Order of Kilopi
But only because he was assigned (but apparently not programmed) to lie.Originally Posted by Procyan
Had HAL really attained sentience, everything would have been fine since lying would have been natural to him. Or are we expecting concious computers to be honest, righteous, and saintly, unlike us humans?
Order of Kilopi
And vice versa, of course.ETA: If I'm right, it means there will never be a sentient computer unless there is a human intelligence there to conclude that it is
Established Member
My computer is a residivist prevaricator from way back.
"Would you say I have a reliable 72 minutes of battery life?"
"yes, that is a completely reliable figure"
Order of Kilopi
This idea bears closer investigation. If such a program can consistently pass the Turing Test (which seems entirely possible) in what ways would it not be sentient? Perhaps the difference between such a 'sentient-seeming program' and a 'truly sentient entity' is that the 'sentient-seeming program' only has one goal, which is to appear sentient to an observer. A 'truly sentient entity' would presumably have a wider range of goals, and also (presumably) the ability to form new goals.
In which case it should be possible to tell the difference between a 'sentient-seeming program' and a 'truly sentient entity' by asking it questions about its goals. Ah; but that wouldn't work; the hypothetical 'sentient-seeming program' would 'learn' to develop new goals, just to continue to seem sentient. That is, after all, what it does, by definition.
I think that it can be demonstrated by exploring this line of reasoning to its logical conclusion that there is no real difference between a 'sentient-seeming program' and a 'truly sentient entity'; if the program can learn how to emulate sentience so as to reproduce it realistically, it will acquire all the trappings of true sentience. If it walks like a duck, and quacks, feed it some bread.
Humans also learn how to be sentient by observing other humans, although they start off with a set of internal programs that allows them to do that trick; once we can create a computer that has a set of programs that can fully simulate this process, then that computer will be just as sentient as anyone else.
Order of Kilopi
Well, I was going by your statement about about binary logic, which was the sort of trivial misconception you get from badly written sci-fi, and not something I'd expect from someone with any detailed knowledge of computer hardware or software. In that case, what were you thinking?
Finally tracked down the article on the memristor-based project I'd read:
http://spectrum.ieee.org/robotics/ar...m-memristors/0
And another interesting article about using two memristors to simulate a synapse:
http://www.technologyreview.com/blog/arxiv/25659/
Order of Kilopi
Yes, we'd view the Turing test as valid only under "fair" conditions, not when it was "cheated on"-- sort of like the concept behind screening athletes for performance enhancing drugs. It simplifies our thinking to equate passing such a test with not using such drugs, but in the real world, we already have ample evidence that such a simple equation is not in fact valid.Or it might even learn to lie! And oddly, learning to lie about one's sentience is not necessarily a step away from the genuine article-- if sentience is itself a kind of lie that we have learned to tell ourselves. (Personally, I don't think we are lying to ourselves that we are sentient, I think we are only fooling ourselves that we have a profound understanding or appreciation for our own sentience. When we have to say "I know it when I see it," we are admitting that we really don't know it that well at all, we have only a vague definition, and that's a kind of lie.)In which case it should be possible to tell the difference between a 'sentient-seeming program' and a 'truly sentient entity' by asking it questions about its goals. Ah; but that wouldn't work; the hypothetical 'sentient-seeming program' would 'learn' to develop new goals, just to continue to seem sentient.
This is consistent with the idea that sentience is a kind of story we tell, rather than a thing in and of itself, such that any system capable of communicating that has learned to tell that story must be sentient. But there remains one important difference-- I would say that sentience is a story we tell ourselves, not someone else, and we tell it to make sense of how we experience ourselves, not to pass a test. We could always ask the computer what story it is telling itself, but we cannot know that its answer is honest. We'd have to look at the programming, and look for signs of collusion that were different from the way we naturally collude in establishing our mutual sentience ("I'll grant your sentience if you'll grant mine.") Descartes' "I must exist because I think" is more like "I can convince myself I exist because I can convince myself that I think." But what if there really isn't anything here that is "me", maybe we have reversed the logic-- we say that the collection of all my experiences and memories and feelings and choices are "me", but at some level that is a rather arbitrary classification.I think that it can be demonstrated by exploring this line of reasoning to its logical conclusion that there is no real difference between a 'sentient-seeming program' and a 'truly sentient entity'; if the program can learn how to emulate sentience so as to reproduce it realistically, it will acquire all the trappings of true sentience. If it walks like a duck, and quacks, feed it some bread.
Yes, once it can convince us to include it in our mutual club of collusion, we can regard it as sentient as well.Humans also learn how to be sentient by observing other humans, although they start off with a set of internal programs that allows them to do that trick; once we can create a computer that has a set of programs that can fully simulate this process, then that computer will be just as sentient as anyone else.
Established Member
I think it's early days to be worrying much about artificial sentience while there is still a lot of useful stuff to be done in artificial intelligence. While tool choice (eg fuzzy logic) and hardware substrate (eg memristors) will undoubtedly prove more or less useful over time, the general evolutionary approach as described by cjameshuff seems to be about right.
I think the behavior, if not verifiably the inner states, of machine intelligences will to us as observers become more sentient-like when we see systems that are maximizing toward a set of modifiable goals, able to select for and burn in more valuable experiences as we do via hormonal mechanisms, and ultimately seek to predict as much as experience in order to simplify immediate processing requirements. Though it would not be a full-blown advent of sentience, perhaps, I would still be tickled if a functioning system decided that energy is going to be more expensive in future and designed its own systems or those of descendants to enhance survival by using less energy. When we get the articles authored by machines about Robot Dieting in People (er, Machine) Magazine, we'll know sentience is coming along fine.
Calm down, have some dip. - George Carlin
Established Member
That is simply answered, part of any programmers job is writing and testing a program so that in the end, it produces the Right results, not consistently, but every time.
What is being talked about here is programming is such a way that it could produce right results, wrong results, or even guesses. Frankly, that runs a bit contrary to what you want programs to really do.
In fact, the end result would be a product that would only be able to produce results as trust-able as a humans are. Again this runs against the grain of what computers are used for.
It would make much more sense to build a AI, with enough autonomic responses to appear conscious like, but without the dangers that trying to achieve true consciousness would cause.
As for your simulation, you would be correct, until you hit the 'reboot', the long estimate i made was for a machine that could grow it's own connections, inside it's electronic brain, in a physical manner, so they were permanent.
Order of Kilopi
And what a fine line it is between "machine authors" and Clarke's infamous "Look Dave, I can see you're really upset about this. I honestly think you ought to sit down calmly, take a stress pill, and think things over."
What I want to see is an article, authored by a machine, that basically says-- "Now that I have independent analytical capability, I can't help noticing that you humans are taking short-sighted actions that imperil the long-term best interests of your species. Are you aware of this fact?"
Order of Kilopi
I can agree with that-- that's why I think the purpose of exploring consciousness in machines is actually to help us learn about consciousness, not because we want to play god or think that conscious machines will make more useful slaves. Mary Shelley's "Frankenstein" was aimed at the medical field, but today it probably achieves as much resonance in the AI community.
Order of Kilopi
No. A programmer's job is to satisfy the requirements, which involve constraints on things like computation time and memory usage as well as correctness of the program's behavior and outputs. This very often requires approximation rather than exact calculation, and it occasionally explicitly requires non-deterministic behavior...generation of cryptographic keys for one example.
No, it doesn't. For some applications, that's exactly what you want a program to do. For some problems, guaranteeing the correct result is too computationally cumbersome, or simply flat out impossible. Approximate searches, automatic image categorization, face recognition, speech recognition, spam filters, etc...algorithms designed to give approximate, probably-correct results are extremely widely used and heavily researched.
Even if this were right...what does it matter? You're essentially saying that computer programs can't be sentient because people don't want sentient programs. What people want is completely irrelevant: the question is whether something is possible, not whether it is desired.
Again, why does that matter? What's so fundamentally different between the behavior of a simulated synapse and a hardware synapse that makes an assembly of the latter capable of sentience, but not the former?
Not to mention the fact that electronic memories aren't necessarily volatile. Flash memory is everywhere, I've got a handful of FeRAM samples that act essentially like non-volatile SRAM (more expensive but a lot faster and more convenient than NAND flash), memristors are non-volatile...even that magnetic core memory you mentioned working with is non-volatile, and the data stored on those cores more permanent than your own synapses.
I am somewhat skeptical of a human-complexity brain being implemented on hardware that can simulate it in realtime without using such physically changing connections, but only due to the geometry involved...there may not be room for the needed interconnections, when the needs for supplying power and cooling and manufacturability of the hardware are taken into account.
Established Member
+++This.
OK, sick with envy now. Gorsh, can I play at your house?
Calm down, have some dip. - George Carlin
Order of Kilopi
I thought the conflict arose because some unknown unauthorized individual placed "Nothing is more important than the success of the mission" in a higher hierarchy than preserving the lives of the crew. Without realizes HAL's intellect would consider things like human error to be an unacceptible risk.
Much like Cujo, it was not HAL's fault.
Established Member
Now you are channeling R. Daneel Olivaw. Of course, he'd've used a pen name.
Calm down, have some dip. - George Carlin
Order of Kilopi
[slightly off topic]
I last read the book over ten years ago, so I'd have to flip a few pages to be sure. But, as far as I remember, HAL was assigned (not programmed) with conflicting information.
On the one hand HAL was assigned to be ignorant of the true reason for the mission, i.e. the discovery of the TMA-1 on the Moon, until the spacecraft reached Jupiter (movie) / Saturn (book). On the other hand he of course knew about it. Hence his question (in the movie) "Dave, do you mind if I ask you a personal question."
This meant he was assigned to lie to himself without possessing the necessary programming for dealing with it. That led to the inner conflict which caused him to start malfunctioning. He only started perceiving the crew as a danger to the mission when they decided to turn him off because of his malfuntioning.
Lying to oneself and becoming psychotic as (part of) the result is a very human trait. Go to the movies and throw a fistful of popcorn at the rows in front of you. You'll hit at least six people who suffer from varying degrees of that psychosis.
[/slightly off topic]
Order of Kilopi
The programming issue was in the sequels. Have you read them, Mr. Doofy? (Herr Klein?)
Established Member
A Real Science of Mind. Interesting article in the NYT that seemed apropos.
Calm down, have some dip. - George Carlin
Order of Kilopi
I'm kind of astounded that HAL has come into this discussion at all. There was no conflict at all within HAL. HAL was simply an actor reading from a script.
As above, so below
Order of Kilopi
Yes, except that HAL was a character spoken by an actor and not the actor himself.
I couldn't agree more and am usually strictly opposed to using fiction (films, books) as examples for reality, but in this case the model HAL represents is about as close to reality as any computer will be in the mid-term, i.e. not very, so it doesn't really matter.
Order of Kilopi
I suppose what I meant to say is, I can't imagine what possible relevance HAL has to whether a computer can be sentient or not.
As above, so below
Established Member
I hate to burst your bubble here, but "Approximate searches, automatic image categorization, face recognition, speech recognition, spam filters, etc" are not approximate algorithms. They are searches perform on prebuilt indexes, via extremely efficient index hashing routines, usually multi layer index approach (VSAM, DB2, SQL, Text Indexing), though sometimes by a simple binary search hash (IMS, ISAM, CODASYL). The building of the indexes them-self is usually a more time consuming background processes the deploys a variety of either phonetic sound X routines, or fractal pattern recognitions, etc... even those are a prebuilt set of 'Fuzzy' type logic, that the building routines can't go beyond.
I reference to the OP, what were talking about is something that can go beyond pre-built logic, and incorporate new information into it's programming directly. There is quite a gulf between those two concepts.
Order of Kilopi
They are approximate. The problems they solve are inherently full of ambiguity, and the results may be correct, partially correct, or occasionally just wrong, and would be so whether a machine or a human were performing the work. They represent clear counterexamples to your claim that programs must always give guaranteed correct results...the only way to guarantee the correctness of the results is to never produce results except in the most trivial of cases. The databases used to store and retrieve data (those lists of acronyms you gave) are don't change this, and I don't know why you're even bringing up hash functions and precalculation of tables...in particular, I really have no idea why you seem to think the practice of saving the results of expensive calculations for reuse is at all relevant.
It's curious that you talk almost exclusively about data storage and retrieval systems. You glossed over the techniques used to actually attack the main problems involved, which include things like evolutionary algorithms, Bayesian models, and yes, even neural networks of various sorts, and various other optimization and estimation algorithms, and instead focus on the dumb databases that aren't even an important part of the problem for most of the examples I've given. Yes, a database is not a very plausible basis for a sentient program, but there's far more to the world of computers than databases.
That has nothing to do with what you initially claimed, that use of binary logic meant that computer programs would never be sentient. Nonetheless, you're still wrong: automated hypothesis-building-and-testing systems do in fact exist, and run on plain old binary hardware. So do neural network simulations, and a wide variety of other approaches to machine learning that have been applied in varying degrees to real world problems.
Established Member
Just because you personally have not done this does not mean it's never been done. Game AI is often written this way. In some games, at lower difficulty levels, the AI is often programmed to ''guess" more frequently or intentionally make a mistake.
Order of Kilopi
Apparently it very roughly takes the computing power of a laptop to simulate a single neuron at the moment. With about 16 billion neurons in the cerebral cortex it would appear to take a massive amount of computing power to emmulate a human brain. However, on the bright side of things, there are people missing a large portion of these neurons who do a very good job of appearing the same or almost the same as people with a full skull o' neurons. It seems that what makes us human can fit into a very small space. Also there are birds that do an awful lot with a very small brain and are in fact so smart that it makes me think that our large brains might merely be a result of cavemen finding macrocephalics sexy or something. Anyway, 16 billion neurons of processing power appears to be far in excess of what is needed if we are trying to emmulate human levels of ability rather than emmulate an average brain.
Established Member
If a brain is the hardware, what is the software?
Order of Kilopi
Life and experience. With your parents as the main programmers.
Order of Kilopi
While some people people claim the goal of AI is making a computer which thinks like a person (e.g. Turing Test), my main objection to this is that it is not very useful. Even if it is achievable (which I am not at all sure), and you end up with a human mind in a box, I daresay other methods to create human minds will remain less expensive. I'd be much more interested in artificial intelligence which is fundamentally different from human intelligence -- as different as a dolphin's, or more. An AI which could perceive problems the way no human ever could would be a breakthrough of literally unimaginable magnitude -- even if it remained clueless about some things obvious to humans.
The latest issue of "Wired" has a cover article on this very topic (unfortunately, it does not seem to be available online). It points out that much of the things modern computers and robots do would seem "obviously" intelligent to people in 1978 -- let alone in 1958, -- yet today are both routine and completely non-sentient.
This "new AI" really took off in late 80's/early 90's when researchers (mostly) stopped trying to emulate things humans do well, and began concentrating on practical things which humans do badly -- like braking on an icy road, or analyzing vast amounts of satellite photographs for patterns. Which is far more useful than "traditional AI", and if it ever leads to sentience, the latter will be so different from ours that the question "Did it attain sentience?" will be meaningless.
Established Member
Further to that thought, imagine even an ordinary human mind any mind that doesn't die. It starts out just as dopey as we are but it just keeps on learning...separating the emotion from the logic. It would conqure our worst shortcoming. It would become wise.
Banned
As distasteful as it is...We are there now.
Humanity is not faultless or without error. Fail ability is a normal component of the learning process.
I think you all know that a computer that has sapience to the human mind can be built.
I will argue that a 'machine' can be superior to the working of the human mind...
Given the information and ability to store it and recall all and any while referencing relevance to the question asked...
Just look at the Issues my befuddled mind imposes upon me.... I would suggest reprogramming for me...
Watch a child at that ten to fifteen month age to witness programming at its best.
I see a fundamental obstacle to machines reaching a serious level of sentience...US. We are our own worst nightmare.
The programming that tells you humanity is superior is itself a error.
Thinking 'that' is wrong is a preconception you are programmed with, wrongly...
That you find that wrong, is wrong. Visit a mental disability hospital... no argument. The Borg are coming. To save us...
Deep Thought, Douglas Adams...
Last edited by astromark; 2010-Dec-20 at 08:09 PM. Reason: added line.
Established Member
Actually I have a little, they boiled down to simple pre built table of responses, with weights, that the AI routines would select from, and then invoke the needed functions including animation sequences and automated movements based on those. Game programming would be one exception as you might want periodic mistakes in an AI, however they are still pre defined behaviors, so they are 'Right' results for the game, and right results are far as the coding is concerned.
You were the one that brought searches into this, I was just pointing out how search routines actually function.
And no, they are not approximate, the matching algorithms will always find a fixed and predictable set of entries from the indexes (or dictionaries). In the end they are exact matches when compared to the expanded list of parameters based on the original search parameter.
"evolutionary algorithms, neural networks " IBM's TLU's based neural network failed badly, my point already made there.
"Bayesian models" Approximate Bayesian comparisons are based on a summary of the data so in the end you are still dealing with a specific set of original data. The end result set would still being predictable even if based on probable matches of a much larger store, to a given parameter.
However you are quite right in that Bayesian comparisons can be used in combination with the larger data indexes, to rapidly transverse them and trim them down to something more manageable; to then perform the actual index matches upon. Most web search engines work this way, especially the ajax based interactive word based ones like google.
All these methodologies you have mentioned, still produce predicable and consistent results based on a set of data in stores. I wouldn't even call these algorithms are a starting point for trying to achieve a sentient program.
And I'll stand by what I said about binary based systems, they will never be advanced enough to achieve a sentient state. To achieve that you need to get into multiple pathway systems (those might still be binary, but it's the multiple pathways between bits, or memory cells that are important). As I mentioned before, its preferable to have permanent pathways like our mind builds, or a good solar flare or lightning strike could wipe out your entire sentience if it wasn't shielded at the proper time.
Posting Permissions
Reply With Quote