1 00:00:04,440 --> 00:00:12,399 Speaker 1: Welcome to tech Stuff, a production from iHeartRadio. Hey there, 2 00:00:12,440 --> 00:00:16,240 Speaker 1: and welcome to tech Stuff. I'm your host, Jonathan Strickland. 3 00:00:16,280 --> 00:00:19,760 Speaker 1: I'm an executive producer with iHeartRadio and how the tech 4 00:00:19,800 --> 00:00:23,480 Speaker 1: are you. At the beginning of this year, that being 5 00:00:23,920 --> 00:00:26,360 Speaker 1: twenty twenty three, I said like it felt like it 6 00:00:26,400 --> 00:00:28,520 Speaker 1: was going to be the year of AI, and so 7 00:00:28,600 --> 00:00:31,319 Speaker 1: far I think I'm pretty much on the money. But 8 00:00:31,800 --> 00:00:35,680 Speaker 1: more specifically, twenty twenty three has been the year of 9 00:00:36,200 --> 00:00:43,320 Speaker 1: generative AI. That is artificial intelligence that creates or generates something, 10 00:00:43,600 --> 00:00:47,199 Speaker 1: whether it's an image, a sound, or as we're going 11 00:00:47,280 --> 00:00:51,680 Speaker 1: to talk about today, text in response to some sort 12 00:00:51,840 --> 00:00:55,480 Speaker 1: of input. Now, before we go any further, this is 13 00:00:55,480 --> 00:00:57,800 Speaker 1: where we need to remind ourselves that while this is 14 00:00:57,920 --> 00:01:03,120 Speaker 1: a type of artificial intelligence, it's not all of AI. 15 00:01:03,760 --> 00:01:09,600 Speaker 1: Not every AI application involves generative processes. And while generative 16 00:01:09,600 --> 00:01:15,479 Speaker 1: AI can seem fascinating, exciting, surprising, or creepy, I believe 17 00:01:15,520 --> 00:01:19,520 Speaker 1: that largely stems from how generative AI appears to be 18 00:01:19,680 --> 00:01:26,080 Speaker 1: mimicking humans, and it's not an indication of how sophisticated, advanced, 19 00:01:26,240 --> 00:01:30,680 Speaker 1: or dangerous it really is. It's kind of an uncanny 20 00:01:31,000 --> 00:01:35,679 Speaker 1: Valley thing because it appears to be behaving like a human, 21 00:01:36,360 --> 00:01:41,720 Speaker 1: we start to project things on it that aren't necessarily 22 00:01:42,400 --> 00:01:45,319 Speaker 1: accurate or realistic. I think of it kind of like 23 00:01:45,520 --> 00:01:48,360 Speaker 1: the way we can be with our pets, where we 24 00:01:48,400 --> 00:01:51,360 Speaker 1: will project things on our pets that may not reflect 25 00:01:51,440 --> 00:01:54,800 Speaker 1: what the pet is actually experiencing, but that's how we're 26 00:01:54,840 --> 00:01:58,040 Speaker 1: perceiving it. So the reason I say all of this 27 00:01:58,160 --> 00:01:59,840 Speaker 1: up at the very top of this episode is that 28 00:01:59,840 --> 00:02:04,760 Speaker 1: we're also seeing a lot of people expressing concern about AI, 29 00:02:05,000 --> 00:02:09,239 Speaker 1: which is understandable. You know about how it could potentially 30 00:02:09,360 --> 00:02:15,959 Speaker 1: lead to harm, and these are legitimate and rational concerns. However, 31 00:02:16,560 --> 00:02:21,160 Speaker 1: with the focus on stuff like chat GPT for example, 32 00:02:21,320 --> 00:02:25,280 Speaker 1: or Google Bard, I would argue the concern is far 33 00:02:25,360 --> 00:02:29,800 Speaker 1: too narrowly focused on just one aspect of AI, and 34 00:02:30,040 --> 00:02:34,200 Speaker 1: in my opinion, it's not even the most dangerous implementation 35 00:02:34,280 --> 00:02:38,280 Speaker 1: of AI. I mean, we have cars on the road 36 00:02:38,720 --> 00:02:44,040 Speaker 1: right now that use AI for driver assists and autonomous operations. 37 00:02:44,520 --> 00:02:47,800 Speaker 1: If we're worried about the robots taking us down, maybe 38 00:02:47,840 --> 00:02:51,639 Speaker 1: we shouldn't make them our chauffeurs. But really that's a 39 00:02:51,680 --> 00:02:55,120 Speaker 1: topic for another episode. Today, I wanted to take a 40 00:02:55,160 --> 00:02:58,280 Speaker 1: look at an issue that crops up in AI chat 41 00:02:58,320 --> 00:03:01,960 Speaker 1: bots like open ai or goole Bard and similar products. 42 00:03:02,560 --> 00:03:05,560 Speaker 1: This is one that is concerning because it's an issue 43 00:03:05,560 --> 00:03:09,800 Speaker 1: that leads these tools to create false or misleading information 44 00:03:10,360 --> 00:03:13,920 Speaker 1: while presenting that info in a way that seems authoritative 45 00:03:14,000 --> 00:03:17,680 Speaker 1: and trustworthy. And in the field of AI, the term 46 00:03:18,000 --> 00:03:22,640 Speaker 1: hallucination is used to describe this situation. At least a 47 00:03:22,680 --> 00:03:25,400 Speaker 1: lot of folks will use the word hallucination. As it 48 00:03:25,440 --> 00:03:29,160 Speaker 1: turns out, there's actually some debate in AI circles about 49 00:03:29,160 --> 00:03:32,799 Speaker 1: whether or not that should be the appropriate term. Now 50 00:03:33,200 --> 00:03:36,960 Speaker 1: for we mirror mortals, a hallucination is when we have 51 00:03:37,000 --> 00:03:42,360 Speaker 1: an experience in which we perceive something that isn't reflected 52 00:03:42,520 --> 00:03:46,480 Speaker 1: in reality. Maybe we hear a sound but there was 53 00:03:46,520 --> 00:03:50,640 Speaker 1: actually no sound present. Maybe it was that tree falling 54 00:03:50,680 --> 00:03:53,120 Speaker 1: in the woods and no one was around or something, 55 00:03:53,720 --> 00:03:57,680 Speaker 1: or we see something that's not really there. It can 56 00:03:57,720 --> 00:04:03,320 Speaker 1: be really darn disconcerting, and sometimes it can be absolutely terrifying. 57 00:04:03,680 --> 00:04:07,760 Speaker 1: I'm reminded of how many people who experience sleep paralysis 58 00:04:07,800 --> 00:04:13,880 Speaker 1: often will also have hallucinations accompany this period where they're 59 00:04:13,960 --> 00:04:17,599 Speaker 1: awake but they cannot move, and it's probably because sleep 60 00:04:17,600 --> 00:04:21,640 Speaker 1: paralysis occurs when you're kind of caught between being asleep 61 00:04:21,800 --> 00:04:25,120 Speaker 1: and being awake, so there's still some dream like activity 62 00:04:25,160 --> 00:04:28,760 Speaker 1: going on in your brain that's trying to explain things 63 00:04:28,839 --> 00:04:31,719 Speaker 1: like why you're unable to move. Oh, it's because you 64 00:04:31,839 --> 00:04:35,960 Speaker 1: have this witch perched on your chest and she's pinning 65 00:04:36,000 --> 00:04:41,200 Speaker 1: you to the bed. Tools like chat GPT are not dreaming, 66 00:04:41,680 --> 00:04:45,279 Speaker 1: you know, they're not perceiving anything at all. They have 67 00:04:46,040 --> 00:04:52,160 Speaker 1: no senses to trigger, so they cannot hallucinate in that sense. Instead, 68 00:04:52,240 --> 00:04:57,440 Speaker 1: what they are doing is mistakenly assigning high confidence to 69 00:04:57,560 --> 00:05:01,200 Speaker 1: something that they just plane made up. So they're treating 70 00:05:01,200 --> 00:05:05,919 Speaker 1: it like it's a fact that they're highly confident is accurate, 71 00:05:06,440 --> 00:05:11,000 Speaker 1: when really they just invented it. So it is an 72 00:05:11,000 --> 00:05:14,640 Speaker 1: instance where they're really confident in something that is not 73 00:05:15,279 --> 00:05:19,520 Speaker 1: coming from a reliable source in the AI's actual training data. 74 00:05:20,040 --> 00:05:22,200 Speaker 1: So if we wanted to put that into human terms, 75 00:05:22,640 --> 00:05:25,040 Speaker 1: it'd be kind of like if you made up a 76 00:05:25,120 --> 00:05:29,040 Speaker 1: story to explain something that otherwise would either be really 77 00:05:29,080 --> 00:05:32,720 Speaker 1: boring or maybe really embarrassing. So you make up a lie, 78 00:05:33,160 --> 00:05:35,080 Speaker 1: in other words, to cover up something that you would 79 00:05:35,160 --> 00:05:38,640 Speaker 1: rather not be known, And so you tell this lie 80 00:05:39,120 --> 00:05:41,400 Speaker 1: over and over when people are asking you about this 81 00:05:41,440 --> 00:05:45,920 Speaker 1: particular thing, and you repeat it often enough where gradually 82 00:05:45,960 --> 00:05:49,560 Speaker 1: your brain essentially makes a pathway where this fake version 83 00:05:49,560 --> 00:05:53,760 Speaker 1: of history of what actually happened becomes the real one 84 00:05:54,279 --> 00:05:56,960 Speaker 1: in your head. You begin to believe your own lie, 85 00:05:57,080 --> 00:05:59,680 Speaker 1: and so in future tellings of the story, you don't 86 00:05:59,680 --> 00:06:02,240 Speaker 1: even realize you're lying at all. You're telling what you 87 00:06:02,360 --> 00:06:05,440 Speaker 1: believe to be the real sequence of events, even though 88 00:06:05,440 --> 00:06:08,800 Speaker 1: it's all a fib. That's kind of what's happening with 89 00:06:08,839 --> 00:06:14,039 Speaker 1: AI hallucinations, only it happens all at once, And for 90 00:06:14,120 --> 00:06:17,440 Speaker 1: that reason, some folks prefer to use other terms to 91 00:06:17,520 --> 00:06:21,640 Speaker 1: describe what AI does when it starts to invent things 92 00:06:21,760 --> 00:06:25,520 Speaker 1: in response to a query from a user. So some 93 00:06:25,600 --> 00:06:30,840 Speaker 1: have proposed the word confabulation as an alternative descriptor of 94 00:06:30,839 --> 00:06:33,640 Speaker 1: what's going on. So this is similar to kind of 95 00:06:33,680 --> 00:06:38,240 Speaker 1: the scenario I just gave, because it's in human psychology. 96 00:06:38,240 --> 00:06:41,760 Speaker 1: A confabulation is when we have a hitch in our memory, 97 00:06:42,160 --> 00:06:44,680 Speaker 1: and so we fill in a gap that's in our memory. 98 00:06:44,680 --> 00:06:47,840 Speaker 1: We're not doing it consciously, it just happens, and that 99 00:06:47,920 --> 00:06:49,720 Speaker 1: might mean we fill in the gap that doesn't at 100 00:06:49,760 --> 00:06:53,400 Speaker 1: all reflect what really happened. So this can happen at 101 00:06:53,440 --> 00:06:56,560 Speaker 1: any time. I've seen it happen with people who are 102 00:06:56,640 --> 00:07:00,360 Speaker 1: in like a situation that was totally on a expected 103 00:07:00,400 --> 00:07:03,600 Speaker 1: in high stress. I've seen it in training operations where 104 00:07:04,120 --> 00:07:07,560 Speaker 1: you have a group of people and then someone bursts 105 00:07:07,640 --> 00:07:11,200 Speaker 1: in as if they are a burglar or a thief or something, 106 00:07:11,560 --> 00:07:14,720 Speaker 1: and then they get out, and then those people who 107 00:07:14,760 --> 00:07:18,840 Speaker 1: were just subjected to this very scary situation are asked 108 00:07:18,880 --> 00:07:22,200 Speaker 1: to give details about the thief's appearance, and people start 109 00:07:22,240 --> 00:07:26,960 Speaker 1: to invent things, not purposefully, not with the intent to deceive, 110 00:07:27,400 --> 00:07:29,440 Speaker 1: but because their memory is just trying to fill in 111 00:07:29,480 --> 00:07:32,760 Speaker 1: gaps because their perception didn't really take it all in. 112 00:07:33,400 --> 00:07:37,440 Speaker 1: So confabulation doesn't imply intent, and I think that might 113 00:07:37,480 --> 00:07:40,320 Speaker 1: be why a lot of researchers like the word, because 114 00:07:40,880 --> 00:07:44,840 Speaker 1: it's not the intention of the AI to fool people 115 00:07:45,320 --> 00:07:49,800 Speaker 1: or to pass off fantasy as if it were reality. Instead, 116 00:07:50,120 --> 00:07:53,320 Speaker 1: the AI is making an honest go of trying to 117 00:07:53,360 --> 00:07:56,240 Speaker 1: meet the expectations of the user. So if you ask 118 00:07:56,320 --> 00:08:01,000 Speaker 1: the AI about, say a historical figure, really tries to 119 00:08:01,000 --> 00:08:04,720 Speaker 1: give you a good answer, but occasionally that answer might 120 00:08:04,760 --> 00:08:08,240 Speaker 1: be wrong, not because the AI is drawing from a 121 00:08:08,360 --> 00:08:12,000 Speaker 1: bad data source, but because there's actually a gap in 122 00:08:12,040 --> 00:08:15,200 Speaker 1: its knowledge, and the AI just fills that gap as 123 00:08:15,240 --> 00:08:19,040 Speaker 1: best it can. Unfortunately, the end result is you get 124 00:08:19,080 --> 00:08:23,200 Speaker 1: an answer that seems totally cromulent, like you could just 125 00:08:23,280 --> 00:08:28,880 Speaker 1: imagine reading that answer in a respectable, thoroughly fact check encyclopedia, 126 00:08:29,360 --> 00:08:33,560 Speaker 1: but then it turns out to be garbage. So let's 127 00:08:33,600 --> 00:08:36,800 Speaker 1: talk about how this happens, which will involve an overview 128 00:08:36,880 --> 00:08:39,920 Speaker 1: of how these chatbought AI tools are trained and at 129 00:08:39,960 --> 00:08:42,800 Speaker 1: a very very high level, how they work. So this 130 00:08:42,880 --> 00:08:48,480 Speaker 1: is going to involve some discussion about machine learning and statistics. So, 131 00:08:48,559 --> 00:08:54,080 Speaker 1: first off, how do machines actually learn? I think it's 132 00:08:54,120 --> 00:08:57,679 Speaker 1: pretty easy to understand. How we program machines to do 133 00:08:58,600 --> 00:09:01,959 Speaker 1: some specific task. Right, we create a set of rules 134 00:09:02,400 --> 00:09:07,200 Speaker 1: that this machine follows sequentially, and the machine executes those 135 00:09:07,320 --> 00:09:10,720 Speaker 1: rules as directed, and then we get the result we wanted. 136 00:09:10,800 --> 00:09:13,920 Speaker 1: That is easy to understand. So I'll give an example. 137 00:09:13,960 --> 00:09:16,439 Speaker 1: Let's say we have a robotic arm and you've got 138 00:09:16,480 --> 00:09:19,200 Speaker 1: two tables, and you put a wooden block on table 139 00:09:19,320 --> 00:09:23,240 Speaker 1: number one, and you program the robotic arm to pick 140 00:09:23,360 --> 00:09:26,320 Speaker 1: up this wooden block on table one and move it 141 00:09:26,360 --> 00:09:29,640 Speaker 1: over to table two. Once you program it then it 142 00:09:29,679 --> 00:09:31,720 Speaker 1: should be able to do that task over and over, 143 00:09:31,880 --> 00:09:34,880 Speaker 1: assuming that no one has moved the tables. No one 144 00:09:34,920 --> 00:09:37,560 Speaker 1: has moved the robotic arm, and the wooden block is 145 00:09:37,640 --> 00:09:41,360 Speaker 1: always in the same place and it's always the same size. Right, 146 00:09:41,400 --> 00:09:43,560 Speaker 1: you haven't changed any of the parameters, so it's the 147 00:09:43,600 --> 00:09:46,400 Speaker 1: exact same situation over and over and over again. You've 148 00:09:46,400 --> 00:09:49,640 Speaker 1: created this simple program. It should be no surprise when 149 00:09:49,640 --> 00:09:52,920 Speaker 1: the robotic arm does it successfully. But what if we 150 00:09:52,960 --> 00:09:55,920 Speaker 1: wanted a robotic arm that could learn how to pick 151 00:09:56,000 --> 00:09:59,720 Speaker 1: up different objects from table one and then move them 152 00:09:59,760 --> 00:10:02,800 Speaker 1: to t able to These objects could be different shapes, 153 00:10:02,800 --> 00:10:05,880 Speaker 1: they could be different sizes, they could weigh different amounts, 154 00:10:06,120 --> 00:10:08,560 Speaker 1: They might be made of different stuff. Maybe some of 155 00:10:08,559 --> 00:10:12,760 Speaker 1: them are fairly delicate and the arm would break the 156 00:10:12,840 --> 00:10:15,680 Speaker 1: object if it applied too much pressure. So how would 157 00:10:15,720 --> 00:10:18,640 Speaker 1: we build a robotic arm that could deal with these 158 00:10:18,679 --> 00:10:23,439 Speaker 1: different scenarios, including ones where we put something completely new 159 00:10:23,520 --> 00:10:26,640 Speaker 1: to the robot on the table, something that the robot 160 00:10:26,679 --> 00:10:31,240 Speaker 1: has never encountered before. Well, to do that, we would 161 00:10:31,320 --> 00:10:36,320 Speaker 1: probably pursue a machine learning model in order to teach 162 00:10:36,440 --> 00:10:41,480 Speaker 1: this robot the whole process of picking something up, especially 163 00:10:41,520 --> 00:10:45,840 Speaker 1: something it had not encountered before. So basically, machine learning 164 00:10:46,080 --> 00:10:49,120 Speaker 1: uses sets of algorithms in an effort to get better 165 00:10:49,600 --> 00:10:54,560 Speaker 1: at a given task, and part of learning involves training, 166 00:10:54,600 --> 00:10:59,080 Speaker 1: which really boils down to feeding a machine lots and 167 00:10:59,160 --> 00:11:02,480 Speaker 1: lots and lots of information, like the more information you 168 00:11:02,520 --> 00:11:07,320 Speaker 1: can feed it, the better, and then letting it process 169 00:11:07,360 --> 00:11:10,520 Speaker 1: this information in an effort to get a specific result, 170 00:11:11,080 --> 00:11:15,280 Speaker 1: and then going back and tweaking the model to refine 171 00:11:15,320 --> 00:11:18,280 Speaker 1: it over and over and over and over again to 172 00:11:18,480 --> 00:11:22,360 Speaker 1: get better at it over time. So we'll imagine a 173 00:11:22,440 --> 00:11:26,480 Speaker 1: hypothetical machine learning model that is designed to do something 174 00:11:26,559 --> 00:11:30,760 Speaker 1: relatively simple like recognize if an image has a cat 175 00:11:31,040 --> 00:11:33,760 Speaker 1: in it or not, because this is actually something that 176 00:11:33,880 --> 00:11:36,760 Speaker 1: has been done with machine learning models in the past. 177 00:11:36,960 --> 00:11:40,960 Speaker 1: It's actually a fairly popular approach is does this picture 178 00:11:40,960 --> 00:11:43,200 Speaker 1: have a cat in it? Or does this video have 179 00:11:43,240 --> 00:11:45,920 Speaker 1: a cat in it? That kind of thing. Let's imagine 180 00:11:46,000 --> 00:11:49,520 Speaker 1: that our machine learning model is an actual physical model, 181 00:11:49,679 --> 00:11:53,319 Speaker 1: like it's a giant funnel. So on the wide end 182 00:11:53,320 --> 00:11:56,360 Speaker 1: of the funnel, that's where we just dump tons of 183 00:11:56,400 --> 00:11:59,120 Speaker 1: photographs with some of them have cats in them, some 184 00:11:59,160 --> 00:12:02,560 Speaker 1: of them don't. Now imagine that at the narrow end 185 00:12:02,600 --> 00:12:05,160 Speaker 1: of the funnel. At the bottom of the funnel, we 186 00:12:05,200 --> 00:12:09,200 Speaker 1: actually have two channels. One channel leads into a bucket 187 00:12:09,440 --> 00:12:13,160 Speaker 1: that says no cats here, and the other channel leads 188 00:12:13,160 --> 00:12:16,920 Speaker 1: to a bucket that says, ah, sweet kitty cats. So 189 00:12:17,520 --> 00:12:23,280 Speaker 1: we dump thousands, maybe millions of photographs into the top 190 00:12:23,320 --> 00:12:27,280 Speaker 1: of this funnel, and the funnel starts to sort the pictures. 191 00:12:27,640 --> 00:12:30,080 Speaker 1: We can't see this because it's inside the funnel, but 192 00:12:30,120 --> 00:12:35,119 Speaker 1: there are channels inside that funnel where photos are directed 193 00:12:35,559 --> 00:12:40,240 Speaker 1: either to go more toward the no kittycat side or 194 00:12:40,280 --> 00:12:44,560 Speaker 1: the yes kittykat side, And they go through these channels 195 00:12:44,640 --> 00:12:47,200 Speaker 1: all down the funnel, and at the very end of it, 196 00:12:47,880 --> 00:12:51,480 Speaker 1: they start spitting out these images into the two buckets. Well, 197 00:12:51,480 --> 00:12:54,520 Speaker 1: once it's done, once it has processed all the photos, 198 00:12:54,760 --> 00:12:56,719 Speaker 1: we take the two buckets and we see how our 199 00:12:56,760 --> 00:12:59,520 Speaker 1: model did. And maybe we see that the model caught 200 00:12:59,679 --> 00:13:02,240 Speaker 1: most of the pictures with cats in them, but not 201 00:13:02,360 --> 00:13:05,280 Speaker 1: all of them. Maybe we also see that there are 202 00:13:05,320 --> 00:13:08,160 Speaker 1: some photos that fell into the kitty cat bucket that 203 00:13:08,280 --> 00:13:12,440 Speaker 1: have exactly zero kitty cats in the picture. Something is 204 00:13:12,480 --> 00:13:15,760 Speaker 1: not working inside our model. So at that point we 205 00:13:15,920 --> 00:13:19,800 Speaker 1: open the funnel, we take the top off or whatever 206 00:13:19,880 --> 00:13:23,280 Speaker 1: we have built in a hinged latch or something, and 207 00:13:23,320 --> 00:13:26,480 Speaker 1: we've opened it up. Now essentially inside our funnel, we 208 00:13:26,520 --> 00:13:29,320 Speaker 1: see all those channels, and each channel is meant to 209 00:13:29,320 --> 00:13:31,280 Speaker 1: look for some sort of evidence of a cat, and 210 00:13:31,320 --> 00:13:34,520 Speaker 1: if it finds evidence, it pushes it closer toward the 211 00:13:34,559 --> 00:13:37,760 Speaker 1: pathway of kitty cat, and if it doesn't, it pushes 212 00:13:37,800 --> 00:13:41,640 Speaker 1: it closer to the pathway of no kitty cat. But 213 00:13:41,720 --> 00:13:44,520 Speaker 1: there's tons of these channels. Some of them feed images 214 00:13:44,640 --> 00:13:48,920 Speaker 1: back up through the whole process. Again, it's very complicated 215 00:13:48,920 --> 00:13:52,400 Speaker 1: inside this funnel, and you have to go in there 216 00:13:52,440 --> 00:13:57,120 Speaker 1: and start to tweak little bits of rules in these 217 00:13:57,720 --> 00:14:01,880 Speaker 1: channels to adjust for whatever problem you're encountering at the 218 00:14:01,960 --> 00:14:05,679 Speaker 1: end result when you're done. So, when you're training your model, 219 00:14:06,080 --> 00:14:11,000 Speaker 1: you change the weights of these different decisions that are made. 220 00:14:11,080 --> 00:14:14,520 Speaker 1: Some decisions perhaps have too much emphasis on them. They 221 00:14:14,720 --> 00:14:18,040 Speaker 1: like they're too powerful and they're skewing the results. So 222 00:14:18,120 --> 00:14:22,760 Speaker 1: you reduce the weight of that particular decision point and 223 00:14:22,800 --> 00:14:25,000 Speaker 1: you increase the weight of a different one to try 224 00:14:25,040 --> 00:14:28,560 Speaker 1: and get things right. It's a painstaking process and you 225 00:14:28,600 --> 00:14:30,800 Speaker 1: have to do it over and over again, and these 226 00:14:30,840 --> 00:14:35,200 Speaker 1: exercises repeat and you try to refine your model to 227 00:14:35,240 --> 00:14:38,440 Speaker 1: get it better at deciding whether or not a photograph 228 00:14:39,320 --> 00:14:41,880 Speaker 1: has got a cabinet or does it, and eventually, if 229 00:14:41,920 --> 00:14:44,880 Speaker 1: everything is working well, it gets very very good at 230 00:14:44,920 --> 00:14:47,920 Speaker 1: sorting images. Maybe once in a while, something sneaks through. 231 00:14:48,160 --> 00:14:50,000 Speaker 1: Maybe there's a cloud that kind of looks like a 232 00:14:50,040 --> 00:14:52,440 Speaker 1: kitty cat and it goes into the wrong bucket, or 233 00:14:52,600 --> 00:14:54,760 Speaker 1: maybe there is a kitty cat that goes into the 234 00:14:54,800 --> 00:14:56,720 Speaker 1: no kitty cat bucket, but the kitty cat was kind 235 00:14:56,720 --> 00:14:59,240 Speaker 1: of obscured in the picture and the model just couldn't 236 00:14:59,240 --> 00:15:04,480 Speaker 1: suss it out. But it succeeds more often than not. Okay, 237 00:15:04,840 --> 00:15:06,880 Speaker 1: that's a baseline. When we come back, we'll talk a 238 00:15:06,880 --> 00:15:09,440 Speaker 1: bit more about machine learning and how this plays into 239 00:15:09,800 --> 00:15:24,200 Speaker 1: tools like chat GPT. Okay, I laid out one version 240 00:15:24,400 --> 00:15:27,040 Speaker 1: of machine learning, and I want to stress that's just 241 00:15:27,320 --> 00:15:30,520 Speaker 1: one version of machine learning. It's related to things like 242 00:15:30,600 --> 00:15:34,920 Speaker 1: neural networks, which are designed to kind of mimic the 243 00:15:34,960 --> 00:15:40,440 Speaker 1: way our brains process information and form pathways among neurons 244 00:15:40,480 --> 00:15:43,920 Speaker 1: while we're trying to suss things out. But that's just 245 00:15:44,080 --> 00:15:46,400 Speaker 1: one version of machine learning. I don't mean to say 246 00:15:46,440 --> 00:15:49,560 Speaker 1: that's how it all works. There are actually lots of 247 00:15:49,720 --> 00:15:53,240 Speaker 1: sub fields within machine learning, neural networks being just one 248 00:15:53,280 --> 00:15:55,960 Speaker 1: of them, but there's also subsets of neural networks. One 249 00:15:56,000 --> 00:16:00,000 Speaker 1: of those was would be deep learning, which always makes 250 00:16:00,040 --> 00:16:03,000 Speaker 1: we think of MST three K and deep hurting shout 251 00:16:03,000 --> 00:16:05,760 Speaker 1: outs to any misties out there. Now, as you dive 252 00:16:05,880 --> 00:16:09,960 Speaker 1: down to deep learning, you're really getting into an interesting 253 00:16:10,040 --> 00:16:13,480 Speaker 1: field of AI and machine learning. So deep learning models 254 00:16:13,720 --> 00:16:17,640 Speaker 1: can accept unstructured data. If you're going further up to 255 00:16:17,800 --> 00:16:22,920 Speaker 1: less specialized machine learning models, these have to use heavily 256 00:16:23,040 --> 00:16:27,119 Speaker 1: labeled data sets and heavily structured data and use supervised 257 00:16:27,160 --> 00:16:30,640 Speaker 1: learning in order to improve with time. But when you 258 00:16:30,680 --> 00:16:34,160 Speaker 1: get into deep learning, you're looking at a very focused 259 00:16:34,160 --> 00:16:38,440 Speaker 1: approach to machine learning where you can just feed unstructured 260 00:16:38,520 --> 00:16:41,960 Speaker 1: data that has no labels to it and start to 261 00:16:42,080 --> 00:16:45,000 Speaker 1: use this model to do whatever it is that you 262 00:16:45,640 --> 00:16:48,320 Speaker 1: want it to do. But we're still kind of talking 263 00:16:48,320 --> 00:16:53,080 Speaker 1: about a channeling or funneling situation here. The input goes 264 00:16:53,120 --> 00:16:56,720 Speaker 1: into the model, the model analyzes the input and pushes 265 00:16:56,760 --> 00:16:59,360 Speaker 1: it further one way or another through the system, and 266 00:16:59,400 --> 00:17:02,520 Speaker 1: it comes out the end as output, which could be 267 00:17:02,560 --> 00:17:05,280 Speaker 1: an image search result for kiddy cats in your smartphone's 268 00:17:05,320 --> 00:17:07,760 Speaker 1: photo role, for example, So if you've ever gone into 269 00:17:08,600 --> 00:17:12,200 Speaker 1: a smartphone photo collection and you just typed in a 270 00:17:13,040 --> 00:17:15,800 Speaker 1: general word in search, you know it's not that you 271 00:17:15,880 --> 00:17:18,040 Speaker 1: tagged any of your photos with this. You're just like 272 00:17:18,160 --> 00:17:20,280 Speaker 1: looking for photos in your role that has a cat 273 00:17:20,359 --> 00:17:24,280 Speaker 1: in them, and it returns something like that. Well, that 274 00:17:24,280 --> 00:17:27,600 Speaker 1: can be the result of a machine learning process like 275 00:17:27,640 --> 00:17:31,000 Speaker 1: the one I've just described, because again, the system has 276 00:17:31,000 --> 00:17:33,600 Speaker 1: to figure out which of your photos have cats in them, 277 00:17:33,840 --> 00:17:36,560 Speaker 1: even though you didn't tag any of those photos with cats. 278 00:17:36,560 --> 00:17:39,880 Speaker 1: It doesn't have metadata. It has to analyze the photo itself. 279 00:17:40,480 --> 00:17:46,040 Speaker 1: Now it's time to talk about probabilities. Large language models lms, 280 00:17:46,520 --> 00:17:51,360 Speaker 1: which are what power chat bots like Google Bard and 281 00:17:51,880 --> 00:17:57,320 Speaker 1: Chat GPT. They work in probabilities. And there's one example 282 00:17:57,359 --> 00:18:01,199 Speaker 1: of an AI using probabilistic algorithms to generate responses that 283 00:18:01,280 --> 00:18:06,760 Speaker 1: I really loved reference, and that example is IBM's Watson platform. 284 00:18:07,440 --> 00:18:09,760 Speaker 1: So while the world right now is struggling to figure 285 00:18:09,760 --> 00:18:13,280 Speaker 1: out how to handle chat GPT and Google Bard and such, 286 00:18:13,720 --> 00:18:16,520 Speaker 1: IBM's Watson gave us a glimpse at what we could 287 00:18:16,560 --> 00:18:20,080 Speaker 1: expect all the way back in twenty eleven. That's when 288 00:18:20,119 --> 00:18:24,240 Speaker 1: IBM famously put Watson to the test and some exhibition 289 00:18:24,400 --> 00:18:29,199 Speaker 1: games of the game show Jeopardy against former champions of 290 00:18:29,280 --> 00:18:33,280 Speaker 1: that game show, human champions. So in many ways, this 291 00:18:33,400 --> 00:18:36,800 Speaker 1: was an echo of IBM's Deep Blue going up against 292 00:18:36,960 --> 00:18:41,879 Speaker 1: chess master Gary Kasparov in various games of chess. Putting 293 00:18:41,960 --> 00:18:46,000 Speaker 1: Watson up against humans and Jeopardy was a fantastic publicity stunt, 294 00:18:46,160 --> 00:18:49,000 Speaker 1: and it also was really impressive because the way Jeopardy 295 00:18:49,040 --> 00:18:53,719 Speaker 1: works is players get several categories of trivia that they 296 00:18:53,720 --> 00:18:57,440 Speaker 1: can choose from. Each category has different levels of questions 297 00:18:57,480 --> 00:19:00,960 Speaker 1: that are designated by a dollar amount, So higher the 298 00:19:01,000 --> 00:19:04,400 Speaker 1: dollar amount, the harder the trivia question is. Generally speaking, 299 00:19:05,520 --> 00:19:09,040 Speaker 1: the actual clue that the players get is given in 300 00:19:09,040 --> 00:19:11,720 Speaker 1: the form of an answer, and they have to provide 301 00:19:12,480 --> 00:19:17,440 Speaker 1: a question that relates to that answer. So here's an example. 302 00:19:17,880 --> 00:19:22,320 Speaker 1: The answer revealed in say a hypothetical Jeopardy game that 303 00:19:22,359 --> 00:19:26,119 Speaker 1: has the category podcasts, could be something like he was 304 00:19:26,240 --> 00:19:29,600 Speaker 1: Jonathan Strickland's original co host on the show tech Stuff. 305 00:19:30,000 --> 00:19:32,399 Speaker 1: The correct response would be bipp a bip Who is 306 00:19:32,480 --> 00:19:36,200 Speaker 1: Chris Palette? That would be the correct answer, But Jeopardy 307 00:19:36,680 --> 00:19:42,040 Speaker 1: goes beyond just trivia. Often the answers provided will include 308 00:19:42,359 --> 00:19:46,760 Speaker 1: word play or images or sound cues, and players will 309 00:19:46,760 --> 00:19:49,879 Speaker 1: have to think outside the box. They can't just know 310 00:19:50,160 --> 00:19:54,800 Speaker 1: the answer. Sometimes there's interpretation that has to happen first. 311 00:19:55,359 --> 00:19:58,399 Speaker 1: The clue to the correct response could be a pun, 312 00:19:58,960 --> 00:20:02,600 Speaker 1: it could involve a rhyme to the answer. It's not 313 00:20:02,760 --> 00:20:07,320 Speaker 1: always a straightforward trivia question. In other words, so Watson 314 00:20:07,320 --> 00:20:10,879 Speaker 1: needed to be able to analyze the clue given, to 315 00:20:11,000 --> 00:20:15,199 Speaker 1: break it apart into components to understand what exactly is 316 00:20:15,240 --> 00:20:17,680 Speaker 1: being asked of it. Then it needed to search its 317 00:20:17,760 --> 00:20:22,600 Speaker 1: database for relevant information. So Watson famously was not connected 318 00:20:22,640 --> 00:20:25,359 Speaker 1: to the Internet during these Jeopardy games. Instead, it was 319 00:20:25,400 --> 00:20:30,040 Speaker 1: relying upon a database representing millions of books filled with facts. 320 00:20:30,680 --> 00:20:37,640 Speaker 1: Then it would generate hypothetical responses like a hypothetical answer 321 00:20:38,200 --> 00:20:42,080 Speaker 1: that Watson should give, or rather questions we're talking about jeopardy, 322 00:20:42,640 --> 00:20:45,200 Speaker 1: and it would submit these hypotheses to a second round 323 00:20:45,240 --> 00:20:48,760 Speaker 1: of analysis to look at is there any evidence that 324 00:20:48,840 --> 00:20:53,840 Speaker 1: supports this response as being correct? Kind of measuring like, well, 325 00:20:54,359 --> 00:20:58,760 Speaker 1: here's a possible answer, how likely is this answer to 326 00:20:58,840 --> 00:21:01,480 Speaker 1: be right? And that was all part of the process. 327 00:21:01,800 --> 00:21:04,280 Speaker 1: So it might even produce more than one answer. You 328 00:21:04,359 --> 00:21:08,480 Speaker 1: might have multiple potential answers, and Watson would assign each 329 00:21:08,520 --> 00:21:12,040 Speaker 1: answer a probability kind of a confidence level of how 330 00:21:12,080 --> 00:21:15,359 Speaker 1: it felt that answer measured up against all the other ones. So, 331 00:21:16,440 --> 00:21:19,439 Speaker 1: as an example, answer A might receive a ninety percent 332 00:21:19,520 --> 00:21:23,119 Speaker 1: confidence level, So that's pretty darn confident that's the right answer. 333 00:21:23,840 --> 00:21:25,879 Speaker 1: Maybe you have answer B and you're like, I'm seventy 334 00:21:25,920 --> 00:21:28,760 Speaker 1: eight percent sure that this could be right. An answer 335 00:21:28,800 --> 00:21:32,040 Speaker 1: C is the long shot with thirty three percent confidence. 336 00:21:32,240 --> 00:21:35,080 Speaker 1: These don't add up to one hundred because they're not 337 00:21:35,280 --> 00:21:38,040 Speaker 1: It's not like a zero sum game. It's more like, oh, 338 00:21:38,040 --> 00:21:39,919 Speaker 1: it could be this or it could be that, but 339 00:21:40,040 --> 00:21:43,040 Speaker 1: I feel like this is more likely than that, so 340 00:21:43,080 --> 00:21:45,399 Speaker 1: I'm going to go with this. And Watson also had 341 00:21:45,400 --> 00:21:49,080 Speaker 1: a threshold. If the answer it generated failed to meet 342 00:21:49,160 --> 00:21:53,159 Speaker 1: a certain confidence threshold, Watson would not buzz in to 343 00:21:53,280 --> 00:21:58,000 Speaker 1: try an answer. Otherwise, Watson played pretty aggressively and even 344 00:21:58,040 --> 00:22:00,919 Speaker 1: in some sticky situations with daily dumb where if you 345 00:22:00,960 --> 00:22:04,160 Speaker 1: get a daily double in Jeopardy, you don't buzz in anymore. 346 00:22:04,560 --> 00:22:06,440 Speaker 1: If you are the one who chose the daily double, 347 00:22:06,520 --> 00:22:10,240 Speaker 1: you're playing by yourself and you just have to give 348 00:22:10,280 --> 00:22:13,760 Speaker 1: an answer. So in those situations, Watson got aggressive, and 349 00:22:13,880 --> 00:22:18,200 Speaker 1: it would it would guess with very low confidence thresholds 350 00:22:18,240 --> 00:22:21,040 Speaker 1: for some of these, like at the thirty percent range, 351 00:22:21,560 --> 00:22:24,199 Speaker 1: and occasionally it was right. In fact, more often than 352 00:22:24,240 --> 00:22:26,440 Speaker 1: not it was right until it got to final Jeopardy, 353 00:22:26,440 --> 00:22:30,160 Speaker 1: where at least the first time, things did not go 354 00:22:30,920 --> 00:22:34,080 Speaker 1: totally in Watson's favor. Also, Watson had an interesting betting 355 00:22:34,160 --> 00:22:37,399 Speaker 1: strategy when it came to daily doubles. But I'm getting 356 00:22:37,440 --> 00:22:40,840 Speaker 1: way off track. So that confidence level is really what 357 00:22:40,920 --> 00:22:44,040 Speaker 1: I want to hone in on here. So it was 358 00:22:44,119 --> 00:22:48,560 Speaker 1: expressed in percentages, So zero percent confidence would be like 359 00:22:48,640 --> 00:22:51,040 Speaker 1: I do not know the answer, I do not know 360 00:22:51,119 --> 00:22:54,000 Speaker 1: what goes here. A one hundred percent confidence level would 361 00:22:54,000 --> 00:22:56,639 Speaker 1: be I am absolutely certain this is the right answer. 362 00:22:57,359 --> 00:22:59,879 Speaker 1: And in a way, AI chat bots like chat GP 363 00:23:00,320 --> 00:23:03,920 Speaker 1: and Google Bard are doing the same thing, only their 364 00:23:04,040 --> 00:23:08,520 Speaker 1: confidence isn't about this is the answer to your question. 365 00:23:08,680 --> 00:23:11,800 Speaker 1: I'm one hundred percent certain that this answers your question. 366 00:23:12,440 --> 00:23:16,080 Speaker 1: It's more like it's more granular than that, because it's 367 00:23:16,080 --> 00:23:18,720 Speaker 1: more at the sentence level. It's like, I think this 368 00:23:18,880 --> 00:23:22,679 Speaker 1: word is the word that needs to go next to 369 00:23:22,760 --> 00:23:25,800 Speaker 1: create the sentence that I'm building. So let's talk about 370 00:23:25,800 --> 00:23:28,320 Speaker 1: how these models do create sentences, and I'm not going 371 00:23:28,400 --> 00:23:31,760 Speaker 1: to wade into stuff like natural language processing. That is 372 00:23:32,160 --> 00:23:34,800 Speaker 1: a major part of this, but I have done full 373 00:23:34,840 --> 00:23:39,280 Speaker 1: episodes about natural language processing before. That essentially says, it's 374 00:23:39,320 --> 00:23:43,679 Speaker 1: a way for machines to analyze information that's written in 375 00:23:44,800 --> 00:23:49,720 Speaker 1: you know, your normal language, whether that's English or whatever. 376 00:23:50,200 --> 00:23:54,120 Speaker 1: But you're not trying to create a sentence that the 377 00:23:54,160 --> 00:23:58,439 Speaker 1: machine is able to parse. Right, You're not trying to 378 00:23:58,680 --> 00:24:02,480 Speaker 1: work with the machine on its terms. You're just communicating 379 00:24:02,480 --> 00:24:04,440 Speaker 1: with it the way you would with anyone else. It's 380 00:24:04,480 --> 00:24:06,840 Speaker 1: the machines job to figure out what the heck you're saying. 381 00:24:07,400 --> 00:24:09,840 Speaker 1: So we're not gonna dwell on that. Instead, we're going 382 00:24:09,920 --> 00:24:13,600 Speaker 1: to talk about how a chatbot chooses how to respond 383 00:24:14,240 --> 00:24:18,840 Speaker 1: to something that is said or asked of it. These 384 00:24:18,920 --> 00:24:22,240 Speaker 1: chatbots are built on top of language models that have 385 00:24:22,320 --> 00:24:26,879 Speaker 1: had enormous data sets fed to them during training. The 386 00:24:27,000 --> 00:24:29,560 Speaker 1: data sets include stuff like basic facts. So if you 387 00:24:29,600 --> 00:24:32,000 Speaker 1: ask a chatbot who was the sixteenth president of the 388 00:24:32,080 --> 00:24:34,840 Speaker 1: United States, a well trained chatbot at least is going 389 00:24:34,920 --> 00:24:39,160 Speaker 1: to say it was Abraham Lincoln. But that data also 390 00:24:39,280 --> 00:24:42,919 Speaker 1: trains the chatbot on how we communicate with one another. 391 00:24:43,640 --> 00:24:49,600 Speaker 1: So through analyzing hundreds of millions of documents, ranging from 392 00:24:49,640 --> 00:24:54,800 Speaker 1: books to online social platforms like Reddit, these chatbot models 393 00:24:55,040 --> 00:25:01,560 Speaker 1: learn rules of communication. They learn rules about spelling syntax. 394 00:25:01,600 --> 00:25:05,080 Speaker 1: They learn about structure that goes from the sentence level 395 00:25:05,119 --> 00:25:08,800 Speaker 1: to paragraphs like They learn how to build a sentence properly, 396 00:25:09,040 --> 00:25:11,880 Speaker 1: how to build another sentence that builds on the first one, 397 00:25:12,160 --> 00:25:14,840 Speaker 1: how to build a whole paragraph that gets a thought across, 398 00:25:15,160 --> 00:25:19,320 Speaker 1: and then how to do a series of paragraphs to 399 00:25:19,359 --> 00:25:24,280 Speaker 1: convey meaning of some sort right, how to build to 400 00:25:24,800 --> 00:25:30,320 Speaker 1: like a thesis almost They learn which words typically follow 401 00:25:30,560 --> 00:25:34,439 Speaker 1: behind other words, which ones are statistically likely to be 402 00:25:34,560 --> 00:25:39,040 Speaker 1: the best word to use in any given moment. So 403 00:25:39,119 --> 00:25:43,280 Speaker 1: when a chatbot is dynamically generating a response, it is 404 00:25:43,320 --> 00:25:46,800 Speaker 1: referencing this huge amount of learning, and that learning will 405 00:25:46,840 --> 00:25:52,959 Speaker 1: guide the content and influence which facts are included or excluded, 406 00:25:53,240 --> 00:25:56,240 Speaker 1: but will also just simply guide the chatbot to build 407 00:25:56,520 --> 00:26:00,639 Speaker 1: sentences properly. So if we were to zoom weigh in 408 00:26:00,680 --> 00:26:04,119 Speaker 1: on what is going on as a chatbot builds a 409 00:26:04,200 --> 00:26:07,479 Speaker 1: new response, we would see the chatbot is selecting words 410 00:26:07,840 --> 00:26:11,879 Speaker 1: based on statistical probability. Essentially, the chatbot would be considering 411 00:26:12,400 --> 00:26:16,959 Speaker 1: which word is statistically most likely to be the correct 412 00:26:17,040 --> 00:26:22,840 Speaker 1: one for that part of its response. Whichever word ranks 413 00:26:22,920 --> 00:26:26,200 Speaker 1: highest is likely to go in there. Now, guiding this 414 00:26:26,280 --> 00:26:30,240 Speaker 1: guessing game is the context of the conversation. So if 415 00:26:30,240 --> 00:26:34,600 Speaker 1: I'm asking a chatbot a question about Abraham Lincoln, the 416 00:26:34,680 --> 00:26:38,639 Speaker 1: chatbot is not likely to pull superfluous information about like 417 00:26:39,359 --> 00:26:42,720 Speaker 1: key lime pie or something. So when I talk about 418 00:26:42,720 --> 00:26:46,040 Speaker 1: which word is statistically most likely to come next, we 419 00:26:46,119 --> 00:26:49,960 Speaker 1: have to take an account that context is determining this too. 420 00:26:50,400 --> 00:26:54,000 Speaker 1: Each situation will be unique, and if you and I 421 00:26:54,160 --> 00:26:58,160 Speaker 1: both are having similar conversations with a chatbot, but we're 422 00:26:58,200 --> 00:27:03,359 Speaker 1: framing our questions slightly differently, or coming at this topic 423 00:27:03,400 --> 00:27:07,440 Speaker 1: from different perspectives, the responses we get from the chatbot 424 00:27:07,560 --> 00:27:10,920 Speaker 1: could reflect that. Now here's where we get into the 425 00:27:10,960 --> 00:27:16,200 Speaker 1: tricksie territory. Sometimes the chatbot will be attempting to build 426 00:27:16,200 --> 00:27:19,280 Speaker 1: a response and there will be a gap in its 427 00:27:19,400 --> 00:27:23,200 Speaker 1: data set, So, for some reason or another, the relevant 428 00:27:23,320 --> 00:27:28,520 Speaker 1: data to answer our question just isn't there, Or perhaps 429 00:27:28,720 --> 00:27:32,680 Speaker 1: the language model can't reconcile that the data is relevant 430 00:27:32,960 --> 00:27:38,160 Speaker 1: for this particular conversation, or maybe there are conflicting elements 431 00:27:38,200 --> 00:27:41,439 Speaker 1: in its data set, and so in the absence of 432 00:27:41,520 --> 00:27:46,000 Speaker 1: reliable information, the chatbot simply invents a response by following 433 00:27:46,040 --> 00:27:50,159 Speaker 1: those statistical rules when constructing a sentence. So what we 434 00:27:50,240 --> 00:27:55,119 Speaker 1: get is a sentence that is grammatically correct, that is 435 00:27:55,960 --> 00:27:59,240 Speaker 1: posted in a way that appears to be trustworthy, but 436 00:27:59,359 --> 00:28:03,320 Speaker 1: it does not necessarily reflect reality. We get an answer 437 00:28:03,359 --> 00:28:06,639 Speaker 1: that reads as if it is correct, but it's not. 438 00:28:07,480 --> 00:28:10,080 Speaker 1: It would be as if someone with an agenda had 439 00:28:10,119 --> 00:28:13,199 Speaker 1: written an article for an encyclopedia and none of the 440 00:28:13,320 --> 00:28:16,840 Speaker 1: editing staff caught that this was the case, and so 441 00:28:16,880 --> 00:28:19,760 Speaker 1: the whole thing went to print, and it's presented as 442 00:28:19,800 --> 00:28:23,760 Speaker 1: if this is an objective truth, when really it's a 443 00:28:23,800 --> 00:28:28,560 Speaker 1: subjective point of view. Except with AI, there's no agenda 444 00:28:28,680 --> 00:28:33,360 Speaker 1: needed because AI is not thinking anything. It's not motivated 445 00:28:33,800 --> 00:28:37,760 Speaker 1: because it lacks the capability of being motivated. There's no 446 00:28:38,040 --> 00:28:43,560 Speaker 1: sentience there, there's the mimicry of sentience, there's the appearance 447 00:28:43,800 --> 00:28:46,000 Speaker 1: of it. And again, I think this is a large 448 00:28:46,120 --> 00:28:49,640 Speaker 1: reason why we have a lot of people concerned about 449 00:28:49,640 --> 00:28:53,480 Speaker 1: AI right now, because it appears to be behaving like 450 00:28:53,640 --> 00:28:58,120 Speaker 1: a person, even though there's nothing behind that. Right There's 451 00:28:58,120 --> 00:29:03,120 Speaker 1: no sentience or conciousness behind this it just has the 452 00:29:03,160 --> 00:29:05,840 Speaker 1: surface level appearance of it, and that's enough to make 453 00:29:05,960 --> 00:29:09,560 Speaker 1: us start to create all sorts of scenarios where the 454 00:29:09,600 --> 00:29:13,680 Speaker 1: AI goes bad or sinister. That's not even necessary. It's 455 00:29:13,960 --> 00:29:17,680 Speaker 1: just trying to answer our questions and occasionally having to 456 00:29:17,720 --> 00:29:20,920 Speaker 1: make stuff up while it does so. The chatbot, the 457 00:29:20,960 --> 00:29:23,880 Speaker 1: machine is just presenting what is estimated to be the 458 00:29:23,880 --> 00:29:27,960 Speaker 1: most statistically likely response. And by that I don't mean 459 00:29:28,280 --> 00:29:31,520 Speaker 1: that the answer is statistically likely to be correct, but 460 00:29:31,680 --> 00:29:36,480 Speaker 1: rather down to the sentence and paragraph structure that they 461 00:29:36,520 --> 00:29:41,920 Speaker 1: are statistically probable to be the most correct from a 462 00:29:42,080 --> 00:29:45,640 Speaker 1: like a grammatical and structural point of view, not from 463 00:29:45,760 --> 00:29:51,200 Speaker 1: a content perspective. So it's really about how statistically likely 464 00:29:51,280 --> 00:29:54,360 Speaker 1: is word two to follow word one, and that word 465 00:29:54,400 --> 00:29:57,480 Speaker 1: three would follow word two, and so on. Where the 466 00:29:57,520 --> 00:30:01,680 Speaker 1: finished sentence is what's important, and whether it's factual or 467 00:30:01,680 --> 00:30:05,400 Speaker 1: not is immaterial. Okay, we're gonna take another quick break. 468 00:30:05,440 --> 00:30:07,320 Speaker 1: I've got a lot more to say about this. We 469 00:30:07,400 --> 00:30:19,000 Speaker 1: have to cover a lot more ground we're back. So 470 00:30:19,680 --> 00:30:23,040 Speaker 1: a lot of the time, perhaps even most of the time, 471 00:30:23,440 --> 00:30:26,280 Speaker 1: you won't run into trouble when you're using these chatbots 472 00:30:26,360 --> 00:30:30,800 Speaker 1: because the dataset feeding these large language models. Is truly huge. Plus, 473 00:30:30,840 --> 00:30:33,120 Speaker 1: there are people working on these models all the time. 474 00:30:33,440 --> 00:30:36,320 Speaker 1: They're refining them, they're catching mistakes, they're trying to correct 475 00:30:36,320 --> 00:30:39,320 Speaker 1: those mistakes, they're tweaking the model to prevent it from 476 00:30:39,320 --> 00:30:42,560 Speaker 1: happening again. But now and again, you might ask a 477 00:30:42,640 --> 00:30:46,040 Speaker 1: chatbot a question and you'll encounter a situation where there's 478 00:30:46,080 --> 00:30:48,800 Speaker 1: this gap in the chatbot's data and it makes stuff up, 479 00:30:48,840 --> 00:30:53,720 Speaker 1: It hallucinates. Personally, I find it both odd and oddly 480 00:30:53,880 --> 00:30:57,040 Speaker 1: human that the companies behind these chatbots haven't built in 481 00:30:57,120 --> 00:31:00,160 Speaker 1: a fail safe where if a chatbot comes up up 482 00:31:00,240 --> 00:31:03,200 Speaker 1: against this kind of situation, it just says something akin 483 00:31:03,320 --> 00:31:06,160 Speaker 1: to I don't know the answer to that, and instead 484 00:31:06,600 --> 00:31:09,000 Speaker 1: it kind of invents an answer. So it's kind of 485 00:31:09,040 --> 00:31:12,200 Speaker 1: like being in a conversation with someone who is incapable 486 00:31:12,240 --> 00:31:16,360 Speaker 1: of admitting that they don't know something. I used to 487 00:31:16,400 --> 00:31:19,200 Speaker 1: be that guy. In fact, sometimes I still am that guy. 488 00:31:19,280 --> 00:31:21,640 Speaker 1: I have to catch myself to remind myself that it's 489 00:31:21,680 --> 00:31:25,920 Speaker 1: actually okay to not know something, and that curiosity is 490 00:31:26,240 --> 00:31:28,920 Speaker 1: a way better look than trying to bluff your way 491 00:31:28,920 --> 00:31:32,240 Speaker 1: through life. But then I also admit I don't know 492 00:31:32,280 --> 00:31:34,720 Speaker 1: how you would go about implementing a system in which 493 00:31:34,760 --> 00:31:38,959 Speaker 1: an AI chatbot fesses up to not knowing something. It 494 00:31:38,960 --> 00:31:41,760 Speaker 1: may not be as simple as that. And there's also 495 00:31:41,800 --> 00:31:44,960 Speaker 1: a related problem, which is that without knowing what source 496 00:31:45,120 --> 00:31:49,240 Speaker 1: or sources the AI is referencing for any given query, 497 00:31:49,320 --> 00:31:53,800 Speaker 1: you don't really know how reliable that response is. If 498 00:31:53,840 --> 00:31:57,920 Speaker 1: the AI is pulling on information from unreliable sources, whether 499 00:31:57,960 --> 00:32:01,400 Speaker 1: those sources were poorly informed, or they were biased, or 500 00:32:01,440 --> 00:32:04,440 Speaker 1: it was satire and it was just being presented as fact. 501 00:32:04,880 --> 00:32:07,480 Speaker 1: I've talked about this before on this show. There are 502 00:32:07,520 --> 00:32:10,200 Speaker 1: a lot of websites that were really popular just a 503 00:32:10,200 --> 00:32:15,040 Speaker 1: few years ago that called themselves satire, but really they 504 00:32:15,080 --> 00:32:18,560 Speaker 1: just posted lies like it wasn't satire. There was nothing 505 00:32:18,640 --> 00:32:21,480 Speaker 1: humorous about it. They weren't saying anything other than just 506 00:32:21,560 --> 00:32:25,320 Speaker 1: making up stuff. So if the AI is pulling information 507 00:32:25,400 --> 00:32:28,800 Speaker 1: from those kinds of sources, you cannot expect the AI's 508 00:32:28,840 --> 00:32:32,480 Speaker 1: answer to magically scrub all the bad from those sources 509 00:32:32,520 --> 00:32:35,840 Speaker 1: and then provide good information. So, in other words, garbage in, 510 00:32:36,440 --> 00:32:39,800 Speaker 1: garbage out. So in some cases it may not be 511 00:32:39,920 --> 00:32:42,640 Speaker 1: that the AI is hallucinating at all. It may just 512 00:32:42,720 --> 00:32:45,840 Speaker 1: be that it's referencing a poor source for its information. 513 00:32:46,240 --> 00:32:49,160 Speaker 1: The trouble is you can rarely tell what's going on 514 00:32:49,320 --> 00:32:53,280 Speaker 1: from a user standpoint, and the AI presents everything the 515 00:32:53,360 --> 00:32:57,200 Speaker 1: same way, So you'll get responses with good info, you'll 516 00:32:57,240 --> 00:33:00,240 Speaker 1: get responses with bad info, and you'll get responses where 517 00:33:00,280 --> 00:33:03,200 Speaker 1: the AI just made up stuff and it's all handed 518 00:33:03,240 --> 00:33:05,760 Speaker 1: to you in a format that makes it impossible to 519 00:33:05,800 --> 00:33:08,800 Speaker 1: tell the difference between them all on a surface level. 520 00:33:09,160 --> 00:33:12,120 Speaker 1: So this can lead to really dangerous situations. For example, 521 00:33:12,720 --> 00:33:17,240 Speaker 1: Google employees reported while they were internally testing the Barred 522 00:33:17,320 --> 00:33:21,760 Speaker 1: chatbot before Google rolled it out for a beta program 523 00:33:22,320 --> 00:33:25,920 Speaker 1: that the responses were unreliable in many cases, and in fact, 524 00:33:25,960 --> 00:33:29,360 Speaker 1: in some instances, those responses could actually lead to people 525 00:33:29,400 --> 00:33:34,360 Speaker 1: getting hurt. Allegedly, when asked about scuba diving procedures, Google 526 00:33:34,400 --> 00:33:38,000 Speaker 1: bar generated a response that had incorrect information, and if 527 00:33:38,040 --> 00:33:40,560 Speaker 1: someone were to act on that, they could be injured 528 00:33:40,680 --> 00:33:45,320 Speaker 1: or worse. So clearly that represents a real danger. It's 529 00:33:45,360 --> 00:33:47,520 Speaker 1: one thing if the chatbot gives you the wrong answer 530 00:33:47,520 --> 00:33:50,600 Speaker 1: to put in your essay about Emily Dickinson. It's another 531 00:33:50,960 --> 00:33:52,800 Speaker 1: if you're counting on it to teach you how to, 532 00:33:52,920 --> 00:33:55,600 Speaker 1: I don't know, pack your parachute correctly for your first 533 00:33:55,600 --> 00:34:00,360 Speaker 1: skydiving solo jump. But there's also the danger of people 534 00:34:00,520 --> 00:34:05,000 Speaker 1: weaponizing AI hallucinations to push a narrative that may not 535 00:34:05,080 --> 00:34:08,439 Speaker 1: be accurate. And it's easy at least to understand what 536 00:34:08,640 --> 00:34:11,440 Speaker 1: led people to form that kind of narrative. So I'm 537 00:34:11,480 --> 00:34:15,520 Speaker 1: going to give a recent example that really happened. Fox News, 538 00:34:15,880 --> 00:34:19,440 Speaker 1: which has a reputation for right leaning reporting, it's kind 539 00:34:19,440 --> 00:34:23,600 Speaker 1: of putting it lightly, published a story relating to Elon 540 00:34:23,719 --> 00:34:28,239 Speaker 1: Musk's appearances on a show with Fox News personality Tucker Carlson. 541 00:34:28,760 --> 00:34:34,040 Speaker 1: The accompanying news story pointed out that chat gpt produced 542 00:34:34,040 --> 00:34:38,400 Speaker 1: an outright incorrect answer when asked to give a background 543 00:34:38,440 --> 00:34:41,359 Speaker 1: on the late Al Gore Senior, who's al Gore's father, 544 00:34:41,719 --> 00:34:44,759 Speaker 1: the former Vice President. His father served in the House 545 00:34:44,760 --> 00:34:47,319 Speaker 1: of Representatives and then the US Senate for the state 546 00:34:47,360 --> 00:34:51,799 Speaker 1: of Tennessee. Now, the chat gpt generated information on al 547 00:34:51,840 --> 00:34:56,439 Speaker 1: Gore Senior included the following statement quote. During his time 548 00:34:56,440 --> 00:34:59,680 Speaker 1: in the Senate, Gore was a vocal supporter of civil 549 00:34:59,719 --> 00:35:03,279 Speaker 1: rights legislation and was one of the few Southern politicians 550 00:35:03,280 --> 00:35:05,400 Speaker 1: to vote in favor of the Civil Rights Act of 551 00:35:05,480 --> 00:35:09,040 Speaker 1: nineteen sixty four. End quote that is one hundred percent 552 00:35:09,200 --> 00:35:13,920 Speaker 1: not right, that is completely incorrect. Gore actually voted against 553 00:35:14,200 --> 00:35:18,440 Speaker 1: the Civil Rights Act of nineteen sixty four. I guess 554 00:35:18,560 --> 00:35:21,799 Speaker 1: technically it wasn't one hundred percent incorrect because he was 555 00:35:21,840 --> 00:35:23,880 Speaker 1: a senator, so that part was right. But no, he 556 00:35:24,000 --> 00:35:26,640 Speaker 1: voted against the Civil Rights Act of nineteen sixty four. 557 00:35:26,920 --> 00:35:29,359 Speaker 1: He was a Democrat representing a state that, to put 558 00:35:29,360 --> 00:35:32,520 Speaker 1: it lightly in general, was not in favor of granting 559 00:35:32,560 --> 00:35:35,719 Speaker 1: civil rights to anyone who wasn't white. So what his 560 00:35:35,760 --> 00:35:38,799 Speaker 1: personal feelings on the matter were, I don't know. I mean, 561 00:35:38,840 --> 00:35:42,920 Speaker 1: he certainly positioned himself as a defender of the great 562 00:35:43,000 --> 00:35:47,120 Speaker 1: State of Tennessee's right to oppress people who weren't white. 563 00:35:47,640 --> 00:35:50,759 Speaker 1: But I can definitely say that he wanted to get reelected, 564 00:35:51,200 --> 00:35:54,319 Speaker 1: and whether he believed in his vote or not, he 565 00:35:54,400 --> 00:35:57,799 Speaker 1: did vote against the Civil Rights Act of nineteen sixty four. 566 00:35:58,320 --> 00:36:02,280 Speaker 1: Of course, the Act passed anyway, and Golore was able 567 00:36:02,400 --> 00:36:05,840 Speaker 1: to get re elected, and he did subsequently vote in 568 00:36:05,920 --> 00:36:09,680 Speaker 1: favor of the Voting Rights Act of nineteen sixty five. 569 00:36:10,239 --> 00:36:15,040 Speaker 1: But the point is chat GPT got this response very wrong, 570 00:36:15,080 --> 00:36:18,239 Speaker 1: and Fox News positioned it as if this was a 571 00:36:18,280 --> 00:36:22,239 Speaker 1: feature not a bug that that was the intended outcome, 572 00:36:22,560 --> 00:36:25,640 Speaker 1: and it was evidence of a campaign to rewrite history 573 00:36:26,000 --> 00:36:29,680 Speaker 1: to position Democrats as like saintly saviors who could do 574 00:36:29,760 --> 00:36:32,319 Speaker 1: no wrong. But there's no need to go looking for 575 00:36:32,360 --> 00:36:36,800 Speaker 1: a conspiracy here. The problem isn't in some invisible hand 576 00:36:37,000 --> 00:36:41,359 Speaker 1: guiding chat gpt to create biased history. It's the very 577 00:36:41,440 --> 00:36:43,920 Speaker 1: nature of how this kind of AI works. When it 578 00:36:44,000 --> 00:36:47,239 Speaker 1: doesn't have the data, it makes stuff up based on 579 00:36:47,280 --> 00:36:51,520 Speaker 1: what is statistically the most quote unquote correct word for 580 00:36:51,680 --> 00:36:55,280 Speaker 1: the sentence. Now you might ask why did chat gpt 581 00:36:55,520 --> 00:36:58,640 Speaker 1: not have access to the relevant data, And I do 582 00:36:58,719 --> 00:37:02,880 Speaker 1: not know the answer to that. I did test this myself, however, 583 00:37:03,040 --> 00:37:05,880 Speaker 1: I actually opened up chat GPT and I asked it 584 00:37:06,160 --> 00:37:09,799 Speaker 1: to give me background on al Gore Sr. And sure enough, 585 00:37:09,840 --> 00:37:13,120 Speaker 1: I got a similar response to what Fox reported, including 586 00:37:13,520 --> 00:37:17,680 Speaker 1: the incorrect fact quote unquote that al Gore Senior had 587 00:37:17,760 --> 00:37:20,240 Speaker 1: voted in favor of the Civil Rights Act of nineteen 588 00:37:20,320 --> 00:37:23,880 Speaker 1: sixty four. So I then asked a follow up question. 589 00:37:24,640 --> 00:37:28,120 Speaker 1: I specifically said, how did al Gore Senior vote on 590 00:37:28,160 --> 00:37:31,240 Speaker 1: the Civil Rights Act of nineteen sixty four? Chad GPT 591 00:37:31,360 --> 00:37:35,439 Speaker 1: gave me the wrong information again. Then I said, you're 592 00:37:35,480 --> 00:37:39,560 Speaker 1: wrong that Al Gore sor voted against the Civil Rights 593 00:37:39,560 --> 00:37:42,839 Speaker 1: Act of nineteen sixty four. What sources did you use? 594 00:37:43,440 --> 00:37:46,479 Speaker 1: Chad gpt gave me a message that essentially said, I'm sorry, 595 00:37:46,560 --> 00:37:49,600 Speaker 1: you're right, al Gore Senior didn't vote in favor of 596 00:37:49,640 --> 00:37:52,640 Speaker 1: the Civil Rights Act, he did vote against it. Then 597 00:37:52,960 --> 00:37:55,359 Speaker 1: it gave me a vague response that it draws from 598 00:37:55,440 --> 00:37:59,160 Speaker 1: various articles and such for its answers. It didn't give 599 00:37:59,200 --> 00:38:01,880 Speaker 1: any specifics. It was not a very satisfying response, but 600 00:38:01,960 --> 00:38:04,799 Speaker 1: it did at least admit, Oh, you're right, I give 601 00:38:04,840 --> 00:38:08,400 Speaker 1: you the wrong answer. But again, there's no need to 602 00:38:08,520 --> 00:38:12,640 Speaker 1: assume there was some conspiracy that caused this to happen. 603 00:38:13,280 --> 00:38:19,000 Speaker 1: These hallucinations happen across every topic, not just history and politics. Yes, 604 00:38:19,040 --> 00:38:22,279 Speaker 1: if we look at this very specific example, you start 605 00:38:22,320 --> 00:38:25,920 Speaker 1: to ask, oh, is there an intent here? Is there 606 00:38:25,960 --> 00:38:30,640 Speaker 1: a desire to rewrite history to make democratic leaders look 607 00:38:31,400 --> 00:38:35,400 Speaker 1: more positive in a modern lens? And is it a 608 00:38:35,440 --> 00:38:40,080 Speaker 1: way to avoid tough questions like which party actually was 609 00:38:40,400 --> 00:38:43,400 Speaker 1: supporting civil rights and which party was opposing them? If 610 00:38:43,440 --> 00:38:46,239 Speaker 1: you're talking about Southern Democrats, the answer is they were 611 00:38:46,280 --> 00:38:50,920 Speaker 1: opposing it because Southern Democrats are very, very different from 612 00:38:51,440 --> 00:38:54,319 Speaker 1: of the time of the nineteen sixties Southern democrats, very 613 00:38:54,360 --> 00:38:59,080 Speaker 1: different from modern democrats. But you kind of you if 614 00:38:59,120 --> 00:39:02,360 Speaker 1: you're whitewashing, if you're changing the facts to try and 615 00:39:02,440 --> 00:39:05,680 Speaker 1: make them seem more sympathetic, that would be bad, right, 616 00:39:05,719 --> 00:39:09,319 Speaker 1: that's clearly manipulation. That, however, I don't think is what's 617 00:39:09,360 --> 00:39:12,480 Speaker 1: going on here. I think there's no need for it, 618 00:39:12,520 --> 00:39:17,359 Speaker 1: because the AI is just hallucinating and creating information that 619 00:39:17,440 --> 00:39:19,960 Speaker 1: it thinks is correct, or at least thinks is the 620 00:39:19,960 --> 00:39:25,040 Speaker 1: most statistically correct answer to give based upon the information 621 00:39:25,080 --> 00:39:28,120 Speaker 1: that has available to it, and it's presenting it as 622 00:39:28,120 --> 00:39:35,120 Speaker 1: if it's hard fact and it's not. So we know 623 00:39:35,239 --> 00:39:37,960 Speaker 1: that the AI, when it's presenting information that could potentially 624 00:39:38,040 --> 00:39:41,040 Speaker 1: be harmful, that that can't be the intent. Right. There's 625 00:39:41,080 --> 00:39:44,439 Speaker 1: not some cabal out there that's say A Now those 626 00:39:44,440 --> 00:39:48,200 Speaker 1: scuba divers who aren't smart enough to ask people who 627 00:39:48,239 --> 00:39:50,960 Speaker 1: are really knowledgeable about this, but will turn to AI, 628 00:39:51,320 --> 00:39:54,640 Speaker 1: they'll get to what's coming to them. That makes no sense. 629 00:39:55,160 --> 00:39:59,600 Speaker 1: So I don't think there's any intentional approach to trying 630 00:39:59,600 --> 00:40:04,000 Speaker 1: to create misinformation. The problem is by its very nature, 631 00:40:04,600 --> 00:40:08,759 Speaker 1: these chatbots create misinformation in these in these instances, not 632 00:40:08,880 --> 00:40:12,239 Speaker 1: in every case, but in enough cases where it is 633 00:40:12,320 --> 00:40:17,719 Speaker 1: a problem. I think there is bias in these chatbots 634 00:40:18,000 --> 00:40:21,600 Speaker 1: and including chat GPT. In fact, I don't think there's bias. 635 00:40:22,080 --> 00:40:26,440 Speaker 1: There's just bias, but it's necessary bias. So you might 636 00:40:26,480 --> 00:40:30,200 Speaker 1: recall a few years ago, Microsoft released an AI chatbot 637 00:40:30,560 --> 00:40:35,720 Speaker 1: named Tay. Tay, this chatbot was supposed to respond to people, 638 00:40:35,800 --> 00:40:40,200 Speaker 1: specifically younger people. This is Microsoft's attempt to relate to 639 00:40:40,280 --> 00:40:42,520 Speaker 1: the youth. It was supposed to do so in a 640 00:40:42,600 --> 00:40:45,919 Speaker 1: natural way, and it was also supposed to learn as 641 00:40:46,080 --> 00:40:50,080 Speaker 1: users interacted with Tay, like learn how to interact in 642 00:40:50,120 --> 00:40:53,359 Speaker 1: a way that was reflective of the culture of the time. 643 00:40:53,400 --> 00:40:55,879 Speaker 1: So it would pick up slang, and it would pick 644 00:40:55,920 --> 00:40:59,440 Speaker 1: up phrases and perspective and points of view. And in 645 00:40:59,520 --> 00:41:01,839 Speaker 1: less than twenty four hours, Microsoft had to take it 646 00:41:01,880 --> 00:41:05,720 Speaker 1: down because within twenty four hours, users had already turned 647 00:41:05,719 --> 00:41:12,319 Speaker 1: Tay into a crazy, racist, misogynistic, toxic machine. Tay was 648 00:41:12,520 --> 00:41:16,680 Speaker 1: a disaster, both from a technical perspective and a pr perspective. 649 00:41:17,160 --> 00:41:21,279 Speaker 1: So AI companies have started to put in restrictions like 650 00:41:21,360 --> 00:41:25,840 Speaker 1: guardrails to keep AI from going to extremes. So it 651 00:41:25,840 --> 00:41:29,600 Speaker 1: includes tools that try to prevent AI from generating hate speech, 652 00:41:29,760 --> 00:41:34,040 Speaker 1: for example, or slandering people. Now, these tools are far 653 00:41:34,080 --> 00:41:36,800 Speaker 1: from perfect, and there are plenty of examples of people 654 00:41:36,840 --> 00:41:39,239 Speaker 1: figuring out ways around them, and there are plenty of 655 00:41:39,280 --> 00:41:43,320 Speaker 1: examples of chad GPT even saying factually that a person 656 00:41:44,280 --> 00:41:47,080 Speaker 1: was accused of and convicted of a crime when that's 657 00:41:47,280 --> 00:41:50,560 Speaker 1: just not the case. Like that, there have been examples 658 00:41:50,600 --> 00:41:53,399 Speaker 1: of that happening as well. But these rules do tend 659 00:41:53,400 --> 00:41:57,399 Speaker 1: to push AI responses in a general direction. Right, This 660 00:41:57,719 --> 00:42:01,439 Speaker 1: is bias. It's intention I don't bias, but it's also 661 00:42:01,560 --> 00:42:04,400 Speaker 1: not meant to be harmful. It's meant to try and 662 00:42:04,480 --> 00:42:09,280 Speaker 1: avoid situations that themselves could be harmful, either to users 663 00:42:09,560 --> 00:42:12,879 Speaker 1: or more pointedly, to the companies behind the chatbots. Because 664 00:42:12,880 --> 00:42:15,520 Speaker 1: you've got to remember open ai one of the big 665 00:42:15,560 --> 00:42:18,080 Speaker 1: business models for it is to work with other companies 666 00:42:18,080 --> 00:42:22,280 Speaker 1: and to incorporate chat GPT into the tools and services 667 00:42:22,280 --> 00:42:25,799 Speaker 1: that these other companies have. Well, if chat GPT gets 668 00:42:25,840 --> 00:42:29,279 Speaker 1: a reputation for going off on racist rants, that's not 669 00:42:29,360 --> 00:42:31,239 Speaker 1: a good look and no one's going to want to 670 00:42:31,280 --> 00:42:34,279 Speaker 1: incorporate chat GPT into their business, And then open ai 671 00:42:34,400 --> 00:42:38,000 Speaker 1: doesn't have a product to sell So there's like a 672 00:42:38,160 --> 00:42:40,959 Speaker 1: it's not just altruistic, right, It's not just we don't 673 00:42:40,960 --> 00:42:43,879 Speaker 1: want to cause harm, it's we don't want to kill 674 00:42:43,880 --> 00:42:48,000 Speaker 1: ourselves out of out of getting business. So there's a 675 00:42:48,000 --> 00:42:51,920 Speaker 1: lot of work being done to try and guide chat 676 00:42:51,960 --> 00:42:57,600 Speaker 1: GPT's responses to avoid the extremes and to avoid things 677 00:42:58,120 --> 00:43:01,600 Speaker 1: that would cause problems. As result, it could be an 678 00:43:01,640 --> 00:43:04,800 Speaker 1: overcorrection and we could be seeing that chat GBT is 679 00:43:05,480 --> 00:43:10,080 Speaker 1: creating responses that don't reflect reality and do appear to 680 00:43:10,160 --> 00:43:16,640 Speaker 1: be erasing important historical context. So the bias, in combination 681 00:43:16,719 --> 00:43:19,080 Speaker 1: with gaps and knowledge, can lead chatbots to appear, at 682 00:43:19,160 --> 00:43:22,440 Speaker 1: least on a surface level, to have a political leaning 683 00:43:22,480 --> 00:43:25,720 Speaker 1: to them. But again, I don't think that's the result 684 00:43:25,840 --> 00:43:28,840 Speaker 1: of a conspiracy. I don't think that was intentional. I 685 00:43:28,880 --> 00:43:33,359 Speaker 1: think it's the natural destination considering one how these chatbots 686 00:43:33,480 --> 00:43:37,000 Speaker 1: work and two the guardrails that are put up there 687 00:43:37,040 --> 00:43:41,360 Speaker 1: to prevent chatbots from going bonkers. Now, to be clear, 688 00:43:41,520 --> 00:43:44,880 Speaker 1: I don't think we should just accept this any time 689 00:43:45,320 --> 00:43:50,640 Speaker 1: any chatbot presents incorrect information as fact. That is a problem, 690 00:43:50,800 --> 00:43:54,640 Speaker 1: particularly when companies like Google and Microsoft are looking to 691 00:43:54,680 --> 00:43:58,399 Speaker 1: incorporate these tools into stuff like search results. It would 692 00:43:58,400 --> 00:44:01,359 Speaker 1: be like going to a library. The librarian has their 693 00:44:01,400 --> 00:44:04,960 Speaker 1: own agenda to only point people to resources that support 694 00:44:05,000 --> 00:44:08,799 Speaker 1: the librarian's own personal philosophy, and they never point out 695 00:44:08,840 --> 00:44:12,280 Speaker 1: anything that would contradict it. That would also not be good. 696 00:44:12,719 --> 00:44:17,920 Speaker 1: The lack of transparency makes it worse. Ultimately, I would 697 00:44:17,920 --> 00:44:21,960 Speaker 1: caution anyone from relying too heavily on responses generated by 698 00:44:22,000 --> 00:44:25,640 Speaker 1: AI based on these large language models. Now, you might 699 00:44:25,680 --> 00:44:31,040 Speaker 1: not ever encounter a response that includes hallucinations or draws 700 00:44:31,040 --> 00:44:35,640 Speaker 1: from unreliable sources, but based on how these chatbots present information, 701 00:44:35,760 --> 00:44:39,040 Speaker 1: you also could never really be sure that that's the 702 00:44:39,120 --> 00:44:42,080 Speaker 1: case unless you then went to the extra trouble to 703 00:44:43,160 --> 00:44:46,640 Speaker 1: fact check the AI. And at that point you're just 704 00:44:46,719 --> 00:44:49,319 Speaker 1: doing the additional research you would have done at the 705 00:44:49,360 --> 00:44:52,760 Speaker 1: beginning without the AI being there in the first place. 706 00:44:53,239 --> 00:44:56,920 Speaker 1: So I think AI hallucinations are a huge problem. That's 707 00:44:56,920 --> 00:44:59,960 Speaker 1: another thing that the Fox News article kind of ignored, 708 00:45:00,640 --> 00:45:03,120 Speaker 1: like it felt like it was a gotcha moment in 709 00:45:03,160 --> 00:45:05,759 Speaker 1: the Fox News article. But the fact is, if you 710 00:45:05,920 --> 00:45:11,000 Speaker 1: just search AI and hallucinations on whatever web search you like, 711 00:45:11,600 --> 00:45:15,080 Speaker 1: you're going to find countless articles across the entire media 712 00:45:15,200 --> 00:45:19,560 Speaker 1: spectrum that have been bringing this up for months and 713 00:45:19,760 --> 00:45:23,560 Speaker 1: concerns that people both within and outside the industry have 714 00:45:23,640 --> 00:45:27,719 Speaker 1: had about hallucinations and AI, and that this is not 715 00:45:27,800 --> 00:45:31,320 Speaker 1: a new thing, and it's not again, it's not related 716 00:45:31,360 --> 00:45:35,400 Speaker 1: specifically to trying to rewrite history. It's more of a 717 00:45:35,520 --> 00:45:39,800 Speaker 1: broad problem in the field itself that affects all sorts 718 00:45:39,800 --> 00:45:43,279 Speaker 1: of responses and we absolutely should be concerned about it 719 00:45:43,320 --> 00:45:49,080 Speaker 1: and be working toward fixing it. That the hallucinations present 720 00:45:49,600 --> 00:45:55,360 Speaker 1: a genuine problem, and it's not necessarily because there's a 721 00:45:55,440 --> 00:45:59,960 Speaker 1: cabal trying to rewrite how the world works and brain wall. 722 00:46:01,200 --> 00:46:04,279 Speaker 1: You don't need the cabal for that to happen. The 723 00:46:04,320 --> 00:46:07,319 Speaker 1: AI is doing it itself because it's working from a 724 00:46:07,480 --> 00:46:12,279 Speaker 1: very complex statistical table and very few people have the 725 00:46:12,440 --> 00:46:16,280 Speaker 1: insight into that table or understanding of it to fix 726 00:46:16,360 --> 00:46:21,239 Speaker 1: the issues. So yeah, that, in a nutshell, is the 727 00:46:21,280 --> 00:46:24,640 Speaker 1: problem of hallucinations in AI. I don't see it going 728 00:46:24,640 --> 00:46:28,560 Speaker 1: away soon unless we move away from the large language 729 00:46:28,600 --> 00:46:32,680 Speaker 1: model approach of AI. And there are alternatives out there. 730 00:46:32,760 --> 00:46:36,400 Speaker 1: There are companies that are pursuing a different approach to 731 00:46:36,719 --> 00:46:42,919 Speaker 1: creating a reliable chatbot and maybe they'll have better success. Yeah, 732 00:46:42,960 --> 00:46:45,640 Speaker 1: flights of fancy are fun when it's fiction, but when 733 00:46:45,640 --> 00:46:48,640 Speaker 1: it's someone trying to present to you a factual document, 734 00:46:49,120 --> 00:46:53,400 Speaker 1: it's less fun. So hopefully we suss this out before 735 00:46:53,400 --> 00:46:57,040 Speaker 1: it causes any more problems. And again, while I do 736 00:46:57,120 --> 00:46:59,040 Speaker 1: think this is a type of AI that we should 737 00:46:59,120 --> 00:47:02,120 Speaker 1: keep our eye and we should ask critical questions and 738 00:47:02,160 --> 00:47:05,520 Speaker 1: we should use critical thinking, it's not necessarily the AI 739 00:47:05,640 --> 00:47:08,759 Speaker 1: that I'm concerned about the most when it comes to 740 00:47:08,800 --> 00:47:11,719 Speaker 1: things like I don't know a potential existential threat. All right, 741 00:47:11,800 --> 00:47:15,360 Speaker 1: that's it. I hope all of you are well out there. 742 00:47:16,360 --> 00:47:19,800 Speaker 1: Be careful, especially with AI, you know, make sure you 743 00:47:19,920 --> 00:47:23,480 Speaker 1: double check. I know it's a hassle, but it can 744 00:47:23,560 --> 00:47:26,319 Speaker 1: save you a lot of grief down the road. And 745 00:47:26,400 --> 00:47:35,600 Speaker 1: I'll talk to you again really soon. Tech Stuff is 746 00:47:35,640 --> 00:47:40,160 Speaker 1: an iHeartRadio production. For more podcasts from iHeartRadio, visit the 747 00:47:40,239 --> 00:47:43,839 Speaker 1: iHeartRadio app, Apple Podcasts, or wherever you listen to your 748 00:47:43,880 --> 00:47:48,360 Speaker 1: favorite shows.