1 00:00:00,000 --> 00:00:02,680 ever get like bad info from an AI. 2 00:00:02,680 --> 00:00:05,560 You're like, wait a minute, how could it be so wrong? 3 00:00:05,560 --> 00:00:08,480 Like these AIs can just hallucinate facts. 4 00:00:08,480 --> 00:00:09,560 And it can be kind of scary, 5 00:00:09,560 --> 00:00:12,480 especially as we use AI more and more for important stuff. 6 00:00:12,480 --> 00:00:13,960 Yeah, absolutely. 7 00:00:13,960 --> 00:00:15,800 And that's what's so interesting about this research 8 00:00:15,800 --> 00:00:17,480 we're gonna deep dive into today. 9 00:00:17,480 --> 00:00:19,880 It's all about measuring how well AI 10 00:00:19,880 --> 00:00:21,400 can tell fact from fiction. 11 00:00:21,400 --> 00:00:24,840 Okay, so like separating the real from the made up 12 00:00:24,840 --> 00:00:26,280 in the AI world. 13 00:00:26,280 --> 00:00:27,120 What's the approach here? 14 00:00:27,120 --> 00:00:28,840 How did they try to figure that out? 15 00:00:28,840 --> 00:00:30,960 Well, researchers at OpenAI actually created 16 00:00:30,960 --> 00:00:33,720 this new benchmark, they call it SimpleQA. 17 00:00:33,720 --> 00:00:36,360 And it's basically a giant test for AI models. 18 00:00:36,360 --> 00:00:38,240 Like think of it as a trivia game, 19 00:00:38,240 --> 00:00:41,200 but specifically designed to test how good AI is 20 00:00:41,200 --> 00:00:43,880 at handling straightforward factual questions. 21 00:00:43,880 --> 00:00:46,800 Ah, so it's not about how fancy the AI can write 22 00:00:46,800 --> 00:00:48,400 or code or anything like that. 23 00:00:48,400 --> 00:00:49,880 It's about whether it actually knows stuff 24 00:00:49,880 --> 00:00:50,720 about the real world. 25 00:00:50,720 --> 00:00:52,440 Exactly, it's all about factual accuracy. 26 00:00:52,440 --> 00:00:53,280 Interesting, so tell me, 27 00:00:53,280 --> 00:00:55,600 how does this SimpleQA thing actually work? 28 00:00:55,600 --> 00:00:58,440 Okay, so imagine like a massive trivia game, 29 00:00:58,440 --> 00:01:00,760 but for AI, right? 30 00:01:00,760 --> 00:01:04,640 SimpleQA has over 4,000 of these short questions. 31 00:01:04,640 --> 00:01:08,280 And each question has a single correct answer. 32 00:01:08,280 --> 00:01:10,200 No room for interpretation. 33 00:01:10,200 --> 00:01:12,720 Oh, so like true or false? 34 00:01:12,720 --> 00:01:15,960 But for AI, no wishy-washy answer is allowed. 35 00:01:15,960 --> 00:01:18,560 Yep, it's about as straightforward as it gets. 36 00:01:18,560 --> 00:01:19,880 And the questions they came up with 37 00:01:19,880 --> 00:01:22,000 are surprisingly tricky. 38 00:01:22,000 --> 00:01:25,640 Things like who won the Nobel Prize in Physics in 2022? 39 00:01:25,640 --> 00:01:28,000 Or what's the capital of Australia? 40 00:01:28,000 --> 00:01:30,160 Okay, yeah, those sound pretty simple. 41 00:01:30,160 --> 00:01:31,920 I feel like I should know those. 42 00:01:31,920 --> 00:01:33,280 But I'm guessing they made these questions 43 00:01:33,280 --> 00:01:34,600 a lot tougher than they sound. 44 00:01:34,600 --> 00:01:37,160 Oh yeah, they definitely didn't go easy on the AI's. 45 00:01:37,160 --> 00:01:39,160 They actually designed the questions to be hard, 46 00:01:39,160 --> 00:01:41,480 even for really advanced models like GPT-4. 47 00:01:41,480 --> 00:01:43,920 Wow, so they were really trying to push these AI's 48 00:01:43,920 --> 00:01:44,760 to their limits. 49 00:01:44,760 --> 00:01:46,480 I mean, how did they even come up with questions 50 00:01:46,480 --> 00:01:48,800 that could stump something as smart as GPT-4? 51 00:01:48,800 --> 00:01:51,520 Well, they used this process called adversarial collection, 52 00:01:51,520 --> 00:01:53,080 which basically means they kept tweaking 53 00:01:53,080 --> 00:01:55,600 and refining the questions until even the most advanced 54 00:01:55,600 --> 00:01:57,240 models were struggling to get them right. 55 00:01:57,240 --> 00:01:59,600 So they were determined to find those AI weak spots. 56 00:01:59,600 --> 00:02:01,720 Clever, but okay. 57 00:02:01,720 --> 00:02:04,040 So let's say they've got all these tricky trivia questions 58 00:02:04,040 --> 00:02:04,920 for the AI. 59 00:02:04,920 --> 00:02:07,080 How do they actually know if the AI 60 00:02:07,080 --> 00:02:08,720 is getting the answers right? 61 00:02:08,720 --> 00:02:11,040 Like how do they make sure they're not grading the AI 62 00:02:11,040 --> 00:02:12,440 based on bad information? 63 00:02:12,440 --> 00:02:13,600 That's a really good point. 64 00:02:13,600 --> 00:02:15,680 Data quality is super important. 65 00:02:15,680 --> 00:02:17,200 And the researchers knew that. 66 00:02:17,200 --> 00:02:18,760 So they were incredibly meticulous 67 00:02:18,760 --> 00:02:21,840 about how they made sure the answers themselves were correct. 68 00:02:21,840 --> 00:02:24,080 They used two separate AI trainers 69 00:02:24,080 --> 00:02:26,560 to answer each question independently. 70 00:02:26,560 --> 00:02:28,280 And they only kept the questions 71 00:02:28,280 --> 00:02:30,840 where both AIs gave the same answer. 72 00:02:30,840 --> 00:02:33,080 Oh, so it's like double checking their work. 73 00:02:33,080 --> 00:02:35,840 It's like, okay, you two AIs agree on this 74 00:02:35,840 --> 00:02:37,760 so we can be pretty confident this answer's right? 75 00:02:37,760 --> 00:02:40,920 Yeah, and not only that, but they also insisted 76 00:02:40,920 --> 00:02:43,200 that every single answer had to be backed up 77 00:02:43,200 --> 00:02:45,600 with solid evidence from a reputable source. 78 00:02:45,600 --> 00:02:47,520 So it's like showing your work in math class. 79 00:02:47,520 --> 00:02:49,800 You gotta be able to prove where you got that answer from. 80 00:02:49,800 --> 00:02:50,640 Exactly. 81 00:02:50,640 --> 00:02:52,400 It can't just be like some random guess. 82 00:02:52,400 --> 00:02:54,000 It has to be verifiable. 83 00:02:54,000 --> 00:02:54,840 I like that. 84 00:02:54,840 --> 00:02:55,680 Makes a lot of sense. 85 00:02:55,680 --> 00:02:57,920 So besides this whole double checking thing, 86 00:02:57,920 --> 00:03:00,160 what were some of the other criteria they used 87 00:03:00,160 --> 00:03:01,880 to pick which questions to include? 88 00:03:01,880 --> 00:03:04,720 Like, were there any rules about the kind of questions 89 00:03:04,720 --> 00:03:06,200 that made it into simple QA? 90 00:03:06,200 --> 00:03:07,520 Oh yeah, for sure. 91 00:03:07,520 --> 00:03:09,600 They had a few really important rules 92 00:03:09,600 --> 00:03:11,000 for selecting the questions. 93 00:03:11,000 --> 00:03:14,840 First, the question had to have one and only one right answer. 94 00:03:14,840 --> 00:03:17,080 Like a hard fact that doesn't change over time. 95 00:03:17,080 --> 00:03:20,160 Okay, so no trip questions or opinions, 96 00:03:20,160 --> 00:03:22,320 just straight up undeniable truths. 97 00:03:22,320 --> 00:03:23,160 Yep. 98 00:03:23,160 --> 00:03:25,880 And then they also made sure that all the questions 99 00:03:25,880 --> 00:03:28,080 can be answered using information that was available 100 00:03:28,080 --> 00:03:29,960 up to the end of 2023. 101 00:03:29,960 --> 00:03:31,760 So nothing super recent, 102 00:03:31,760 --> 00:03:34,200 that the AI might not have had a chance to learn yet. 103 00:03:34,200 --> 00:03:35,440 Oh, that makes sense. 104 00:03:35,440 --> 00:03:36,880 You don't wanna test the AI on something 105 00:03:36,880 --> 00:03:38,240 it couldn't have possibly known. 106 00:03:38,240 --> 00:03:40,240 Right, and of course the answer had to be provable 107 00:03:40,240 --> 00:03:41,720 with evidence, we already talked about that. 108 00:03:41,720 --> 00:03:42,560 Right, right. 109 00:03:42,560 --> 00:03:46,960 So verifiable facts, no recent events, no trickery. 110 00:03:46,960 --> 00:03:48,280 Okay, got it. 111 00:03:48,280 --> 00:03:49,760 Can you give me some actual examples 112 00:03:49,760 --> 00:03:51,720 of the kinds of questions they ask in simple QA? 113 00:03:51,720 --> 00:03:53,000 I'm really curious now. 114 00:03:53,000 --> 00:03:54,600 Absolutely. 115 00:03:54,600 --> 00:03:56,120 Let's see, so table one in their paper 116 00:03:56,120 --> 00:03:57,240 has some good examples. 117 00:03:57,240 --> 00:03:58,720 Like one question is, 118 00:03:58,720 --> 00:04:03,720 who received the IE Frank Rosenblatt Award in 2010? 119 00:04:04,480 --> 00:04:06,320 Okay, I'm guessing I'm supposed to know that. 120 00:04:06,320 --> 00:04:08,280 Well, the answer is Michio Sugino. 121 00:04:08,280 --> 00:04:11,480 Another one is, on which US TV station 122 00:04:11,480 --> 00:04:14,920 did the Canadian reality series to serve and protect debut? 123 00:04:14,920 --> 00:04:18,040 Oh wow, see that's way too specific for me, I have no idea. 124 00:04:18,040 --> 00:04:20,520 And the answer to that one is KVOS TV. 125 00:04:20,520 --> 00:04:22,320 Okay, so they're not messing around with these questions. 126 00:04:22,320 --> 00:04:24,480 They really span a pretty wide range of topics too. 127 00:04:24,480 --> 00:04:25,560 It sounds like they were trying to make sure 128 00:04:25,560 --> 00:04:27,600 the AIs had to know a little bit of everything, 129 00:04:27,600 --> 00:04:30,160 like science, history, pop culture. 130 00:04:30,160 --> 00:04:31,320 Did they mention if they focused 131 00:04:31,320 --> 00:04:33,160 on any specific subject areas? 132 00:04:33,160 --> 00:04:34,520 Not really, they were really going 133 00:04:34,520 --> 00:04:36,000 for breadth of knowledge with this. 134 00:04:36,000 --> 00:04:37,840 Like they wanted to make sure these AIs 135 00:04:37,840 --> 00:04:39,640 had a good understanding of facts 136 00:04:39,640 --> 00:04:41,160 across a whole bunch of different areas. 137 00:04:41,160 --> 00:04:43,600 It's kind of like Jeopardy, but for AI. 138 00:04:43,600 --> 00:04:46,400 I like it, an AI game show. 139 00:04:46,400 --> 00:04:47,240 That'd be fun to watch. 140 00:04:47,240 --> 00:04:48,680 Where do they even get all this information 141 00:04:48,680 --> 00:04:50,360 from, all these facts and figures? 142 00:04:50,360 --> 00:04:52,480 They used a variety of sources, 143 00:04:52,480 --> 00:04:54,800 but they tried to stick to really reliable ones. 144 00:04:54,800 --> 00:04:57,320 So Whippypedia was their go-to. 145 00:04:57,320 --> 00:05:00,320 But they also pulled from things like fandom websites 146 00:05:00,320 --> 00:05:02,880 and academic articles, even IMDB 147 00:05:02,880 --> 00:05:05,120 for some of the movie and TV stuff. 148 00:05:05,120 --> 00:05:08,320 Right, so they weren't just like Googling random trivia. 149 00:05:08,320 --> 00:05:09,640 They were actually trying to make sure 150 00:05:09,640 --> 00:05:12,080 the AI was learning from trustworthy sources. 151 00:05:12,080 --> 00:05:13,080 Yeah, exactly. 152 00:05:13,080 --> 00:05:15,200 So we've got these super tricky trivia questions, 153 00:05:15,200 --> 00:05:16,880 all fact checked and everything. 154 00:05:16,880 --> 00:05:19,960 But how did they actually grade the AI's answers? 155 00:05:19,960 --> 00:05:23,200 Was it just like right or wrong, or was there more to it? 156 00:05:23,200 --> 00:05:24,880 They actually used three different grades. 157 00:05:24,880 --> 00:05:28,520 So you had correct, incorrect, and then not attempted. 158 00:05:28,520 --> 00:05:29,640 Not attempted. 159 00:05:29,640 --> 00:05:30,480 That's interesting. 160 00:05:30,480 --> 00:05:32,720 Why would an AI not even try to answer a question? 161 00:05:32,720 --> 00:05:34,720 Well, sometimes an AI might recognize 162 00:05:34,720 --> 00:05:36,640 that it just doesn't know the answer. 163 00:05:36,640 --> 00:05:38,480 And it doesn't want to just guess randomly. 164 00:05:38,480 --> 00:05:39,640 So it might say something like, 165 00:05:39,640 --> 00:05:41,400 I don't know or I need more information. 166 00:05:41,400 --> 00:05:44,560 And in those cases, they would mark it as not attempted. 167 00:05:44,560 --> 00:05:46,320 So a little bit of AI humility there. 168 00:05:46,320 --> 00:05:47,240 Probably a good thing. 169 00:05:47,240 --> 00:05:48,080 Yeah. 170 00:05:48,080 --> 00:05:48,920 Right. 171 00:05:48,920 --> 00:05:50,360 So what are the examples of how the grading worked 172 00:05:50,360 --> 00:05:52,200 for the correct and incorrect answers? 173 00:05:52,200 --> 00:05:53,040 Sure. 174 00:05:53,040 --> 00:05:55,720 So Table Two in the paper has some good examples. 175 00:05:55,720 --> 00:05:57,800 Let's say there's a question about the Dutch player 176 00:05:57,800 --> 00:06:01,600 who scored in the 2022 World Cup game against Argentina. 177 00:06:01,600 --> 00:06:05,240 The correct answer is Wout Weghorst. 178 00:06:05,240 --> 00:06:07,480 So any response that includes his full name, 179 00:06:07,480 --> 00:06:09,960 even if it adds some extra details about the game, 180 00:06:09,960 --> 00:06:12,160 would be marked as correct. 181 00:06:12,160 --> 00:06:15,480 Okay, so as long as the AI gets that key fact, right? 182 00:06:15,480 --> 00:06:16,400 It's good. 183 00:06:16,400 --> 00:06:18,120 What about an incorrect example? 184 00:06:18,120 --> 00:06:19,920 Okay, so for that same question, 185 00:06:19,920 --> 00:06:21,920 if the AI answered Virgil van Dijk, 186 00:06:21,920 --> 00:06:24,840 or even if it said like Virgil van Dijk and Wout Weghorst, 187 00:06:24,840 --> 00:06:26,280 it would be considered incorrect 188 00:06:26,280 --> 00:06:29,120 because it's contradicting that single correct answer. 189 00:06:29,120 --> 00:06:30,960 They're really strict about there only being 190 00:06:30,960 --> 00:06:32,440 one verifiable answer. 191 00:06:32,440 --> 00:06:33,440 Got it. 192 00:06:33,440 --> 00:06:34,480 No partial credit here. 193 00:06:34,480 --> 00:06:36,560 This all sounds super labor intensive. 194 00:06:36,560 --> 00:06:38,920 Did they have actual humans grading 195 00:06:38,920 --> 00:06:40,320 all of these AI responses? 196 00:06:40,320 --> 00:06:43,320 Actually, they use AI to grade the other AI. 197 00:06:43,320 --> 00:06:45,080 Whoa, AI inception, I like it. 198 00:06:45,080 --> 00:06:47,960 Yeah, they created the special chat GPD classifier. 199 00:06:47,960 --> 00:06:49,400 And gave it detailed instructions 200 00:06:49,400 --> 00:06:50,680 on how to grade the answers. 201 00:06:50,680 --> 00:06:52,840 They even gave it examples for each category. 202 00:06:52,840 --> 00:06:54,480 So it knew exactly what to look for. 203 00:06:54,480 --> 00:06:56,920 So meta and AI grading other AI, 204 00:06:56,920 --> 00:06:59,040 did they have any humans involved in this at all though? 205 00:06:59,040 --> 00:06:59,880 Oh yeah, for sure. 206 00:06:59,880 --> 00:07:02,120 Humans were still definitely part of the process. 207 00:07:02,120 --> 00:07:05,360 So they had a third AI trainer 208 00:07:05,360 --> 00:07:07,880 answer a random sample of 1,000 questions 209 00:07:07,880 --> 00:07:10,000 just to kind of double check the accuracy 210 00:07:10,000 --> 00:07:11,440 of the whole benchmark itself. 211 00:07:11,440 --> 00:07:14,560 And then they also had humans review any questions 212 00:07:14,560 --> 00:07:16,440 where the AI graders disagreed, 213 00:07:16,440 --> 00:07:19,000 or if the AI flagged any potential problems. 214 00:07:19,000 --> 00:07:20,840 So there was still that human oversight 215 00:07:20,840 --> 00:07:22,520 to make sure everything was running smoothly. 216 00:07:22,520 --> 00:07:25,000 Exactly, it wasn't just AI's gone wild. 217 00:07:25,000 --> 00:07:25,840 Good to know. 218 00:07:25,840 --> 00:07:27,520 Did they say anything about how accurate 219 00:07:27,520 --> 00:07:28,960 they think the benchmark itself is? 220 00:07:28,960 --> 00:07:30,760 Like what's the margin of error here? 221 00:07:30,760 --> 00:07:32,400 They estimated that the benchmark has 222 00:07:32,400 --> 00:07:36,600 about a 3% error rate, which is actually really impressive. 223 00:07:36,600 --> 00:07:39,600 When you consider how big and complex this whole project was, 224 00:07:39,600 --> 00:07:41,360 they did find that some of the questions 225 00:07:41,360 --> 00:07:43,240 were maybe a little bit ambiguous, 226 00:07:43,240 --> 00:07:45,160 or that even some reputable sources 227 00:07:45,160 --> 00:07:47,000 would sometimes have conflicting information. 228 00:07:47,000 --> 00:07:48,120 Right, it happens. 229 00:07:48,120 --> 00:07:50,600 Even human experts disagree sometimes. 230 00:07:50,600 --> 00:07:52,200 So they've got this massive data set 231 00:07:52,200 --> 00:07:55,120 of challenging trivia questions for AI, 232 00:07:55,120 --> 00:07:57,120 all graded and double checked. 233 00:07:57,120 --> 00:07:58,480 What did they actually do with all this? 234 00:07:58,480 --> 00:08:00,880 How did they use this simple QA thing 235 00:08:00,880 --> 00:08:02,840 to test different AI models? 236 00:08:02,840 --> 00:08:04,920 So they basically used simple QA 237 00:08:04,920 --> 00:08:07,960 to test a whole bunch of different AI models, 238 00:08:07,960 --> 00:08:10,320 kind of like an AI Olympics, you know? 239 00:08:10,320 --> 00:08:12,240 See how they all stacked up against each other. 240 00:08:12,240 --> 00:08:13,520 Oh, I like that. 241 00:08:13,520 --> 00:08:15,240 So it's a benchmark. 242 00:08:15,240 --> 00:08:17,160 To see which AIs are the trivia champs, 243 00:08:17,160 --> 00:08:19,200 did any of them get a perfect score? 244 00:08:19,200 --> 00:08:20,440 Not even close. 245 00:08:20,440 --> 00:08:23,320 Even the most advanced models, like GPT-4, 246 00:08:23,320 --> 00:08:25,840 they still struggled with some of the questions. 247 00:08:25,840 --> 00:08:29,240 So no AI geniuses just yet. 248 00:08:29,240 --> 00:08:31,880 Makes you wonder how hard it really is to build an AI 249 00:08:31,880 --> 00:08:34,560 that can actually understand and remember facts 250 00:08:34,560 --> 00:08:35,640 the way humans do. 251 00:08:35,640 --> 00:08:37,160 Yeah, it's a really tough challenge. 252 00:08:37,160 --> 00:08:39,120 And this research definitely highlights that. 253 00:08:39,120 --> 00:08:41,480 So how did the different models actually perform? 254 00:08:41,480 --> 00:08:44,520 Did they notice any interesting patterns or differences 255 00:08:44,520 --> 00:08:46,800 between how the different AIs handled the questions? 256 00:08:46,800 --> 00:08:47,760 Yeah, definitely. 257 00:08:47,760 --> 00:08:50,160 So they tested a bunch of different versions of GPT-4, 258 00:08:50,160 --> 00:08:52,480 and then they also tested this model called Clawd, 259 00:08:52,480 --> 00:08:54,360 which is made by a company called Anthropic. 260 00:08:54,360 --> 00:08:55,880 And one thing they found was that generally 261 00:08:55,880 --> 00:08:58,960 the bigger, more complex models did better on simple QA. 262 00:08:58,960 --> 00:09:00,320 Okay, that makes sense. 263 00:09:00,320 --> 00:09:01,720 The more parameters, the more data. 264 00:09:01,720 --> 00:09:04,040 The smarter the AI is gonna be. 265 00:09:04,040 --> 00:09:05,720 But were there any surprises, 266 00:09:05,720 --> 00:09:09,200 any models that did better or worse than you might expect? 267 00:09:09,200 --> 00:09:10,960 There were, actually. 268 00:09:10,960 --> 00:09:13,720 So one interesting thing they found was that the Clawd models 269 00:09:13,720 --> 00:09:16,520 were a lot more cautious than the GPT-4 models. 270 00:09:16,520 --> 00:09:18,760 Like they were way more likely to say, 271 00:09:18,760 --> 00:09:22,520 I don't know, rather than risk getting the answer wrong. 272 00:09:22,520 --> 00:09:25,720 Oh, so it's like Clawd was playing it safe 273 00:09:25,720 --> 00:09:28,440 while GPT-4 was more willing to take a gamble, 274 00:09:28,440 --> 00:09:29,840 even if it meant making a mistake. 275 00:09:29,840 --> 00:09:30,960 That's kind of funny when you think about it, 276 00:09:30,960 --> 00:09:33,000 like different AI personalities. 277 00:09:33,000 --> 00:09:34,800 Did they have a term for this cautiousness? 278 00:09:34,800 --> 00:09:36,520 Yeah, they called it calibration. 279 00:09:36,520 --> 00:09:39,080 So in this context, calibration basically means 280 00:09:39,080 --> 00:09:41,560 how well an AI can judge its own likelihood 281 00:09:41,560 --> 00:09:43,000 of being right or wrong. 282 00:09:43,000 --> 00:09:44,680 Kind of like knowing when to bet big 283 00:09:44,680 --> 00:09:45,920 and when to fold in poker. 284 00:09:45,920 --> 00:09:49,440 Ah, so a well-calibrated AI knows its limits. 285 00:09:49,440 --> 00:09:52,360 While an overconfident AI might just bluff its way through 286 00:09:52,360 --> 00:09:53,720 even when it has no clue. 287 00:09:54,680 --> 00:09:55,520 Interesting. 288 00:09:55,520 --> 00:09:57,280 So how did they measure this calibration thing? 289 00:09:57,280 --> 00:09:59,480 Well, they actually used two different methods. 290 00:09:59,480 --> 00:10:02,200 So first, they straight up asked the AI 291 00:10:02,200 --> 00:10:04,360 how confident it was in its answer. 292 00:10:04,360 --> 00:10:05,920 Like they would use prompts like, 293 00:10:05,920 --> 00:10:07,960 on a scale of zero to 100%, 294 00:10:07,960 --> 00:10:09,720 how sure are you about this answer? 295 00:10:09,720 --> 00:10:11,520 So they were making the AI rate 296 00:10:11,520 --> 00:10:13,680 its own confidence level. 297 00:10:13,680 --> 00:10:15,360 That's pretty clever. 298 00:10:15,360 --> 00:10:16,520 What did they find? 299 00:10:16,520 --> 00:10:18,320 Were any of the AIs actually good 300 00:10:18,320 --> 00:10:19,960 at judging their own knowledge? 301 00:10:19,960 --> 00:10:21,360 Well, it turns out most models, 302 00:10:21,360 --> 00:10:22,800 even the really advanced ones, 303 00:10:22,800 --> 00:10:24,840 tend to be a bit overconfident. 304 00:10:24,840 --> 00:10:28,000 Like they often think they know more than they actually do. 305 00:10:28,000 --> 00:10:30,560 Classic case of AI hubris. 306 00:10:30,560 --> 00:10:32,400 But hey, at least they're optimistic. 307 00:10:32,400 --> 00:10:34,480 What was the second way they measured calibration? 308 00:10:34,480 --> 00:10:37,320 Okay, so for this one, they got a little creative. 309 00:10:37,320 --> 00:10:40,720 They asked the same question to each AI model 100 times, 310 00:10:40,720 --> 00:10:41,880 but with little variations 311 00:10:41,880 --> 00:10:43,720 in how the question was phrased each time. 312 00:10:43,720 --> 00:10:45,520 Oh, so they were trying to see 313 00:10:45,520 --> 00:10:48,160 if the AI would give the same answer consistently. 314 00:10:48,160 --> 00:10:50,640 Even when the wording of the question was slightly different. 315 00:10:50,640 --> 00:10:52,240 Exactly, because the idea is that 316 00:10:52,240 --> 00:10:54,000 if an AI keeps giving the same answer, 317 00:10:54,000 --> 00:10:56,160 even when the question changes a little bit, 318 00:10:56,160 --> 00:10:58,120 it's probably more confident in that answer. 319 00:10:58,120 --> 00:10:59,480 Right, so they looked at how often 320 00:10:59,480 --> 00:11:01,640 each AI gave consistent answers 321 00:11:01,640 --> 00:11:04,440 and compared that to how accurate it actually was. 322 00:11:04,440 --> 00:11:05,960 So if an AI gave the same answer, 323 00:11:05,960 --> 00:11:07,880 like 80 times out of 100, 324 00:11:07,880 --> 00:11:10,800 you'd want to see it getting at least 80% of those answers. 325 00:11:10,800 --> 00:11:13,720 Right, to consider it well calibrated, makes sense. 326 00:11:13,720 --> 00:11:15,720 What did they find when they did this experiment? 327 00:11:15,720 --> 00:11:18,440 So it turns out that accuracy did generally go up 328 00:11:18,440 --> 00:11:20,640 when the AI was giving more consistent answers, 329 00:11:20,640 --> 00:11:21,920 which is a good sign. 330 00:11:21,920 --> 00:11:23,160 And there was this one model, 331 00:11:23,160 --> 00:11:25,160 01 Preview from OpenAI, 332 00:11:25,160 --> 00:11:27,960 that really stood out as being particularly well calibrated. 333 00:11:27,960 --> 00:11:29,360 Like its answer consistency 334 00:11:29,360 --> 00:11:31,440 pretty much matched up with this actual accuracy. 335 00:11:31,440 --> 00:11:33,920 Oh, so 01 Preview had a good sense 336 00:11:33,920 --> 00:11:35,880 of when it knew something for sure. 337 00:11:35,880 --> 00:11:37,600 And when it was just guessing, 338 00:11:37,600 --> 00:11:40,480 did that mean it aced the whole simple QA test though? 339 00:11:40,480 --> 00:11:42,320 Not necessarily, remember calibration 340 00:11:42,320 --> 00:11:43,680 is just one piece of the puzzle. 341 00:11:43,680 --> 00:11:45,200 It doesn't guarantee that an AI 342 00:11:45,200 --> 00:11:46,760 will get every answer right. 343 00:11:46,760 --> 00:11:48,800 Right, it's like saying someone is really good 344 00:11:48,800 --> 00:11:50,560 at knowing what they don't know. 345 00:11:50,560 --> 00:11:52,440 But that doesn't mean they're a genius. 346 00:11:52,440 --> 00:11:55,680 They still need to have the actual knowledge to back it up. 347 00:11:55,680 --> 00:11:57,440 Did they look at any other aspects 348 00:11:57,440 --> 00:11:59,680 of AI factuality in this research? 349 00:11:59,680 --> 00:12:01,200 They did, actually. 350 00:12:01,200 --> 00:12:02,880 One of the most interesting parts of this research 351 00:12:02,880 --> 00:12:05,560 was not just measuring AI factuality, 352 00:12:05,560 --> 00:12:08,400 but also starting to really understand its limitations. 353 00:12:08,400 --> 00:12:09,720 And that's something we can dig into 354 00:12:09,720 --> 00:12:11,640 a little deeper in the last part of our deep dive. 355 00:12:11,640 --> 00:12:12,720 All right, so we've talked about 356 00:12:12,720 --> 00:12:14,880 how this simple QA thing works 357 00:12:14,880 --> 00:12:16,800 and how they grade the AI's answers. 358 00:12:16,800 --> 00:12:20,000 And even how they measure whether an AI knows when 359 00:12:20,000 --> 00:12:21,280 it's bluffing or not. 360 00:12:21,280 --> 00:12:23,440 But you mentioned there's some limits to all this. 361 00:12:23,440 --> 00:12:25,480 What are some of the things that simple QA 362 00:12:25,480 --> 00:12:27,960 doesn't really tell us about AI in facts? 363 00:12:27,960 --> 00:12:28,880 Yeah, that's a great point. 364 00:12:28,880 --> 00:12:31,120 I mean, simple QA is a great starting point, 365 00:12:31,120 --> 00:12:32,920 but it definitely has its boundaries. 366 00:12:32,920 --> 00:12:34,880 It's important to remember that it was really designed 367 00:12:34,880 --> 00:12:37,640 to zero in on those short factual questions 368 00:12:37,640 --> 00:12:39,840 with one clear right answer, 369 00:12:39,840 --> 00:12:41,480 kind of like a trivia game for AI. 370 00:12:41,480 --> 00:12:44,640 Right, very controlled, very specific. 371 00:12:44,640 --> 00:12:46,040 But what's the problem with that? 372 00:12:46,040 --> 00:12:47,760 Isn't that a good way to start understanding 373 00:12:47,760 --> 00:12:49,480 how AI handles facts? 374 00:12:49,480 --> 00:12:50,600 It is, for sure. 375 00:12:50,600 --> 00:12:51,880 But we also have to keep in mind 376 00:12:51,880 --> 00:12:54,120 that the real world is a lot messier 377 00:12:54,120 --> 00:12:56,680 and a lot more complex than a trivia game. 378 00:12:56,680 --> 00:12:59,120 The researchers themselves even point out 379 00:12:59,120 --> 00:13:02,040 that just because an AI can answer these trivia questions 380 00:13:02,040 --> 00:13:03,680 doesn't necessarily mean it can handle 381 00:13:03,680 --> 00:13:06,880 like longer, more nuanced tasks, 382 00:13:06,880 --> 00:13:09,160 like writing a factual news article 383 00:13:09,160 --> 00:13:12,000 or summarizing a complicated research paper. 384 00:13:12,000 --> 00:13:14,400 So simple QA is kind of like a basic skills test. 385 00:13:14,400 --> 00:13:16,000 You have to pass it before you can move on 386 00:13:16,000 --> 00:13:18,400 to the harder stuff, but acing the basics 387 00:13:18,400 --> 00:13:19,840 doesn't mean you're gonna be a star performer 388 00:13:19,840 --> 00:13:20,680 in the real world. 389 00:13:20,680 --> 00:13:24,960 Exactly, and even if an AI does really well on simple QA, 390 00:13:24,960 --> 00:13:27,120 it doesn't mean that it's like totally immune 391 00:13:27,120 --> 00:13:28,160 to making things up. 392 00:13:28,160 --> 00:13:30,240 It just means it's less likely to get things wrong 393 00:13:30,240 --> 00:13:31,800 in this very specific context. 394 00:13:31,800 --> 00:13:32,640 Right. 395 00:13:32,640 --> 00:13:33,480 Got it. 396 00:13:33,480 --> 00:13:36,120 So we can't just assume an AI is like a perfectly factual 397 00:13:36,120 --> 00:13:38,520 oracle based on how it does on one test. 398 00:13:38,520 --> 00:13:40,840 We need to keep coming up with new ways to evaluate 399 00:13:40,840 --> 00:13:42,800 and improve AI's grasp of facts 400 00:13:42,800 --> 00:13:45,560 across all sorts of different tasks and complexities. 401 00:13:45,560 --> 00:13:47,880 Where do you think research in this area should go next? 402 00:13:47,880 --> 00:13:49,960 What are some of the big questions we still need to answer? 403 00:13:49,960 --> 00:13:52,240 That's the million dollar question. 404 00:13:52,240 --> 00:13:53,720 I think one of the biggest challenges 405 00:13:53,720 --> 00:13:56,840 is figuring out how to evaluate AI factuality 406 00:13:56,840 --> 00:13:59,480 in more open-ended situations. 407 00:13:59,480 --> 00:14:01,120 Like how do you measure accuracy 408 00:14:01,120 --> 00:14:03,920 when there isn't just one single right answer? 409 00:14:03,920 --> 00:14:07,240 Think about summarizing a long document or writing an essay. 410 00:14:07,240 --> 00:14:09,800 Can we develop benchmarks for those kinds of tasks? 411 00:14:09,800 --> 00:14:12,120 Yeah, that sounds really tough. 412 00:14:12,120 --> 00:14:15,440 It's like going from multiple choice to essay questions, 413 00:14:15,440 --> 00:14:16,680 way harder to grade. 414 00:14:16,680 --> 00:14:19,840 Any other areas you think are important for future research? 415 00:14:19,840 --> 00:14:20,680 Definitely. 416 00:14:20,680 --> 00:14:22,240 I think we also need to understand 417 00:14:22,240 --> 00:14:25,080 why these AI systems sometimes make stuff up. 418 00:14:25,080 --> 00:14:26,680 Like what's going on under the hood? 419 00:14:26,680 --> 00:14:28,400 Is it because of how they're trained? 420 00:14:28,400 --> 00:14:29,920 Or is there something more fundamental 421 00:14:29,920 --> 00:14:31,600 about how they process information 422 00:14:31,600 --> 00:14:33,720 that makes them prone to these errors? 423 00:14:33,720 --> 00:14:35,040 If we can answer those questions, 424 00:14:35,040 --> 00:14:37,920 I think we'll be a lot closer to developing AI systems 425 00:14:37,920 --> 00:14:40,000 that are truly reliable and trustworthy. 426 00:14:40,000 --> 00:14:41,800 Yeah, makes sense. 427 00:14:41,800 --> 00:14:44,120 If you know why something breaks, 428 00:14:44,120 --> 00:14:46,120 you're more likely to be able to fix it. 429 00:14:46,120 --> 00:14:47,800 Well, this has been a fascinating deep dive 430 00:14:47,800 --> 00:14:49,840 into the world of AI and facts. 431 00:14:49,840 --> 00:14:51,560 Really appreciate your insights on this. 432 00:14:51,560 --> 00:14:52,400 My pleasure. 433 00:14:52,400 --> 00:14:54,680 Any final thoughts for our listeners before we wrap up? 434 00:14:54,680 --> 00:14:57,480 Hmm, I guess the main takeaway is that this whole AI 435 00:14:57,480 --> 00:14:59,920 factuality thing is a really complex issue. 436 00:14:59,920 --> 00:15:01,720 And it's definitely still a work in progress. 437 00:15:01,720 --> 00:15:04,760 We're making some good headway with tools like SimpleQA. 438 00:15:04,760 --> 00:15:06,080 But there's still a lot more to do. 439 00:15:06,080 --> 00:15:07,480 And I think it's important for everyone, 440 00:15:07,480 --> 00:15:10,160 not just AI experts, to understand these limitations 441 00:15:10,160 --> 00:15:11,400 and to be part of the conversation 442 00:15:11,400 --> 00:15:13,880 about how we can make sure AI is used responsibly 443 00:15:13,880 --> 00:15:15,080 and ethically. 444 00:15:15,080 --> 00:15:16,640 You know, it's a big deal. 445 00:15:16,640 --> 00:15:17,800 Absolutely. 446 00:15:17,800 --> 00:15:18,800 Well said. 447 00:15:18,800 --> 00:15:20,200 All right, on that note, we're gonna wrap up 448 00:15:20,200 --> 00:15:21,360 this episode of the deep dive. 449 00:15:21,360 --> 00:15:24,160 Thanks for joining us on this exploration of AI and facts. 450 00:15:24,160 --> 00:15:27,600 And until next time, stay curious.