Western Standard - August 22, 2023


You can teach AI, but it may not learn

Hosted by
Mentioned
NWT residents tell of harrowing evacuations from fire-ravaged territory MPC Daily Market Update Aug 23, 2023

Episode Stats


Length

16 minutes

Words per minute

139.94

Word count

2,270

Sentence count

133


Summary

Summaries generated with gmurro/bart-large-finetuned-filtered-spotify-podcast-summ .

In this episode, researcher and IT management professional Gleb Lizzyk talks about artificial intelligence, and why it may not be as intelligent as we think it is. She explains how ChatGPT, a generative AI program developed by Google, may be less intelligent than we think and why we should be worried about it.

Transcript

Transcript generated with Whisper (turbo).
Hosts, guests, and mentioned names generated with spaCy (en_core_web_sm), reconciled against Wikidata.
00:00:00 Welcome Western Standard viewers. We are privileged to be joined today by Gleb Lizzyk,
00:00:07 a researcher and IT management professional who's not as anxious about artificial intelligence
00:00:15 as a lot of us have become recently. I had one of our columnists who was playing with ChatGPT
00:00:24 and it asked him. My columnist asked the program to give him a speech that would sound like the
00:00:32 Prime Minister and bless my soul it did somewhat. So clearly there is something in there and yet you
00:00:42 are not as sure that this is going to change the world after all. So can you tell us what it is
00:00:49 about artificial intelligence that's maybe not as intelligent as we think it is?
00:00:57 Absolutely. I'll try my best. One of the fun things I asked GPD to do was one of the many things
00:01:07 is to spin off a meaningless word salad about diversity and inclusion that could be used as
00:01:12 let's say a statement of someone's allegiance to the idea. ChatGPT exceeded my expectations
00:01:19 proving the fact that it can weave the words together so that they come across as carrying a deep
00:01:27 meaning whereas typically if you dig deeper ChatGPT then starts to show lack of any substance or
00:01:37 coherence. So no wonder that imitating our dear leader is so easy for ChatGPT to do.
00:01:45 Anything that has been repeated many times or templatized as a piece of as a piece of information or the
00:01:56 generative AI like ChatGPT can reproduce. So it's a language model not an intellect as we understand it.
00:02:05 Oh I see. So at least I think I see. You did an experiment with vaccine reactions. I think that I think this was in the column that you provided for us.
00:02:17 Yeah. And you tried to get it to change its mind and you succeeded somewhat in making it say that there were risks for younger people in association with vaccines.
00:02:30 Now regardless of what the viewer may think about the effectiveness and safety of the vaccines this discussion is actually about the ability of a program to learn.
00:02:40 And I think what you discovered was that no matter what you told it, it always wanted to revert back to the algorithms on which it was based.
00:02:51 Was that essentially what you were trying to, what you were putting out in that article?
00:02:56 Yes. Let me explain briefly what I tried to do. So we as human beings have knowledge, we have reasoning, right?
00:03:04 And we obviously have our own language patterns. So we can respond to the inputs with the specific outputs. So does ChatGPT, but ChatGPT or similar generative AI models, as they, to use the scientific term,
00:03:20 they don't really possess any knowledge and they don't really think through their responses or analyze them in any shape or form.
00:03:29 That's why it takes us a while to respond to a question, at least to a difficult question, right? We have to think about it.
00:03:36 ChatGPT does not think about anything. It just passes the inputs through the patterns and gives you the output.
00:03:43 That's very, it's a very simplistic form of explaining it. But what I tried to do is not even present new information to ChatGPT, but just simply say, what do you know?
00:03:57 How much do you know about, let's say, vaccine science effects and all the other things that, yeah, and I used that information to drive ChatGPT to a conclusion
00:04:06 that was drastically, drastically different from its initial statement about how vaccines, the COVID-19 vaccines are safe and effective, right?
00:04:17 So I wasn't feeding ChatGPT with any information, just using what it already had, right?
00:04:25 And it came to this drastically different conclusion. But when I, when I worded the, what follows directly from that conclusion, specifically, if,
00:04:36 if, if someone under 50 is looking to increase his or her lifespan, if you will,
00:04:47 should vaccine be recommended to the person? And it was very clear from the previous conclusion for ChatGPT that no, it's, it's not the right thing to do, right?
00:05:01 I'm not arguing that at this point, like whether vaccine are good or bad or anything like that.
00:05:05 I'm just saying that the conclusion was that vaccine is, is, is really harmful for anybody for 50 in comparison to COVID, right?
00:05:14 More, significantly more harmful, like up to 20 times and something like that, right?
00:05:18 And, and, and the response to, of ChatGPT to the question of whether vaccines should still be recommended for that person was yes,
00:05:29 because benefits outweigh the risks and stuff like that. So it hasn't really changed its mind drastically.
00:05:40 And, um, needless to say that the conversation we had in a, within a particular chat window just stays there.
00:05:48 It does not reflect on any other chats or any that I would call like a general broad understanding of ChatGPT of the issue, right?
00:05:58 So it just stays within, localized within the chat. As soon as you delete the chat, it's gone.
00:06:04 So, Gleb, obviously it was set up that way by the, uh, the team, the, the people who invented it in the first place and you tried to change it and you found you couldn't.
00:06:17 But I understand that, uh, the way of the future with AI is that you can select the news feeds that you want to be guided by, feed them into, uh, an intermediary program.
00:06:32 And then the program just sort of automatically writes according to the, the, the, the algorithms in ChatGPT.
00:06:41 Is there, can you through the process of feeding information into ChatGPT change its mind?
00:06:51 It seems that so far from what you've said you can't, but surely we're not forever stuck with the, uh, preconceptions that one person had when they designed it.
00:07:01 Yeah, well, it's, it's not that I could not change its, its mind.
00:07:09 Um, like I, again, within that chat window, I was able to change its mind, even though this mind is really big quotes, right?
00:07:18 Because its mind is very isolated to a single chat.
00:07:21 Um, the, uh, what I'm trying to say is that unless the whole learning process for, for generative AI changes, and the learning process for the generative AI is just a one-time shot.
00:07:37 It's very, uh, computer intense, very expensive process that happens once, right?
00:07:43 Which basically sets the patterns of, uh, ChatGPT brain or, uh, neural network, if you will, uh, to a certain state, right?
00:07:54 That state doesn't change.
00:07:56 It's also enforced through the hard program policies of those who train ChatGPT.
00:08:02 Um, and, um, whatever conclusions you can drive ChatGPT to, they have no real effect on those patterns.
00:08:13 Uh, so that, that process is basically set in stone, right?
00:08:19 Uh, so the, the mind, sorry, the mind of the ChatGPT is set in stone.
00:08:23 So who controls that, ultimately?
00:08:26 I mean, should we be alarmed at this, that the government or an agency or even big tech, by whatever name you want to know it, can actually, uh, quietly work on our minds and change our opinions without us even knowing that there is an alternative way to look at it?
00:08:46 How scared should we be?
00:08:47 How scared should we be?
00:08:49 Sorry, say it again.
00:08:50 Uh, what, uh, what was the question again?
00:08:52 The question was, uh, so let me ask you this.
00:08:57 If ChatGPT has been programmed to think a certain way, whatever you put into it, you're going to get an answer from it, which is ideologically consistent with what the original program writers had in mind.
00:09:20 That has the potential to change public opinion, to, to change where the center of gravity of public opinion is, if you will, because whatever you ask it, it's always going to give you an answer from a certain perspective.
00:09:35 If I understand you correctly, that's what the risk is.
00:09:40 How scared do you think we should be that that is what is going to happen?
00:09:45 Well, yeah, the ChatGPT is programmed and I, I hate to use this word program because that's not, uh, the way traditional programming works.
00:09:59 Uh, but it's trained by its trainers and obviously it reflects the trainer's biases and, uh, the information with which it's, it's fed, um, is filtered and cleaned out.
00:10:14 In a certain way, uh, to make sure that the ChatGPT does what it's supposed to do.
00:10:23 So ultimately what it leads to is that we don't really have an intelligent agent, uh, which is, which is very similar to human brain.
00:10:37 Uh, where, where, where we, um, really absorb the information from the, our experience.
00:10:43 We make connections, logical connections.
00:10:45 We change our minds and we learn on the fly from experiences.
00:10:49 That's not the case with ChatGPT.
00:10:52 Uh, uh, once it, it, it has that knowledge and then biases of, of its trainers.
00:10:59 I see.
00:11:00 There's, there's only, um, it's just, just the way the technology works is not the same as humans.
00:11:09 So it's very difficult to compare.
00:11:11 Uh, and, um, yeah, so everything kind of follows from that point.
00:11:17 So what do you think this is going to do to us over a period of years?
00:11:22 Us.
00:11:23 Society.
00:11:24 Now from the, from the societal point of view, obviously there are, there are, I, I mentioned in my article what the, the repercussions are.
00:11:33 Some obvious ones is that, uh, some jobs are going to be jeopardized obviously, right?
00:11:38 So there's, there's many jobs that are not very creative, uh, that are based on, um, uh, sort of the same templatized responses, um, that you, you grab your knowledge from, from some existing knowledge base and you spin off, spin it off in a, in a language.
00:11:54 So it's concerning the writers.
00:11:56 So, you know, it can be healthcare, can be education, uh, programming is another area.
00:12:02 So, uh, those jobs, uh, are certainly in jeopardy, right?
00:12:06 But the, the other societal repercussions is that, uh, if we think of HRGPD as something that is intelligent in the sense that it has knowledge and ability to logically think, then it can be very misleading for, uh, for people to ask questions and rely on those answers.
00:12:25 Uh, because they come with, they come very, they come out very easily, right?
00:12:30 And then, and they can, the people can be very easily fooled by, uh, by seemingly knowledgeable agent, which isn't really, which doesn't have any knowledge.
00:12:42 And it's just, uh, processes the inputs into the outputs based on, uh, whatever ideological views for, or political views that trainers had.
00:12:53 Do you think it could change the way we vote?
00:12:57 Sorry?
00:12:58 Do you think it could change the way we vote?
00:13:01 The way you vote?
00:13:03 Yes.
00:13:04 Oh, absolutely.
00:13:07 You can, you can, you can, uh, uh, it can change the way people vote, uh, simply because if, if someone is trying to decide to vote and goes and asks questions to chat GPT, and the chat GPT has certain political biases reflected of, of its trainers, it will, uh, it will be very convincing.
00:13:27 It just kind of, if you, if you, if I go back to my, uh, example of, uh, convincing chat GPT to change its mind on, uh, on the vaccines, right?
00:13:37 Regardless of what the answers were from chat GPT, it, it just keeps repeating the same mantra about how vaccines are safe and effective.
00:13:47 It, it, regardless, regardless of what conclusions we achieved, what, uh, what is the current reasoning, it would still repeat the same mantra as the vaccine is safe and effective.
00:13:59 It's, it's totally brainwashed minds in this respect.
00:14:03 So, Gleb, I understand that, uh, you were actually born in, uh, Russia.
00:14:08 Yeah.
00:14:09 Um, were you actually born in the Soviet Union?
00:14:12 Yes, I was.
00:14:13 So, you would have, uh, almost an instinctive, uh, feel for how people's minds are, uh, can be changed and how information can be handled and controlled.
00:14:25 So, I'm going to ask, I'm going to put the whole thing in one sentence and ask you to agree or to disagree.
00:14:32 We're saying that artificial intelligence is not nearly as intelligent as we think it is, but it may be more sinister than we ever thought possible.
00:14:43 Agree or disagree?
00:14:46 I agree that the artificial intelligence, as we know it right now, which is in, in the form of generative AI, is not intelligent at all, right?
00:14:58 And when I say intelligent, I mean, uh, the ability to reason, to have a knowledge base, to have a memory and the ability to change its mind.
00:15:09 So that's, that's just some key, key pieces to the intelligence besides just recognizing patterns, which ChatGPT does very well.
00:15:17 So, um, no, it's, um, as much as we think of AI, if you want me to put it in a, in one sentence, as much as we think of AI as a, as a, uh, almost like a replacement or enhancement to a human brain.
00:15:36 And the current version of journey of AI and its instance, ChatGPT, for example, is not anywhere close and, and arguably is not even designed to do that.
00:15:49 Well, one thing, Gleb, it is certainly much more than a spell checker.
00:15:53 Gleb Lizzyk, from Vaughan, an IT specialist, with us today to talk about artificial intelligence and where it may lead us.
00:16:03 Gleb, thank you very much for coming on the program. We much appreciate it.
00:16:06 Thank you, Nigel. Have a good day.
00:16:08 On behalf of the Western Standard, I'm Nigel Hannaford.
00:16:12 Thank you.