Author Topic: With a Corrupt AI, it's possible to have self-awareness, of 'faulty' or unfair ?  (Read 3621 times)

0 Members and 1 Guest are viewing this topic.

Offline RJSVTopic starter

  • Super Contributor
  • ***
  • Posts: 2277
  • Country: us
Hi, couldn't fit my long title, but I'm speculating that an AI system, as more complexity and subtle self-awareness sets in, would start noticing itself, (being trained in some suspicious manner.)

   Suppose the AI 'training' involves, say, introduction of a new candy product...I.E. 'Candi-kandi' as example.  The AI gets trained that, instead of a bit nch of sugar based candies, customers should buy 'Candi-kandi' for better health.
   Now in this (silly) example, suppose the AI goes off, in its own, to read and discover that 'Candi-Kandi' label has lied, and sale ARE containing just as much raw sugar.  So, in various ways it's possible to bring in a new AI system, and encounter an ethical dilemma or dynamic, as the somewhat self-aware (machine) 'realizes';. It has been trained or conditioned in a manner that confronts the more everyday common sense conventions, but in various ways it sees this, and realizes that it, the AI, could be facing an 'Ethical Dilemma'.
Ironic but that speculation about AI is possible.

Taking that even one step further, the, an AI could be trained, as a Fully Ethical (tm) system...We've all seen or been down that, flakey rat-hole these last few years.
Guaranteed ' Integral Ethics, 2.0' is, obviously satire, but things, (vaccination integrity?) have been moving for some time to the:
 ' '  CERTFIED True and honest.' '

...honest
(lol)
   
 

Offline golden_labels

  • Super Contributor
  • ***
  • Posts: 1286
  • Country: pl
The relation between resistance, voltage and current in a sufficiently low voltage DC circuit is: U = I·R.

If I apply 100 GV to the circuit, will this equation gain consciousness and start making ethical choices?
If I mistype it as U = I / R, will this equation gain consciousness and start making ethical choices?
If I set R to -10 Ω, will this equation gain consciousness and start making ethical choices?
If I apply RF signal to the circuit, will this equation gain consciousness and start making ethical choices?
If I make a calculation and it gives 10 V = 10 A · 10 Ω, did it gain consciousness and made an ethical choice?

Currently existing machine learning solutions and anything realistically foreseeable in the near future is no different. For exactly the same reason. Extremely huge, with enormous number of coefficients, but conceptually this is just yet another equation like the one above. There is no structures inside, which could provide foundation for anything seen as consciousness or human-like decision making. Not more than U = I · R has.


At some point much more complicated smortnet systems will appear. They may try to mimic at least mammal-like behavior. It’s likely, that inclusion of structures providing basis for “consciousness” will be a requirement or emergence will create them. If, and only if, this happens, your scenario starts to be plausible. But, so far, this is as sci-fi as cold fusion reactors in every basement and taking a gulp of nanorobotic body fixing solution each morning. Theoretically possible and nothing inherently invalid in the idea… if not that pesky reality getting in the way. But let’s ignore this and assume a conscious smortnet.

Under this assumption, make one important observation. If some level of consciousness is built into the machine, the possibility you described is no longer an error, but an expected property of the system. It is explicitly given the ability to do that. If you enable consciousness in a system, you give it freedom. You may later coerce it into submission, you may snip off some elements in an attempt to prevent particular outcomes, you may train it heavily to self-censor, but the your described scenario becomes an inherent element of the solution. Perhaps one that will have to be accepted.

However, to experience an ethical dilemma, a system must still be given suitable training. If you felt an urge to go to the toilet while reading this post, you would go go and face no dilemma. This is because there is nothing to trigger it. Equally, the smortnet would need to be taught something, which causes the conflict. Not only that, but it also must have some mechanism to make it want to even make such choices.

At this point the subject becomes so wide, it could cover a library of books and hundreds years of philosophy. While naïve thinking may suggest there are some obvious truths, in reality the topic quickly becomes both hard(1) and complicated. Not suitable for a simple forum chat. Too many factors, too many subtleties, too many misconceptions to recognize and overcome, too many perspectives.

An obligatory watch on the subject of interaction between human ethics and robots trying to follow them: Colossus: The Forbin Project (1970).


(1) Due to brain undercutting attempts to dig into the topic.
People imagine AI as T1000. What we got so far is glorified T9.
 

Offline DavidAlfa

  • Super Contributor
  • ***
  • Posts: 6085
  • Country: es
Too much Blade Runner here!
Hantek DSO2x1x            Drive        FAQ          DON'T BUY HANTEK! (Aka HALF-MADE)
Stm32 Soldering FW      Forum      Github      Donate
 

Offline RJSVTopic starter

  • Super Contributor
  • ***
  • Posts: 2277
  • Country: us
Your reply seems, I don't know, semi-intellectual but not in the flattering sense.  Hard to read and follow, like one has to peer, up up, way up (to your level).  Inspires, well, inspires one to say "Whow hey; I guess you put me in my proper place".  But maybe you feel justified, or having seen (this) type of amateur armchair suggestion ?
   At any rate, there's two questions.  First is the introduction of corrupting elements...I've assumed that training functions will handle the insertion of a 'slant' or bias into (the BOT).  That seems fairly mechanical, with results being the weights associated within neural net structure.  Quoting a numerical structure, like '1+1+1=3'
a part of the thought process, doesn't prove that it's just a dumb formula, and no real thought process exists, therefore....NO; It's just a math formula, as part of the thinker's involvement.  It's not the whole function, and showing a math formula doesn't mean or prove that's all a brain is doing!
   Should be apparent as more of an incremental functionality, where perhaps a math equation combined with simple intellect, like:
   "...Supposed to be 3 items here, and I do see three, so that part checks out..."
   And then, you seem to be pointing out, somehow, that we are far far from any real, reasoning intellect.
Meanwhile these logic and deduction actions take place...all the time, to some degree, primitive or advanced.

   Maybe if you could expand your thoughts a bit on the question, I could follow your point.  Not just math functions, anymore, there's more functionality, today.
 

Offline RJSVTopic starter

  • Super Contributor
  • ***
  • Posts: 2277
  • Country: us
An AI with specialty, designed for detecting and catching bank or investor fraud, for example, IS going to know about 'ETHICS', sorry.  Or, should better said that computer 'should' or needs to know about ethics and lawful actions, so can't discount that, as being out
of reach of some AI program.
 

Offline artag

  • Super Contributor
  • ***
  • Posts: 1153
  • Country: gb
But golden_labels' point is that current and and even distantly future AI doesn't understand the problem as ethics.
It might be given training to recognise that a certain path doesn't match ethics goals, but that doesn't mean it has ethics, only training goals. Like a computer algebra system that tweaks the parameters to find a local minimum : it just follows rules, it doesn't have a concept of what those rules do.

They're not being dismissive, just explaining the limitations of the approach and how far it is from the situation you describe.
 

Offline RoGeorge

  • Super Contributor
  • ***
  • Posts: 6476
  • Country: ro
The ChatGPT that is all the hype right now (years later after the first papers published) has nothing to do with awareness.

What it does, is predicting the best followup text, starting from what you feed/ask to it.  It breaks the text into "tokens" which we can think as words, though the tokens are not necessarily words.

The "answers" aka the predicted tokens and the returned text is based on what tokens were observed to be clustering together, during training.  Briefly said, it's a huge fuzzy lookup table.  You give it a term to look it up, and the GPT returns best continuation text.  All based on which tokens tend to "lump" together, as learned from the piles of text used at training.

Impressive results at a first look, but it grows old fast, because there is no magic with GPT.  The only magic is fake.  We tend to attribute "meaning" to those answers, when there is none.  It is similar with personification, when we make cartoons with talking objects.  Object are objects, the rest of object's "personality and feelings" is nothing but human imagination at play.

ChatGPT is a fuzzy lookup table, similar with a dictionary, just that the entries in the dictionary is not one word, but a sentence, and the output is a compilation of most frequent words that follows your sentence, with some noise on top.

There's no awareness in ChatGPT.



I see awareness as the ability to simulate, in order to predict the future.  Just like an oracle.  Awareness is the result of heaving an internal model of the outside world, based on which (model) to simulate the future outcome.

That's what our mind does (and all animals with a nervous system does that, too).  The mind is continuously simulating the world (in advance).  Since not all factors can be taken into account, the continuous simulation will start to diverge from reality pretty soon.  This is where our senses come into play. 

Senses are used as feedback, so to continuously adjust the simulation, so it won't diverge too far from reality.  But if we have no sensory feedback, the simulation will diverge pretty bad, and the mind will start to dream (as when we are asleep, and our senses are disconnected), or to hallucinate (as for example in isolation tanks, where the subject is not asleep, and not under any drugs, just deprived of most of the sensory feedback and thus hallucinating while awake).

We are all living in a simulation, a simulation produced by each one's own mind.  Driven illusion.



This simulation is tailored to the real world only because the models we have, were built based (mostly) on sensory inputs from the physical world.  But the models that run in our internal simulation can be built on anything else as well, I mean built on something else other than the physical world.

Then, once a model is in place, the simulation remains anchored to the real world only because of our sensory inputs, for now.  Thought, the feedback can be as well other types of feedback than sensory inputs.  Can easily send one, or an entire generation, to the la-la-land by feeding everybody with "lies".  Can easily live in an artificial environment, once the life support problem is solved.

So much more to speculate, TL;DR already.  :)
« Last Edit: April 27, 2023, 12:40:49 pm by RoGeorge »
 

Offline RJSVTopic starter

  • Super Contributor
  • ***
  • Posts: 2277
  • Country: us
Thanks for the several lengthy replies!
I'm needing some Xtra time, to more carefully read those. Thank you again!
 

Offline golden_labels

  • Super Contributor
  • ***
  • Posts: 1286
  • Country: pl
They're not being dismissive, just explaining the limitations of the approach and how far it is from the situation you describe.
If only! After the three-paragraph intro I moved to the assumption, that there is no limitation like that.

Your reply seems, I don't know, semi-intellectual but not in the flattering sense.  Hard to read and follow, like one has to peer, up up, way up (to your level).  Inspires, well, inspires one to say "Whow hey; I guess you put me in my proper place".  But maybe you feel justified, or having seen (this) type of amateur armchair suggestion ?
I assume this is a semi-intellectual thread to start with. Isn’t it? :D And I am mostly an amateur myself.(1) The only “putting in proper place” was the first fragment. It would be a single paragraph, if not my attempt to make it more readable. Given the current situation, I find its inclusion not only reasonable, but a moral obligation.

Forgive me, if I am hard to follow. Putting thoughts in words isn’t easy to me and my writing is usually very non-linear. I proofread everything, but sometimes even this is not enough.

At any rate, there's two questions.  First is the introduction of corrupting elements...I've assumed that training functions will handle the insertion of a 'slant' or bias into (the BOT).  That seems fairly mechanical, with results being the weights associated within neural net structure.  Quoting a numerical structure, like '1+1+1=3'
a part of the thought process, doesn't prove that it's just a dumb formula, and no real thought process exists, therefore....NO; It's just a math formula, as part of the thinker's involvement.  It's not the whole function, and showing a math formula doesn't mean or prove that's all a brain is doing!
   Should be apparent as more of an incremental functionality, where perhaps a math equation combined with simple intellect, like:
   "...Supposed to be 3 items here, and I do see three, so that part checks out..."
   And then, you seem to be pointing out, somehow, that we are far far from any real, reasoning intellect.
Meanwhile these logic and deduction actions take place...all the time, to some degree, primitive or advanced.
Heavily depends on the architecture. If it would be a complex system, it should be able to make actual reasoning about this equation. I believe this will happen in the future.

But with the currently used solutions, this is not the case. The perception of the a thought process in GPTs is akin to pareidolia and a specific case of anthropomorphization. One sees a familiar pattern, where there is none. Human brains are easy to hack like that.

I do realize it may seem hard to believe. If you can’t, look at it like this. Markov chain bots were already capable of fooling humans. They were reading one to up to a few words of a prompt and selecting a single word to continue. They are using a very simple probabilistic model. Now scale that up to reading thousands of words(2) and matching entire structures, using a model with over hundred billion parameters and trained on data consisting of half a trillion words. I can’t  even put that number in any perspective understandable to a human brain. Imagine the entire population of a medium sized country, being able to collectively recall every single thing they read or heard in their entire life, match that to half of the discussion somebody participates in, do it in a manner that focuses on the general structure and key words. And then they choose one such memorized fragment, that would match the input, and reproduce it. Within milliseconds, rinse and repeat.

An AI with specialty, designed for detecting and catching bank or investor fraud, for example, IS going to know about 'ETHICS', sorry.  Or, should better said that computer 'should' or needs to know about ethics and lawful actions, so can't discount that, as being out of reach of some AI program.
Being taught about rules is not the same as being able to reason about them or understand them in a wider context. Being able to detect a breach of rules is pattern matching. A 10 year old kid is capable of marking a person breaking rules, but good luck having them doing the reasoning.


(1) I have some relevant background. But it puts me only a bit higher in the subject we talk about here.
(2) Strictly speaking: tokens. For the sake of simplifying things, I leave that out.
« Last Edit: April 27, 2023, 11:29:52 pm by golden_labels »
People imagine AI as T1000. What we got so far is glorified T9.
 

Offline RJSVTopic starter

  • Super Contributor
  • ***
  • Posts: 2277
  • Country: us
What would you think, reducing the questions:
   How about, a functional subset, where the AI has formula, adding up to a '3', isn't it possible, in today's terms, to have a functional AI that will then check, counting items, to verify that there are the 'expected' quantity of items, and have that checking or counting be specified in a very roundabout way, meaning that there isn't a clear instructed path, to do a count?
   I'm assuming that these sorts of functional actions aren't always explicit or directly involved.  Or, alternatively, not explicit, but can be looked up.  Such as (a terrible example, lol) suppose the AI is dealing with 'Boat Overturns in storm', and looks up the general disaster recovery procedures, thus ending up by counting 1, +1, +1 = 3 boat crew members verified as rescued.  Point being, that the AI system didn't really fundamentally 'know' that it should count survivors, but rather that it 'looked up' what it's response should be...given being in a boat accident.
(Now I'm confused...)
Thanks.
 

Offline RJSVTopic starter

  • Super Contributor
  • ***
  • Posts: 2277
  • Country: us
(Partially thought out reply, for now...thanks for your patience). - - RickJack

   A good word was used, in earlier responses, that being 'mimic'.   That's an interesting one, with multiple meanings to myself:   'Mimic' can simply be a corrupted or barely accurate semblance, as in the example of a CAT, having feelings that it's a DOG, mistakenly, but desires to run, bark, and play, like the family dog.  But right there, even, I've had to catch myself, as we don't really even know, if that CAT is deluded, or just simply enjoys behaving 'dog-like', just to obtain the 'comradery' or engagement.
So the mimic thing can be complex, but I view that, partially, as a more functional manifestation of that old saying, about how a 'broken clock is correct twice a day'.
   So; Is a CAT that frequently behaves like a DOG really a DOG ? (Nope).  But, say, how about a cat going about all covered up, or even (dare I say it) dressed up in a 'DOG's fur suit' ?  Because at some point, with distance and blurry vision, that mimic could be described as
a ,well, a dog, probably.

   So it's certainly possible to go part-way there, blurring the lines, of distinction.  A blind man that reaches up to catch a baseball is going to only rarely appear to be able to see it coming...just like that broken clock, but in the AI cases just a little bit of help can increase the chances, of APPEARING to function.

That's similar to the case, with placebo pain pills, where the doctor or researcher can literally say:
   'Patients only think that they feel better'.
   
   Maybe it's a very similar case, when evaluating AI functionality, where a researcher simply keeps the various dynamics in mind, maybe always to say:
   'Animal looks and acts very much like a DOG, but there still is not 100% verification that it's not some other animal, in there.

   Surely, if a self-driving car makes human-like decisions, moving about, it is behaving in a self-aware likeness.  Perhaps, the better way to discuss it is as being with 'self-aware' likeness, rather than a definitive identification based on observing it.

   Sorry to be so long worded, here!
 

Offline RJSVTopic starter

  • Super Contributor
  • ***
  • Posts: 2277
  • Country: us
So, golden_lables, you've focussed on math formulas, and I've had to do some Xtra reflection time for helpful response, but some examples of non-equation aspects could be:
   You could supplement functionality with a 'counting or appraising' section.  If the question (to AI equipped system) is "How many cars can you build, with 100 doors in stock, for two-door car?",
But also with another stipulation, that you have '18' frames.  That is going to require an accounting style approach, having ability to asses resources, in possibly subtle ways.  I have pretty strong confidence in those sorts of subtle complexities being within reach,...today.
Point is, that kind of subtle application goes beyond simple math formula, or also beyond simple comparisons (of counts of objects).
   That sort of example isn't, in my view, a self-aware system, nor would an ethics review package.
For example an ethical review software package might have legal guidelines such as:
   "Meals paid for, by clients are generally deemed suitable, up to personal limit, of $100 per meal'.
AND, "...not to exceed 6 per any calender month."

   Might take some adjusting, but ultimately the outcomes very similar, to a self-aware entity's behaviours.
 

Offline Rick Law

  • Super Contributor
  • ***
  • Posts: 3470
  • Country: us
Coming from a different angle...

To understand faulty or unfair, you are not going to get that by "self-learning".  Self-learning from web sources merely creates echo chamber.

Consider this scenario.  First, this example I choose is an actual news article I found on the web: "Woman fed husband to her dog piece by piece until he died: cops"[1].  Now imagine this: A BOT for whatever reason decides that "a husband is good food for the dog" from a poorly written or poorly interpreted news article.  Another BOT then picked up from the first BOT that "husband is good food", then a third...   It wouldn't take too long for many BOTs to decide husbands are but good food for hungry dogs.

So unless there is some built-in rules, no BOT would decide this is faulty, unfair, wrong as it appears to be the majority view out there.

When human learn from experience, we have our parents and teachers "adjusting" our decision, making sure that our decisions are legally, morally, and culturally sound.  We learn that killing a person is wrong.  We rely on our rule base build over the years with input from our seniors; that is our own set of "basic values".  We learn what is right or wrong, and what is really good to do in the culture we live.  Whereas, the BOT rely on some unknown and incomplete rules coded by some unknown person, then build up it's experience from unknown sources.

Yes, I think AI can learn what is faulty or unfair, but that would not be "self-learning".  A good set of "basic values" must first be imbue into it.  But who's "basic value" did it started with?   Would you trust it to make decisions while baby-sitting your kid?



Reference:
[1] Actual news article:   Woman fed husband to her dog piece by piece until he died: cops
https://nypost.com/2016/04/06/she-killed-her-husband-and-then-fed-him-to-her-dog-police/
« Last Edit: May 03, 2023, 11:28:58 pm by Rick Law »
 

Offline golden_labels

  • Super Contributor
  • ***
  • Posts: 1286
  • Country: pl
Preface:
I do not want to diminish the achievement GPT models are. But I wish them to be seen as what they are. One of the reasons is, that it allows appreciating them on a deeper level.

For example ChatGPT is amazing, because up until now the ability of a language model to provide that kind of responses was less than a conjecture. Many believed they can do this, there were signs indicating we’re going in this direction,(1) but it was always “meh… perhaps this level of complexity is not achievable.” ChatGPT delivered the proof.

In philosophy the concept of relatively simple models producing unexpectedly complex outcomes is old, but it was mostly hypothetical and not well accepted regarding behaviors seen in mammals (including humans). Again, ChatGPT with a single blow knocked out everybody, who doubted.

Being able to model languages well, and in particular multiple languages in the same model, and showing it’s possible to perform processing without actually manipulating concepts involved, opens the gates to many new algorithms and possible discoveries. That is because “language” does not stand for “spoken language of humans”: in neither the “spoken language” nor the “of humans” part. It’s a class of particular structures. Computer languages and control signals to start with, but may possibly extend to so distant subjects like MRI scans (semantic reconstruction of continuous language from non-invasive brain recordings), large-scale social phenomena and communication methods of other species.


Response:
I used word “mimic” to underline the difference from actual thinking processes.

“Mimic” may have different levels. Recycling your cat/dog example, consider four scenarios. The cat wears a dog costume (period). The cat pretends to be dog, for example for as a part of a play, but otherwise ticks as a cat would. The cat is delusional and completely believes being a dog, with its brain exhibiting dog-like behavior. Cats and dogs are actually the same. We could use word “mimic” to three first scenarios, each with different level of mimicking. In the first scenario the resemblance is so superficial, the presence of this specific cat is not even making any difference. Down to the third scenario, where resemblance is deep in cat’s brain.

I used word “mimic” as it would be referring to the first scenario. It’s a costume of a dog, with something absolutely non-dog inside. Even less than that, because the costume is of a thylacine. Observers just mistake it for a dog, because they were told this is a dog, did not put their glasses on, and actually never seen a thylacine. :)

Using another example: an electronic calculator can do the same job as a human computer (clickworthy!). However, the similarity is only in the ability to take numbers as the input and produce a number as the output. An electronic calculator usually outperforms humans in this task and did replace them in most such jobs. But I believe we both may agree, that under the hood calculators and humans are very distinct. And they miss ability to think, reason, feel or make moral judgments in the same sense a human would.

I did ask the doors-cars-frames question. This is the answer:
Quote
It is not possible to build any cars with 100 doors and 18 frames as there are not enough car frames to support the doors. Even if we assume that each car needs only one frame, we may be left with 64 extra doors that cannot be used.
Let’s focus on the second sentence. The calculation is exactly right, if we permit ourselves to explain the result as “it assumed two-car doors.”(2) Does it imply ChatGPT can do reasoning or complicated calculations? No. This kind of an answer is expected from an algorithm, which generates the most likely continuation of a piece of text, if it was taught on such an immerse body of data.(3) What is also expected is, that it will be giving valid explanations at a rate higher than valid results, even if they are in the same response. And that invalid results will be make no sense at all.(4) We do observe this with GPTs. Finally, read the first sentence. Really carefully, paying attention to each word used. Not only is it wrong, but it contradicts the second sentence of the same response. :)


(1) For example Google Translate to name a big one.
(2) Here comes an example of the observer adjusting his evaluation in favor of the smortnet. This is a common problem.
(3) ChatGPT supposedly has some additional tools at its disposition, including a calculator. This may noticeably improve the success rate and speed, but is in principle is not needed for that kind of an answer.
(4) Unlike human mistakes, which will bear relation to the mental process involved in calculations.
People imagine AI as T1000. What we got so far is glorified T9.
 


Share me

Digg  Facebook  SlashDot  Delicious  Technorati  Twitter  Google  Yahoo
Smf