(I seem to have wandered into a number of digressions in composing this piece, but they all seem to tie together, so I hope you'll bear with me ...)
Decades ago, I was at a teacher's conference. I was in a session dealing with computers in education. The morning paper had published an article about computers in education, and, particularly, using computers to teach, and, therefore, replacing teachers. Someone easked about this. The presenter thought for a moment, and replied that any teacher who could be replaced by a computer, *should* be replaced by a computer. His point was that teaching was a complex task, and that any teacher who taught in such a rote manner that he (or she) could be replaced by a machine would be better off out of the profession, and the profession (and the education system) would be better off without him (or her).
Which story I am relaying to lead into:
We are worrying about the wrong thing with regard to AI.
The programs DALL-E, ChatGPT, and others that rely on "machine learning" and pattern models derived from large data sets, have recently racked up an impressive series of accomplishments. They have produced some amazing results. Everyone is now talking about artificial intelligence as if it is an accomplished fact. It isn't.
These programs have been able to produce some absolutely amazing results. But they have been able to produce amazing results for people who have been able to learn how to use them. That does not fit my definition of any kind of intelligence, let alone an artificial one. If the impressive results can only be obtained by people who are willing to put in the time to learn how to use these tools, then they *are* tools. Just tools. Complicated and impressive tools, yes. But just tools. They do not have their own intelligence.
Intelligence would require that the system would be able to provide satisfactory results for pretty much anybody. A person, and intelligence, is able to query the requestor as to whether the results provided are satisfactory. If the results are not satisfactory, the intelligence is able to query the requester and find out why not, and use this information to modify the results until the results *are* satisfactory. And that is, of course, only one of the aspects of intelligence. There are many others, such as motivation. So, while I'm willing to grant that these tools are very sophisticated, complicated, and definitely useful developments, they don't get us that much closer to actual artificial intelligence.
The results from these tools have created a great deal of interest, even in the general populace. It has particularly created interest within the business community, and new investment artificial intelligence projects and companies is probably a good thing. (Unless, of course, we are all on a hiding to nothing and we never *will* get real artificial intelligence. But let's assume for the moment that we will.) It has also engendered a good deal of discussion on the wisdom of pursuing artificial intelligence, and the dangers of artificial intelligence. Since my particular field is dangers associated with information systems, I have been very interested in all of this, and think it's a good thing. We should be considering the dangers, particularly the dangers, with regard to machine learning, that we have created, and are perpetuating, bias in our systems, particularly when the data sets that we use to train machine learning systems are, themselves, collected, collated, and maintained, by artificial intelligence systems. Which may already be affected by various forms of bias that we engendered in the first place, and have never realized are even there.
There is, however, one fairly consistent theme that appears in discussions of the dangers of artificial intelligence, and which DALL-E, ChatGPT, and their ilk have indicated is a false concern. While it is primarily a screaming point of the conspiracy theory and tin foil hat crowd, many people are concerned about the possibility of what tends to be referred to as "The Singularity." This is the hypothesis (and it is a fairly logical hypothesis), that when we do, actually, get artificial intelligence, that is truly intelligent, and can work on improving itself, that such a system would advance so rapidly that there would be absolutely no way that we could keep up, and it would, from our perspective, almost immediately become so intelligent that we would have no chance of controlling it. It would rapidly become intelligent enough that any of our protections, which are never perfect, would leave open a vulnerability which the system itself could exploit, and therefore it would, again, almost immediately, from our perspective, be beyond our control. What happens at that point is open to a variety of conjectures. This intelligence could turn evil, from our perspective, and wipe out the human race. (Some people would consider this a good thing.) Or, it might create a kind of benevolent dictatorship, managing our lives and having pretty much complete control of the entire human race, since it would be able to commandeer all information systems, which means basically every form of business, industry, entertainment, and any other human activity. Or, the artificial intelligence may simply take off, without us, leave us behind, and disappear from any involvement with us. Or, well, there are all kinds of other options that people have explored and theorized.
None of these options particularly scare me.
That's the wrong thing to worry about. What we should be worrying about is relying on artificial intelligence, and, particularly, these recent examples. These tools are not really intelligent. They do not understand. They do not comprehend. They do not appreciate. They just predict the likelihood of the next piece of output from patterns, in masses of data, that they have being fed. (I have mentioned with, elsewhere, the fact that what we are feeding them is possibly biasing them, and that the bias is probably self reinforcing. And we'll come back to that point.)
I asked ChatGPT to write a sermon. It did a very banal, pedestrian job. When I pointed out some of the flaws, ChatGPT basically gave me back the same thing, all over again. It didn't understand my complaint: it just responded based upon my statement. It didn't understand my statement: my statement was just a prompt to the system, and had similar enough terms to the first prompt that the output was, basically, identical.
I gave a friend an opportunity of a trial with it. He said that it produced a reasonable Wikipedia article.
I think this is illustrative in ways that most people wouldn't. I have never thought highly of Wikipedia. While I applaud the general concept, I feel that, in actual implementation, Wikipedia is the classic example of the pooling of ignorance. When I first set out to assess Wikipedia, I, of course, as an expert in the field, looked up the entry on computer viruses. It was terrible. As far as I know, having checked it several times in the intervening years (although I haven't looked at it recently), it's still terrible. At one point it had more than one factual error per sentence. And, of course, in those early, carefree, bygone days when I still have some thought that maybe Wikipedia might be a useful exercise, I made corrections to these errors. Corrections which were, of course, immediately rescinded by Wikipedia's editorial staff.
Wikipedia does not rely on expert opinion. How could it? The editorial staff of Wikipedia do not know how to judge who is expert, and who is not, on a given entry, or topic. The original computer virus entry did, and as far as I know still does, contain the common received wisdom on computer viruses, with all of the mistakes, errors, and misconceptions, that the common man holds about computer viruses. Therefore, when I tried to correct these errors, the Wikipedia staff felt that I was introducing errors, and so they reverted back to their original mistake-ridden text. For an actual expert, there is, actually, no point in even attempting to correct the errors in Wikipedia. Wikipedia relies upon the common man's perception, and, therefore, it's pretty close to social media as a source of information. There is an enormous quantity, but there is not necessarily very much quality.
(My take on, and attitude towards, Wikipedia, while formed many years ago on the basis of the number of mistakes in the technical entries may be [possibly unfairly] reinforced by the fact that after Gloria died, Wikipedia removed all entries to her from my entry in Wikipedia. I found this very personally hurtful, and, to this day, I have no idea why they did it.)
Wikipedia relies upon entries available on the web, and therefore may rely heavily on social media. Wikipedia also goes by seniority, not by expertise. If you are higher up on the Wikipedia editorial food chain, you can reverse any entry or correction that an expert makes. Therefore, it is no surprise that Wikipedia is riddled with errors, particularly in recent discoveries, and in any area where expert opinion is of value. Wikipedia has become the Funk and Wagnalls of the information age. It's widely available, possibly useful in general cases, and very often wrong.
This is why my friend's further comment that it made "the classic error," was also illustrative. "The classic error" will be repeated, in many articles, and postings, made on the Web, by those who think they know the case, but are not necessarily fully informed. This type of material will be repeated, ad nauseam, on social media, thereby reinforcing the truth and validity of this erroneous material.
And, of course, ChatGPT has been trained on social media. ChatGPT has been trained on material, and text, that could be gathered to give an indication of how we humans speak in response to queries. Or challenges. (This is also why ChatGPT is likely to become obnoxious and abusive if you challenge it. That's the way people react on social media, and it's social media that provides the material that has trained ChatGPT.)
ChatGPT, and DALL-E, the graphic, or art, generating version of the pattern model tool, are simply responding, with patterns that they can predict from a massive database that they have assessed, of what is to be produced in response to any prompt. It's simply using statistical models (very complex statistical models, to be sure), to generate what the average human being would generate, if challenged in the same way. There is no understanding on the part of either ChatGPT, or DALL-E, or any others of those pattern model tools. They do not understand. They do not comprehend. They don't have to. They just churn out what it is likely that a human being would churn out in response to the same prompt.
I asked ChatGPT to produce various materials in recent tests. What I got was pedestrian and uninspired. Well, of course it was. ChatGPT is not understanding, and doesn't have any way to obtain inspiration. It's just going to generate something in response to a prompt. And it is going to generate what most human beings would generate. And most human beings are, let's face it, lazy. So, what most human beings would produce, when challenged to produce a an article, or a sermon, or a presentation outline, would be pedestrian, banal, and uninspired. It's the type of article that you read in most trade magazines. Vendors go to professional authors and ask them to produce an article on blat. The professional author does a quick Google search on the topic, feels that they are expert, and turns out banal, pedestrian, uninspired text. There is nothing innovative, and there is nothing in the material that leads to any item or idea that would spark creative thought. That's not what most human beings do, that's not what most of the material on social media is, and so that's what ChatGPT produces.
Many years ago, I ran across a quote which said that creativity is allowing yourself to make mistakes. Art is knowing which ones to keep. ChatGPT does make mistakes. But most of them simply are not worth keeping. ChatGPT doesn't think about what it's doing: it just predicts the next, most likely next, probable word that a human being would write in this stream of text. So, ChatGPT isn't going to create anything that's inspired, isn't going to create anything that's creative, isn't going to produce much of anything that is much of use for anything, and if we fail to understand this, we fail to realize what relying on ChatGPT can produce for us. Which is, basically, so much dross.
I have recently read many articles which assert that ChatGPT can provide for us mundane letters, mundane article outlines, and mundane articles themselves, which will be of a help in business. But that is only because we, as a society, have become accustomed to the mundane, and accept it. And, if we continue to use ChatGPT for these types of purposes, we will, in fact, produce more mundane dross, and, increasingly find that garbage acceptable. We are training ourselves to accept the banal, and the uninformative. Eventually we will train ourselves to accept a word salad which is completely devoid of any meaning at all.
ChatGPT is becoming more capable, or at least more facile. It is being trained on larger and larger data sets. Unfortunately, those data sets are being harvested, by and large from social media, and by and large with the aid of existing artificial intelligence tools. Therefore, the fear that some have raised, that we have already biased our artificial intelligence tools by the data that we gave to them, is now being self-reinforced. The biased artificial intelligence tools that we created with biased data, are now being used to harvest data, in order to feed to the next generation of pattern model tools. This means that the bias, far from being eliminated, is being steadily reinforced, as is the bias towards meaningless dross. If we rely on these tools, that is, increasingly, what we are going to get.
And, with the reliance on artificial intelligence in the metaverse, that is what we are going to get in the metaverse. The metaverse is an incredibly complex undertaking. It is, if all the parts that we have been promised, are included, a hugely complex system, orders of magnitude more complex than any we have yet devised, with the possible exception of the Internet, and the World Wide Web and social media itself. We will need to have artificial intelligence tools to manage the metaverse. And these tools are going to have our existing biases, and are going to have the bias towards uncreative, uninspired garbage. And therefore, that's what the metaverse is going to give us.
Increasingly readable, and convincing, garbage to be sure, but garbage nonetheless. Do we really want to be convinced, by garbage?
At any rate, in another test, I complained to ChatGPT that I was lonely. I mean, most people don't listen anyways, and most people don't listen very well. So I figured that ChatGPT would be at least as good as one of my friends, who, after all, have disappeared, since they are terrified that I'm going to talk about Gloria, or death, or grief, or pain, all of which are taboo subjects in our society.
The thing is, ChatGPT doesn't know about the taboo subjects in our society. So, it gave me an actually reasonable response. Now, it wasn't great. ChatGPT cannot understand what I am going through, and cannot understand or appreciate the depths of my pain and loneliness. But at least it was reasonable. It suggested a few things. Now, they are all things that I have tried. But they were reasonable things. It said to talk to my friends. As previously mentioned I can't. When challenged, ChatGPT fairly quickly goes into a loop, basically suggesting the same things over and over again. But it also suggested that I take up volunteer work. Now, of course, I knew this. It is something that I suggest to people who are in depression. And I have done it. And, it does help, to a certain extent. So, a half point, at the very least, for ChatGPT.
I can give more points to that than that to ChatGPT. It doesn't give me facile and stupid cliches. It didn't say anything starting with "at least." It didn't tell me that Gloria was in a better place. It didn't tell me that bad things wouldn't happen to me if I only had more faith. All of which people have said to me. And it's all very hurtful. So ChatGPT at least gets another half point for not being hurtful. (If we are still trying for the Turing test, at this point, I would say that, in order to pass, we would have to make ChatGPT more stupid and inconsiderate.)
But I'm not willing to give ChatGPT very much credit at this point. It's not very useful. It wasn't very analytical. And I did challenge some of its suggestions, to see what kind of response I got when I challenged ChatGPT on various points. I did sort of challenge it on the friend's point, and it didn't get defensive about that. So, at least another half point to ChatGPT.
But, as I say, it's not very good. It's as good as a trade rag article, and it's probably as good as any Wikipedia article. In other words, not very good. The material is pedestrian and. I don't think that bereavement counselors have anything to worry about, quite yet.
I should also note that so far, I have the free version of ChatGPT, and therefore I am not talking to GPT-4. This is GPT-3. So it's not as good as the latest version. And I would like to give the latest version a try, but I strongly suspect that it wouldn't do all that much better. But it would be an interesting test.
Relying on ChatGPT, for anything but the absolute, most pedestrian tasks is asking for trouble. It can't understand. It is going to make mistakes. If you present it as an interface, and, talking about my test about loneliness and bereavement, I realize that I may have prompted some idiot with a grief account to try and tie ChatGPT on to a grief account, as a kind of automated bereavement counselor, well, that's really asking for trouble. Trying to use ChatGPT with people who are, in fact, in real trouble, could create a disaster. Please, those of you with grief accounts, do not try this at home. This is only for trained idiots, who actually know that there is no such thing as artificial intelligence, and realize that ChatGPT isn't that much more of an event from ELIZA. (If you don't know who ELIZA was, it passed the Turing test more than four decades ago, and it only took two pages of BASIC code.)
There is concern that adding the appearance of an emotional component to computer systems, and particularly artificial intelligence systems, will create dangerous situations for users. This is a very realistic concern. We have seen a number of instances, over at least half a century, where individuals have attributed to, sometimes very simple systems, intelligence, personality, and even concepts of a soul.
As only one aspect of the difficulties, but also the importance, of looking at emotive, or affective, artificial intelligence, or any kind of intelligence in any computer system, consider the case of risk analysis. In information security, we need to teach students of the field that penetration testing, and even vulnerability analysis, does not lead you directly to risk analysis. This is because penetration testing, auditing, and vulnerability analysis are generally performed by outside specialists. These people may be very skilled, and may be able to produce a great deal that is of value to you, but there is one thing that they, signally, do not know: the value of the assets that you are protecting. The value, that is, to you. The value of an asset, whether a system, piece of information, or a database of collected information, has a value to the enterprise that holds it. But it is only that enterprise, and the people who work there, who really do understand the value of that asset. The value in a variety of ways, and therefore the protections that must be afforded to that asset. Therefore, no outside firm can do a complete risk analysis, since they do not understand, or fully comprehend, the value, or values, and the range of different types of value, that the asset holds. For the company.
Currently, our so-called artificial intelligence tools, may be able to perform some interesting feats. But they do not understand. And, particularly in regard to affect and emotion, they do not understand, even what these are, let alone how important they are. Now, we can certainly make some effort to instruct artificial systems as to certain aspects of human behavior, and the indicators that the human may be in high states of emotion. However, the systems will have no understanding, no comprehension, of these emotional states. They will not understand the subtleties and nuances of emotional states. We can give them a set of directives as to how to behave with regard to people, but they will not understand, they will only behave. This is a backstop solution, and it cannot be complete. It is akin to the difference between justice and law, in all of our human societies. Supposedly, we think of our legal systems as providing justice. We even call institutions related to the legal system departments of justice. But we all know, in our heart of hearts, that there is a difference between legal and right. We all know that there are times when our laws come to an unexpected situation, and are then unjust. In the same way, we cannot simply give a set of commands to a computer, as to how to deal with a human that is in an emotional state, and expect that this will address all possible situations. Because the computers do not have an understanding of emotion.
In this latter regard, I highly recommend reading "Affective Computing," by Rosalind Picard. Her work looks not only at human factors engineering, but also at the significance of affect, or some similar analogue, in regard to motivation and decision in automated systems.
No comments:
Post a Comment