Did you know that large language models are described in the Bible? And not just in the New Testament. Isaiah 28:10 literally translates as, "For it is precept upon precept, precept upon precept; line upon line, line upon line; a little here, a little there." But, in the original Hebrew, it reads more like, "Tzav la-tzav, tzav la-tzav, kav la-kav, kav la-kav z‘eir sham, z‘eir sham." In other words, what it is really saying is, "He keeps telling us blah blah blah blah blah."
Both literally and idiomatically this is a *really* great description of large language models.
The bloom seems to be coming off the AI rose. Yes, corporations are still investing millions, and even billions, of dollars in artificial intelligence, mostly in companies that are pursuing the "Large Language Model" chimera. Initially, I thought that the large language models were a cute trick, and might possibly have some uses. As time went on, we found about out about "hallucinations," disinformation, "jailbreaking," and a whole bunch of other problems with a large language models. What we *didn't* find was any activity or process where the large language models were really useful.
Businesses said we can use these large language models to automate mundane tasks. But, really, how mundane does the task have to get before we entrust it to a large language model? And if the task gets that mundane, is it really a task that we need to do? I am reminded of Peter Drucker's famous quote that there is nothing so useless as doing, efficiently, that which should not be done at all.
So, for quite a while, my take on the large language models has been that they are a solution in search of a problem.
Given the money being thrown at them, that search seems to have become more desperate.
I have been doing presentations, to various groups, on artificial intelligence and the different types of approaches to the field that preceded the large language models, and would seem to be considerably different in approach. As well as the various risks, both in not pursuing artificial intelligence, and of pursuing artificial intelligence too avidly. Recently I was given a date for a presentation to a group that I knew would want a bio.
I hate writing bios.
I hate *listening* to bios, for that matter. They always seem to be full of stock phrases and puff pieces, and rather short on actual facts or reasons that I should listen to this particular person, who is doing this presentation not out of any particular interest in or insight into the topic, but as a means of reaching for the next rung on the ladder of fame and success.
I hate doing them, on myself. So, I thought it would be amusing to have ChatGPT write a bio of me. ChatGPT is, after all, the tool that pretty much everybody thinks about when they think about the large language models. In order to see how much the technology has improved over the past few months, I decided to also submit the same request to Claude, the tool from Anthropic. (Claude is supposed to have better "guard rails" against jailbreaking, than does ChatGPT.) And, today, Meta announced Llama 3.1, so I included Meta AI.
Well, It was somewhat amusing.
But it has also become part of my presentation. The bios that all three systems produced point out, in large measure, the problems associated with the large language models of artificial intelligence.
Rob Slade was born, variously, in 1948, 1952, and "the 1950's." That last is accurate, but rather imprecise. I had not known that there was so much controversy over the date of my birth (although I *was* very young, at the time), especially since it is not exactly a secret. So, some of the material is purely in error. I have absolutely no idea where they got 1952 and 1948 from. I also wonder why all three systems decided that it is important to give the year of my birth, but none mentions where I was born, or lived most of my life, or, indeed, where I live now. (Claude *did* manage to figure out that I am Canadian.) Again, there is no particular secret about this.
I gave them a three hundred word limit, and, somewhat to my surprise, given the weird and wonderful errors that LLMs seem to be capable of making, all three did come in "under budget," at 246, 268, and 279 words. All three systems wasted an awful lot of their word count on what could primarily be called promotional or sales material. I had noted that this is a tendency in the large language models. This isn't terribly surprising, given that most of the material that they would have been able to abstract from the Internet, would have been primarily sales, marketing, or other promotional material. I don't know whether this speaks to the tendency, on the part of the large language models, to hallucinate.
It is nice to know that I am renowned, with a career spanning several decades, have made significant contributions to the field of cybersecurity, authoring numerous books and papers, with a solid foundation for my expertise, I'm influential and my publications have served as essential resources for both novices and seasoned professionals, I give engaging presentations, and my ability to demystify complex security concepts make me a sought-after speaker and educator, with a career marked by significant achievements and a commitment to advancing the field of information security, my work has been instrumental in shaping the understanding of digital threats and has left an indelible mark on the information security landscape. My legacy serves as a testament to the importance of dedication, expertise, and innovation in the ever-evolving landscape of information security. You will note that none of these claims are really verifiable, and so they are also basically unchallengeable. On the other hand, my contributions have been recognized with several awards. (Well, I *did* get a mug from IBM, at an event ...)
I am also known as "The Father of Viruses." Oh, gee, thanks, Meta.
ChatGPT found three of my books, Claude two, and Meta one. Nobody found all five. There are other Robert Slades on the Internet. Over thirty years ago we had the "Robert Slade Internet Club, with a membership of about a dozen. There is a Robert Slade who publishes on Greek pottery and inscriptions, another who publishes on fishing lures, another who teaches mathematics, and another who is a DJ for events and parties. In order to give AI the best chance, I specified that I wanted a biography of the Robert Slade who was an information security expert. To their credit, none of the models came up with specific references to the publications of these other Robert Slades.
However, I have been credited with degrees from a number of universities which I did not attend, so their backgrounds may be the explanation. (Meta found one right university, but the wrong degree.) But that doesn't explain my "memberships" in a number of tech associations to which I don't, and have never, belonged, and the failure to list the *one* organization to which I did, for some time. (Apparently I also participated in numerous committees and working groups aimed at improving global cybersecurity standards and practices. Unfortunately, not true.)
Claude says that I developed an early interest in computers and technology during my university years, which is pretty much false. Although I did explore one computer system, and taught myself programming, at the time I felt that computers were pretty much useless except for crunching data, and, even at that, they were a lot of trouble. Both Claude and Meta insist that I started with mainframes: again, pretty much completely false. (Meta insists that I started with computers at age twelve: I was almost thirty before I really got interested.) Claude managed to figure out that it was my reviews of antivirus software that first got me any attention. (Overall, I'd say Claude definitely gets top marks for accuracy. But being the best of a bad lot is not exactly a stunning accolade.)
Yes, it has been somewhat amusing to have the LLMs write my bio. But that's because I hate writing bios anyways. If this bio were, in fact, important to me--if it were some type of resume--I would not be able to accept the errors that have been made. Granted, it has automated the task of generating marketing and promotional fluff; with claims that are both unverifiable, but unchallengeable. But do we want to produce this kind of material, for any reason? Is this not the slippery slope to the post-truth world in which we live, where major figures are able to lie, blatantly, provably, without regard or interest in whether or not people are going to be able to say that is a complete and utter lie? I have, in another piece, commented that the large language models could not be better at producing disinformation than if they had been designed that way.
Have they?
(I also asked Meta AI to do a headshot of me from the available images of me. The result:
)
No comments:
Post a Comment