and knows what each of those objects is.  They have nothing but the sensory perceptions of light, but they construct an entire imaginary world with three dimensional space around them, and this (as far as any human can know) matches up with the actual real space around them, which cannot be perceived by humans.  A human looks at an area of green with blue above it and divides it up into "this is grass, this is sky".

Meanwhile a computer has nothing but the sensory information.  It can be programmed (as face recognition software does, albeit badly a lot of the time) to recognize specific objects.  But this has to be done on an object by object basis, because the computer isn't actually thinking, its just doing what its programmer told it to.  So sure, you could program it to divide the world that those grey pixels on its video feed are road and the yellow bits are the line it shouldn't cross.  Then you show it at a night and will have no idea, so you have to program it to account for lighting differences.  Then take it off road, or to a dirt road, and it will again need to be reprogrammed.  This is because it will never actually understand the concept of road, because it doesn't think, it just reads from a script.

As for Watson, trivia contests are not the same thing as, say, Starcraft 2.  Building a huge database of information and then accessing it on the fly plays to a strength of computers, which is performing a series of simple tasks thoughtlessly and very, very, very quickly.  Its creating the illusion of thought using brute force and processing speed.  Try to invent a program that takes two arbitrary related facts and reaches an intelligent conclusion using them, I guarantee a supercomputer running said program will be less profound than the average twelve year old.  Because computers are fucking stupid.

Reading from a script is fundamentally how every modern computer program works.  No computer program can go beyond its script.  Sure, you could write a program that edited its own code and the relaunched itself, but that would still be limited by the ways in which it was programmed to rewrite itself.  Computers do not think.  If you believe a computer is thinking, that's because its programmer designed it to imitate a specific human behavior using smoke and mirror tricks.

So none of that's actually true. The reason Watson was created for Jeopardy is specifically because it was a weakness of AI. Jeopardy questions are difficult; much more difficult than common speech. They have puns, wordplay, and very unorthodox sentence structure. It wasn't about answering the questions, it was about understanding them. Computers are good at structured databases. Pulling out information and understanding from plain text, which is what the Watson AI does, is incredibly difficult. It involves learning synonyms, meaning, common sense knowledge; all topics which are very difficult for computers. Starcraft AI, on the other hand, is trivial.

Again, you don't need to 'reprogram the AI' for every new circumstance; nor has that been the case for decades. The 60's era game AI, sure. But ever since the 80's era reign of the messies, with neural networks, genetic algorithms, and so on, that has absolutely not been the case. And since the 90s Victory of the Neats, we can even peek inside the AI's head at what it learned, and learn from it ourselves.

Beyond that, as much as we like to self-congratulatorily talk about how humans are more 'spontaneous' and open to 'improvisation' than AI, that's no longer actually true. And, in fact, growing bodies of evidence show just how severely we overestimated our own ability. These days, AI can out-improvise us.

They basically grew a mind in a rough analogue to human childhood, only grooming her to have the traits desired for her role in society. Once they have the fully grown mind they can copy it into as many bodies as they desire. The newer book features an accountant grown in the same way who has then forked off new instances who take decades of specialist training in similar roles.

Yep; that's pretty much the result of 'messy' AI and embodied cognition. And more or less highlights the main problem with AI created with that paradigm. Without the transparency of the neat approach, it becomes incredibly difficult to create something that actually works. If you make a programming mistake rendering the entire thing ultimately fruitless, it doesn't become apparent until after extensive training attempts. Or if it's a subtle mistake, it may just degrade results with no provable way of showing it's doing sub-optimally.

ALSO, DF2014 RELEASE THE KRAKENS, OPEN THE VAULTS TO THE LOWER BOARDS AND STREAM FORTH TO THE ABOVE-LANDS!

[Aseaheru]

The thing you get as the game starts to make a new world, before the world gen options
It also made me laugh for a while...

[Arcvasti]

I MISSED the DF update. Or at least spent several minutes browsing the forums before noticing it. HAO.

[EnigmaticHat]

-snip-

First of all, I said nothing about improvisation or being spontaneous.  Spontaneity is stupid and easy, even in humans.  I was talking about understanding.

Watson's goal is to interpret questions, OK.  Well let's look at its resources:

$3 million in hardware.
The assumption that each question is a Jeopardy question, which is a VERY specific context
Thousands upon thousands of questions from exactly that same context
Massive portions of the internet, including all of wikipedia, which effectively means it has the combined memory of a million humans
Multiple programmers working for years to get it to do this one thing correctly.

Watson is using brute force.  Its very impressive, but that doesn't change that fact that its merely imitating something humans can do using many many times the resources humans use to do it.

Secondly, I wasn't saying that you had to reprogram the computer each time, I was saying that A even if it learns its going to learn in a way specifically dictated by its programming and B the computer doesn't actually understand what its doing, its just doing what its told.  What I'm saying is that its not a sliding scale from calculator to human, its a sliding scale from calculator to much much much faster calculator, because that's fundamentally all a computer is doing.  Performing simple actions ultra-fast.  You can have those simple actions change the code that dictates which simple actions it does, but that is a fundamentally different beast than actually thinking.

Going to back to game AI, massive application of data is a fundamentally different question than taking the same amount of data, and using it well.  That's why Watson isn't a step forward, because winning a Trivia contest with all of wikipedia at your fingertips is liking plugging an aimbot into an oldstyle shooting gallery game.  Its, if not cheating, winning using inhuman capabilities that have nothing to do with intelligence.

Look, what is one thing all good chess programs do?  They weight the AI towards putting its pieces in specific places.  Why do they do this?  Because the AI does not understand the point of the game.  It doesn't understand anything.  It literally plays by going through every possible combination of move, every possible combination of enemy move, ect. (with some pruning of course) and weighting the outcome based on pre-set value judgements.

A computer cannot think.  It cannot understand.  It can only act, because its a tool.  Its like a hammer, but a billion times more complex.  Now, maybe you could take electronics and metal and build something that does think.  Maybe you could even build a sci-fi sentient machine.  But it would be a fundamentally different beast than modern hardware.

Finally, because it needs to be said an arbitrarily large number of time: computers are fucking stupid.

[Sergarr]

Now go play Dwarf Fortress.

[Worldmaster27]

There's nothing new in the Bug Reports board.

I guess everyone is too busy playing DF.

That or just lazy, like me.

[Lord Shonus]

That's because bug reports have belonged on the bug tracker for several versions now, and the bug board is no longer supposed to be used.

[Worldmaster27]

Oh, yeah. That thing that I probably should have remembered. Whelp. Also explains why there are no recent threads.

[EnigmaticHat]

Now go play Dwarf Fortress.

On it

A while back I said the first thing I was going to do for this release was punch a man in bar and see what happens.  When worldgen completes, this dream will be realized.

[Aseaheru]

There's nothing new in the Bug Reports board.

I guess everyone is too busy playing DF.

That or just lazy, like me.

i posted two things allready...

Dont use bug reports, use the tracker.