| |
| |
| |
|
Page: 1 2 3 4 5 6 7 8 9 10
Comments:
<0> <hybrid_@ef> i'm an ai researcher
<1> Ok yes right now :)
<2> kanzure What you need to do for what you are working on is learn higher mathematics. athena mentioned algorithmic information theory
<3> paros: i think you're ***uming that a self-modification means that the agent does not but that... that's not what i mean
<0> <hybrid_@ef> mostly optimisation
<1> Kanzure, are you familiar with Eurisko?
<4> Star-glider, Nope
<1> That's one of the self-modifying AI cl***ics.
<4> `Provides offices, warehouses, residential, and retail locations for sale or rent.`
<4> Oh, wrong Google query I guess :)
<3> star-glider: there are several who claim to have built ai, but they're all brittle chatbots
<3> add "lenat" to that query :)
<4> I think that the first thing that we must do is quantify 'action' and the units of 'behavior' before any sort of behaving engine can be built
<1> Yes, well, that's a narrow AI application.
<2> kanzure There are mechanical ways to map a set of points in a domain to those in a range.
<4> I think I'm going for tonight, spent too much of this day in #ai :)
<4> thanks for your support, Star-glider
<0> <hybrid_@ef> http://www.loebner.net/Prizef/loebner-prize.html
<2> ya know..all programming code does it produce a mapping from input file to output file
<2> (algorithmic information theory 101)
<5> kanzure, the term behavior is not used as anything fundamental in AI. once familiar with simple neural nets and how they learn you get a better insight into what slant to take on AI. without practice it tends to become empty romantic notions
<4> haha
<4> went over my notes from this morning
<4> still have a few questions
<4> How is it that humans set goals and tasks without knowing how far they need to understand a system before thinking about tasks involving that system or process?
<2> while( x > 0 ) { x = F(x); }
<2> does this loop terminate?
<2> things like that
<2> that's a question for math really
<2> yudkowsky calls for a method in which a program can 'understand' the workings of a program from its code. Sort of meta-program that can reason about how code works.
<2> Correct me if i'm wrong, but didnt Alan Turing have something to say about that?
<2> Namely, that it is impossible.
<1> Where I come from we call that 'formal methods software engineering'.
<0> <hybrid_@ef> good ol alan
<4> Star-glider, In response to my question?
<1> Though at the moment the tools are fairly basic.
<2> Indeed, no program can be written that can tell if a program HALTS, let alone does something more exotic.
<1> They can only prove relatively simple constraints on your code.
<4> guess not
<5> paros, that's the general belief at least. not sure about the proof for it though
<1> True, paros, but I think that's a straw-man when it comes to AIs writing code.
<1> Humans can't tell if an arbitrary program will halt either.
<1> But we write code fine.
<2> Star-glider If i write a program who branching behavior is based on a floating points operation, such as e^x, how can a program reason about that function? It would have to know theoretical things about mathmematics.
<1> e^x is a smooth function.
<1> As such it's actually pretty easy to reason about.
<2> In fact, for a given program to know everything about its code, it would have to be more than a Phd in mathematics.
<1> If you know the range of possible input values, then you can work out the possible range of output values.
<2> I wonder how Yudkowsky says this hurdle is overcome...
<1> Well, it would need the appropriate metadata anyway.
<2> Star-glider Consider a loop which contains a division by zero. But this happens in a very odd place, namely like this:
<1> Paros, how do humans write code without having PhDs in mathematics?
<2> a = 0.0001; while( x > 0 ) { a = a*a; x = 1/a; )
<2> Here x grows without bound...but i know that because I know theoretical things about math.
<2> How is this meta-programming software going to acquire this knowledge, and how will it organize and use it?
<0> <x00q@ef> Doesn't a shrink without bound?
<2> Star-glider - THat's not what I said. WHat i said was, to understand the flow of code a in ROBUST WAY you have to be a Phd in math.
<5> girls girls girls (no offense athena :) you really should take the time to play with neural nets. it woud so discipline your train of thought. help you understand how basic it actually is when it comes down to it
<1> Well, to know what a program will do without running it, yes you'd have to have domain expertise.
<1> That's pretty clear.
<2> Star-glider But its not good enough that Singinst can write an app that can WRITE CODE. It needs one that can understand code and modify it. I am harping away on the "understand" part.
<5> these human level considerations will never get you to the goal that is a drooling baby
<1> If you want an AI to understand code, your options are programming in a knowledge base containing the domain expertise, trying to get it to learn the neccessary knowledge from example or experimentation, or trying to reason from first principles.
<5> AI=machine learning. learning is it
<2> Star-glider Wait a minute... I think this whole conversation is bunk. Semantics is not inside of a corpus of text no matter how large. It follows that programming code can never be understood by something that merely looks at lots of code. I think Yudkowsky is a snake-oil salesmen in this sense.
<1> Uh-oh, a semantics debate.
<2> Generally, there are some people out there like DanF_DrC and I who know that semantics is not inside of the symbols that represent them.
<1> Where are the semantics then?
<1> Program code is actually a nice neat example here, because its semantics are defined pretty clearly by a language spec.
<1> (or hardware spec)
<2> But Yudkowsky and the singinst is playing a big trick on us. Namely, by saying that some program can be written that UNDERSTANDS CODE on a theoretical level (ie.. how it branches and grows or shrinks without bound). But this is nonsense. You need a human to read human code.
<2> Star-glider Yes the operations are nice and neat, but only a human knows that a certain summation gets closer to zero and does to "forever"
<2> does so forever*
<1> Do you mean 'code inevitable requires knowledge about the world it models or interacts with to understand' ?
<5> guys like Yann LeCun and Richard Sutton tends to agree with me though
<1> An AI would know that if you programmed it in as an inference rule.
<2> Star-glider NO...im saying that code that contains floating-point operations can only be reasoned about by humans who are aware of things like infinite summs.
<1> Or if it had a learning or reasoning mechanism that could generate such a rule.
<3> danf_drc: can't say none taken! :)
<1> Why can't an AI know about infinite sums?
<5> athena :)
<2> Star-glider Some infinite sums diverge...other converge...there is NO GENERAL RULE to this. If I programmed a loop of code that sums a series of fractions to 2.718 and then the exit condition for the loop was S= 3.0 HOW, I ask you, can a program know this without knowing higher math?
<1> It couldn't, but my point is that if it needs to know about particular maths to reason about some code, then you could teach it about those maths.
<2> n =1; while( S < 3.0) { S = (1+1/n)^n; n++ };
<2> does this loop TERMINATE or not?
<2> It does not
<1> I agree that there isn't going to be some simple fixed algorithm that can reason about code.
<2> HOw can a meta-application tell this? (it would have to know higher math)
<1> Gah, so /teach it higher math/, if that's what you need.
<2> Star-glider Not so fast! How can you teach it higher math unless it already knows english?
<5> drooling baby remember
<2> See.. "oh well just teach it higher math" aint gonna fly here, because the app has to know how to USE math,
<1> By programming in the axioms of maths as a logic system.
<1> This is actually one of the more successful existing AI domains.
<1> (automated proving systems)
<5> learning is fundamental
<1> That's been done too.
<1> Lenat's project right before Eurisko was AM ('A Mathematician').
<1> Which tried to learn about maths by experiment.
<1> But this is a bit off topic for self-improving AI, if that's what you're criticising.
<2> No its right on topic
<1> Because this kind of maths-heavy code isn't necessarily going to be in an AI.
<2> What about the issue of not having the code in the first place
<6> it's very interesting
<1> Do you mean in these sense that no one knows how to build the initial code base of an AI that can understand its own processing yet?
<6> Star-glider, *i* mean as in breaking the bounds of the programming language
<1> That's certainly true. We don't know if it's even possible, and even if it is it might require an impractical amount of computing power.
<2> no even if you had that, what is this slow code doing (that will be optimized?)
<1> The simple answer to that is to take a look at Newell's SOAR system.
<1> The cannonical symbolic AI, and one which some people (most noticably Newell himself) claim is a valid basis for self-improving AI.
<1> I don't think it is of course, most people don't.
<1> But it does provide a nice model.
<1> SOAR is supposed to work by finding solutions to problems using very general reasoning mechanisms (including a broken kind of almost-reflection), and then when it does so generating some 'shortcut' code so that next time the problem is encountered only a narrower, more relevant space is searched.
<1> As I understand it, all self-improving AI works the same way; by adding special-case code onto a general base.
<2> oh really?
<1> In addition to optimising out any design inefficiencies, but obviously there are fairly strict limits to that.
<1> (unless you start taking the Singularity stuff seriously, where AIs will develop nanobots and improve their /hardware/, but that's not really a sensible research plan :) )
<2> actually singist does not mention nanobots
<2> although they have articles about self-modifying code
<1> I suppose you could think of it as a human programmer with a computer but no software.
<1> As they encounter real-world tasks, they solve them with human general intelligence.
<1> But when they encounter the same tasks again, they write special-purpose programs to do the tasks.
<1> The programs are much faster and more accurate than the human at a given task, but limited to that one thing, and need the human to run them, feed them input, and do something with their output.
<1> Over a long period of time, a good hacker will write lots of programs that link together, including some scripts that handle some of the linking and chosing what to run.
<1> Self-modifying AI is supposed to work like that, but the programs are being written to handle small cognitive tasks, not big application areas.
<1> That and for general intelligence you're need a /lot/ of layering and recombination into bigger modules.
<1> But yeah, you'd have to have the initial AI filling the role of the human programmer, and no-one knows how to build one.
<5> one word for selfmodification could be.. learning
<1> Most AI learning algorithms don't alter the code-base, or even generate executable code.
<1> (at least, not in the same format that the AI is written in)
<5> if a neural net is a computing paradigm what is the code for it if not the learned...
<6> one of my bots wrote code into $eval() strings
<3> adding new rules to a rule base is a way of self-modification isn't it?
<1> Neural nets generally have fixed topology and neuron functions.
<6> one wrote code into disk files and then launched a shell to run it
<1> Most NNs don't even have loops (recurrent NNs seem to be the exception not the rule in AI research).
<6> no one was interested
<5> Star-glider, you mean like the brain
<1> It's certainly true that anything that acts like a universal Turing machine can act as a platform for AI (or any other software, in principle), but that's a point lots of people mention with it very rarely being relevant.
<5> turing completeness is not helpful here
<1> Kinda like Godel incompleteness. It's a cool bit of maths, but it's not normally as relevant as people seem to think.
<5> do try to understand that learning is fundamental. that's how it's done. machine learning.
<1> Fundamental to general intelligence yes.
<1> Part of what makes direct self-modification so interesting as a concept is that humans can't do it.
<6> ?
<1> We can't consciously rewire our neurons.
<5> backpropagation is the base example of a useful paradigm able to provide solutions on its own. something from nothing
<1> We can only do things that we hope will have certain gross effects on our brains.
<3> we can't rewire them and we can't upgrade the wetware either as is possible with hardware when intel release a new chip
<1> No, though that's not so relevant when the brain still has a lot more raw computing power than our computers.
Return to
#ai
or
Go to some related
logs:
#perl
noapictimer
#linux
craigslist.co
#oe
#asm
linux gentoo unmerge sendmail
Linux unrar -p
ohgood
gentoo :command not found2:
|
|