1 00:00:00,000 --> 00:00:04,640 Okay, so can AI actually predict elections? 2 00:00:04,640 --> 00:00:05,920 What's a really interesting question? 3 00:00:05,920 --> 00:00:08,720 And this paper you sent over from Brigham Young University 4 00:00:08,720 --> 00:00:09,840 dives right into that. 5 00:00:09,840 --> 00:00:10,680 Yeah, caught my eye. 6 00:00:10,680 --> 00:00:14,200 It's about using something called distribution-based prediction 7 00:00:14,200 --> 00:00:17,200 to forecast election results. 8 00:00:17,200 --> 00:00:20,320 Right, it's using large language models or LLMs 9 00:00:20,320 --> 00:00:22,440 to try and analyze voting trends. 10 00:00:22,440 --> 00:00:23,800 LLMs. Yeah, LLMs. 11 00:00:23,800 --> 00:00:26,400 For those of us who aren't AI experts, 12 00:00:26,400 --> 00:00:27,800 can you break that down a little? 13 00:00:27,800 --> 00:00:33,200 Yeah, so you can think of LLMs as really powerful AI systems 14 00:00:33,200 --> 00:00:36,200 that are trained on tons and tons of text data. 15 00:00:36,200 --> 00:00:36,520 Okay. 16 00:00:36,520 --> 00:00:38,480 And from that, they're able to understand language, 17 00:00:38,480 --> 00:00:40,880 generate text, and even make predictions 18 00:00:40,880 --> 00:00:43,240 based on the patterns they've learned. 19 00:00:43,240 --> 00:00:45,080 So is this basically like instead of trying 20 00:00:45,080 --> 00:00:46,680 to predict how each person will vote, 21 00:00:46,680 --> 00:00:48,400 it looks at broader voting patterns? 22 00:00:48,400 --> 00:00:49,400 Exactly. 23 00:00:49,400 --> 00:00:51,880 It's not about simulating individual voters, 24 00:00:51,880 --> 00:00:55,160 which is really hard and can be like really prone to bias. 25 00:00:55,160 --> 00:00:58,200 It's more about understanding the overall distribution of votes 26 00:00:58,200 --> 00:01:00,320 across different states and demographics. 27 00:01:00,320 --> 00:01:01,360 OK, that makes sense. 28 00:01:01,360 --> 00:01:03,520 So how did they actually test this method out? 29 00:01:03,520 --> 00:01:06,680 Did they use a real election? 30 00:01:06,680 --> 00:01:07,360 They did. 31 00:01:07,360 --> 00:01:10,480 They used the 2020 US presidential election. 32 00:01:10,480 --> 00:01:13,960 They fed the LLM all kinds of data about the election, 33 00:01:13,960 --> 00:01:18,040 so things like demographics, historical voting patterns, 34 00:01:18,040 --> 00:01:19,680 even social media trends. 35 00:01:19,680 --> 00:01:21,160 OK, so what happened? 36 00:01:21,160 --> 00:01:24,120 Was the AI able to predict who won? 37 00:01:24,120 --> 00:01:27,440 The results were surprisingly accurate. 38 00:01:27,440 --> 00:01:29,560 On average, the model's predictions 39 00:01:29,560 --> 00:01:32,080 were only off by less than half a percentage point 40 00:01:32,080 --> 00:01:35,000 for both Biden and Trump and in each state. 41 00:01:35,000 --> 00:01:36,800 Wow, that's really close. 42 00:01:36,800 --> 00:01:37,080 Yeah. 43 00:01:37,080 --> 00:01:38,960 They must have been pretty excited about those results. 44 00:01:38,960 --> 00:01:39,720 Definitely. 45 00:01:39,720 --> 00:01:42,360 I mean, it suggests that this could be a really powerful tool 46 00:01:42,360 --> 00:01:44,920 for understanding and predicting elections. 47 00:01:44,920 --> 00:01:48,520 But like with anything, there are limitations. 48 00:01:48,520 --> 00:01:49,720 Of course, like what? 49 00:01:49,720 --> 00:01:52,280 Well, for one thing, LLMs are always learning and changing, 50 00:01:52,280 --> 00:01:52,640 right? 51 00:01:52,640 --> 00:01:53,000 Yeah. 52 00:01:53,000 --> 00:01:54,960 So their predictions are only as good as the data 53 00:01:54,960 --> 00:01:56,080 they were trained on. 54 00:01:56,080 --> 00:01:58,280 If there's a major shift in public opinion 55 00:01:58,280 --> 00:02:00,880 or, I don't know, some unforeseen event, 56 00:02:00,880 --> 00:02:03,080 then those predictions could be way off. 57 00:02:03,080 --> 00:02:05,080 Right, like a global pandemic, maybe. 58 00:02:05,080 --> 00:02:06,720 Yeah, exactly. 59 00:02:06,720 --> 00:02:08,840 No one could have predicted how COVID would impact 60 00:02:08,840 --> 00:02:10,120 the 2020 election. 61 00:02:10,120 --> 00:02:10,880 That's true. 62 00:02:10,880 --> 00:02:13,000 It's a good reminder that we need 63 00:02:13,000 --> 00:02:15,640 to be cautious about making predictions about the future. 64 00:02:15,640 --> 00:02:18,480 Especially about something as complicated as politics. 65 00:02:18,480 --> 00:02:20,000 Especially about politics. 66 00:02:20,000 --> 00:02:20,880 Absolutely. 67 00:02:20,880 --> 00:02:21,380 Yeah. 68 00:02:21,380 --> 00:02:24,360 So I mean, it's amazing that AI can predict an election 69 00:02:24,360 --> 00:02:26,160 with that level of accuracy. 70 00:02:26,160 --> 00:02:27,800 Yeah, it really is. 71 00:02:27,800 --> 00:02:29,720 And it makes you think about what 72 00:02:29,720 --> 00:02:32,520 this means for the future of AI and how it might impact 73 00:02:32,520 --> 00:02:33,760 politics as a whole. 74 00:02:33,760 --> 00:02:34,720 Totally. 75 00:02:34,720 --> 00:02:36,480 Speaking of the future, did they try 76 00:02:36,480 --> 00:02:39,160 to predict any future elections with this model? 77 00:02:39,160 --> 00:02:39,960 They did. 78 00:02:39,960 --> 00:02:44,240 They used the model to simulate a hypothetical 2024 election 79 00:02:44,240 --> 00:02:46,120 between Trump and Kamala Harris. 80 00:02:46,120 --> 00:02:47,240 Oh, that's interesting. 81 00:02:47,240 --> 00:02:49,240 So who did the AI think would win? 82 00:02:49,240 --> 00:02:51,160 Well, I don't want to spoil the surprise 83 00:02:51,160 --> 00:02:53,360 just yet, but let's just say the results are pretty interesting. 84 00:02:53,360 --> 00:02:54,080 OK, all right. 85 00:02:54,080 --> 00:02:55,560 You've definitely piqued my interest. 86 00:02:55,560 --> 00:02:58,280 But before we get into the 2024 predictions, 87 00:02:58,280 --> 00:02:59,640 can we back up a bit? 88 00:02:59,640 --> 00:03:02,800 I'd love to hear a little more about how this distribution 89 00:03:02,800 --> 00:03:05,520 based prediction method actually works. 90 00:03:05,520 --> 00:03:05,880 Sure. 91 00:03:05,880 --> 00:03:07,680 So one of the key concepts here is 92 00:03:07,680 --> 00:03:11,280 that instead of focusing on individual voters, 93 00:03:11,280 --> 00:03:14,120 it's all about the overall probability distribution 94 00:03:14,120 --> 00:03:15,080 of votes. 95 00:03:15,080 --> 00:03:16,360 Probability distribution. 96 00:03:16,360 --> 00:03:18,080 Yeah, probability distribution. 97 00:03:18,080 --> 00:03:20,120 Can you explain that a bit more? 98 00:03:20,120 --> 00:03:20,800 Sure. 99 00:03:20,800 --> 00:03:22,600 Imagine you're flipping a coin. 100 00:03:22,600 --> 00:03:25,040 You have a 50% chance of getting heads, 101 00:03:25,040 --> 00:03:27,120 50% chance of getting tails. 102 00:03:27,120 --> 00:03:29,440 That's a probability distribution. 103 00:03:29,440 --> 00:03:32,160 In this case, the AI is trying to predict 104 00:03:32,160 --> 00:03:34,080 the probability distribution of votes 105 00:03:34,080 --> 00:03:36,160 for each candidate in each state. 106 00:03:36,160 --> 00:03:38,280 Instead of saying, you know, this person 107 00:03:38,280 --> 00:03:39,640 will vote for this candidate, it's 108 00:03:39,640 --> 00:03:42,120 saying there is a certain percentage chance 109 00:03:42,120 --> 00:03:44,080 that this candidate will win this state. 110 00:03:44,080 --> 00:03:44,560 Gotcha. 111 00:03:44,560 --> 00:03:48,240 So it's like it's a more nuanced way 112 00:03:48,240 --> 00:03:50,560 of thinking about election forecasting. 113 00:03:50,560 --> 00:03:51,080 Yeah, I guess. 114 00:03:51,080 --> 00:03:53,000 It's like you need to account all the uncertainties and stuff. 115 00:03:53,000 --> 00:03:53,560 Exactly. 116 00:03:53,560 --> 00:03:54,600 It's not just black and white. 117 00:03:54,600 --> 00:03:55,040 Right. 118 00:03:55,040 --> 00:03:57,320 It acknowledges that elections are complex. 119 00:03:57,320 --> 00:03:58,680 OK, I think I'm starting to get it. 120 00:03:58,680 --> 00:03:59,360 Good. 121 00:03:59,360 --> 00:04:01,320 But this is all based on historical data. 122 00:04:01,320 --> 00:04:01,760 Right. 123 00:04:01,760 --> 00:04:03,680 Like how do we know that these same patterns will 124 00:04:03,680 --> 00:04:04,920 hold true in the future? 125 00:04:04,920 --> 00:04:05,880 That's a great question. 126 00:04:05,880 --> 00:04:07,320 And it's one of the biggest challenges 127 00:04:07,320 --> 00:04:10,920 for any election forecasting model. 128 00:04:10,920 --> 00:04:13,640 The political world changes constantly. 129 00:04:13,640 --> 00:04:16,280 And there's always unforeseen events, 130 00:04:16,280 --> 00:04:19,720 things that disrupt even the most established trends. 131 00:04:19,720 --> 00:04:20,880 That's a good point. 132 00:04:20,880 --> 00:04:22,520 But it's still really cool to see 133 00:04:22,520 --> 00:04:27,360 how AI is being used to analyze and predict elections. 134 00:04:27,360 --> 00:04:28,800 It feels like, I don't know, we're 135 00:04:28,800 --> 00:04:31,920 entering a whole new era of political forecasting. 136 00:04:31,920 --> 00:04:33,280 It does feel that way. 137 00:04:33,280 --> 00:04:35,680 This research is really just the beginning, I think. 138 00:04:35,680 --> 00:04:39,160 We're probably going to see even more advanced AI models being 139 00:04:39,160 --> 00:04:40,800 developed in the future. 140 00:04:40,800 --> 00:04:43,000 You know, that could completely change how we understand 141 00:04:43,000 --> 00:04:44,280 and predict elections. 142 00:04:44,280 --> 00:04:46,360 OK, so we've talked about the 2020 election. 143 00:04:46,360 --> 00:04:49,360 We've talked about the limitations of AI prediction models. 144 00:04:49,360 --> 00:04:51,120 And we've got this glimpse into the future. 145 00:04:51,120 --> 00:04:51,640 Right. 146 00:04:51,640 --> 00:04:54,040 Now I'm dying to hear what the AI predicted 147 00:04:54,040 --> 00:04:58,280 for that hypothetical 2024 matchup between Trump and Harris. 148 00:04:58,280 --> 00:04:59,560 OK, let's get into it. 149 00:04:59,560 --> 00:05:02,480 What's really cool is that the AI didn't just give one answer. 150 00:05:02,480 --> 00:05:04,800 It generated a whole range of possible outcomes. 151 00:05:04,800 --> 00:05:06,880 And each one had its own probability. 152 00:05:06,880 --> 00:05:08,600 So it's not like a crystal ball where 153 00:05:08,600 --> 00:05:10,080 you get a definite answer. 154 00:05:10,080 --> 00:05:10,960 No, not at all. 155 00:05:10,960 --> 00:05:12,440 It's more like a weather forecast. 156 00:05:12,440 --> 00:05:12,760 It's OK. 157 00:05:12,760 --> 00:05:15,400 Where you see a range of possible temperatures, 158 00:05:15,400 --> 00:05:17,600 you know, precipitation levels. 159 00:05:17,600 --> 00:05:19,880 It's about probabilities, not certainties. 160 00:05:19,880 --> 00:05:20,680 That makes sense. 161 00:05:20,680 --> 00:05:21,160 Yeah. 162 00:05:21,160 --> 00:05:23,440 So walk me through those 2024 predictions. 163 00:05:23,440 --> 00:05:26,560 Like, what did the AI see as the most likely scenario? 164 00:05:26,560 --> 00:05:27,120 Sure. 165 00:05:27,120 --> 00:05:29,760 But first, let me quickly explain what the AI considered 166 00:05:29,760 --> 00:05:31,160 when it was making these predictions. 167 00:05:31,160 --> 00:05:31,560 OK. 168 00:05:31,560 --> 00:05:33,800 So it looked at a ton of factors, things 169 00:05:33,800 --> 00:05:37,520 like historical voting patterns, demographic trends, 170 00:05:37,520 --> 00:05:40,280 even stuff like social media sentiment. 171 00:05:40,280 --> 00:05:41,840 It used all that to create what's 172 00:05:41,840 --> 00:05:45,080 called a probabilistic model of the 2024 election. 173 00:05:45,080 --> 00:05:46,480 So it's not just based on one thing. 174 00:05:46,480 --> 00:05:48,360 It's looking at all these different factors. 175 00:05:48,360 --> 00:05:48,860 Exactly. 176 00:05:48,860 --> 00:05:51,880 It's not just extrapolating from past election results. 177 00:05:51,880 --> 00:05:52,280 Right. 178 00:05:52,280 --> 00:05:53,720 It's more complex than that. 179 00:05:53,720 --> 00:05:54,160 OK. 180 00:05:54,160 --> 00:05:54,360 OK. 181 00:05:54,360 --> 00:05:57,320 So are you ready to share what the AI predicted for each 182 00:05:57,320 --> 00:05:57,800 state? 183 00:05:57,800 --> 00:05:58,480 I am. 184 00:05:58,480 --> 00:05:59,480 But are you ready to hear it? 185 00:05:59,480 --> 00:06:00,160 Absolutely. 186 00:06:00,160 --> 00:06:01,080 Lay it on me. 187 00:06:01,080 --> 00:06:05,800 What did the AI see happening in those key battleground 188 00:06:05,800 --> 00:06:06,680 states? 189 00:06:06,680 --> 00:06:10,080 Well, in states like Pennsylvania, Michigan, Wisconsin, 190 00:06:10,080 --> 00:06:12,400 it predicted really tight races. 191 00:06:12,400 --> 00:06:14,520 The margins were so small that it's 192 00:06:14,520 --> 00:06:17,560 almost impossible to say for sure who would win. 193 00:06:17,560 --> 00:06:20,600 So those states are still just as unpredictable as ever. 194 00:06:20,600 --> 00:06:22,200 It seems that way. 195 00:06:22,200 --> 00:06:25,160 The AI also predicted some surprises, though, 196 00:06:25,160 --> 00:06:29,240 in states that were considered safe bets for one party. 197 00:06:29,240 --> 00:06:30,320 Really? 198 00:06:30,320 --> 00:06:31,000 Like what? 199 00:06:31,000 --> 00:06:32,640 Well, in some cases, it predicted 200 00:06:32,640 --> 00:06:35,600 that states that have always voted a certain way 201 00:06:35,600 --> 00:06:39,000 might actually flip to the other party in 2024. 202 00:06:39,000 --> 00:06:39,640 Whoa. 203 00:06:39,640 --> 00:06:41,240 That would be a major shakeup. 204 00:06:41,240 --> 00:06:44,560 What do you think is causing those potential shifts? 205 00:06:44,560 --> 00:06:46,800 It's hard to say for sure. 206 00:06:46,800 --> 00:06:48,720 But it seems like the AI is picking up 207 00:06:48,720 --> 00:06:52,200 on these subtle shifts in demographics and voting patterns. 208 00:06:52,200 --> 00:06:53,600 You know, things that could mean a bigger 209 00:06:53,600 --> 00:06:55,960 realignment in American politics is happening. 210 00:06:55,960 --> 00:06:56,880 Fascinating. 211 00:06:56,880 --> 00:06:59,840 So this distribution-based prediction method, 212 00:06:59,840 --> 00:07:02,360 it's not just about predicting the overall winner. 213 00:07:02,360 --> 00:07:04,960 It's also about understanding these nuances and potential 214 00:07:04,960 --> 00:07:06,480 surprises at the state level. 215 00:07:06,480 --> 00:07:07,000 Exactly. 216 00:07:07,000 --> 00:07:09,320 It gives us a much more detailed view of what's going on. 217 00:07:09,320 --> 00:07:09,560 OK. 218 00:07:09,560 --> 00:07:11,600 So we've talked about the state level predictions. 219 00:07:11,600 --> 00:07:13,920 But I'm still waiting to hear who the AI thinks 220 00:07:13,920 --> 00:07:16,240 will win the whole thing in 2024. 221 00:07:16,240 --> 00:07:18,480 Are you ready to reveal the big prediction? 222 00:07:18,480 --> 00:07:19,160 I am. 223 00:07:19,160 --> 00:07:22,160 But remember, this is just one possible outcome. 224 00:07:22,160 --> 00:07:22,480 Right. 225 00:07:22,480 --> 00:07:24,920 The AI created a bunch of different scenarios. 226 00:07:24,920 --> 00:07:27,560 And this is just the one it considered the most likely 227 00:07:27,560 --> 00:07:29,280 based on what it knew at the time. 228 00:07:29,280 --> 00:07:29,800 OK. 229 00:07:29,800 --> 00:07:30,760 Fair enough. 230 00:07:30,760 --> 00:07:32,600 So drumroll, please. 231 00:07:32,600 --> 00:07:36,440 Who did the AI predict would win the 2024 presidential election? 232 00:07:36,440 --> 00:07:39,560 Well, according to the AI, the most likely scenario is that 233 00:07:39,560 --> 00:07:42,040 Kamala Harris would win the 2024 election 234 00:07:42,040 --> 00:07:44,800 with 303 electoral votes. 235 00:07:44,800 --> 00:07:45,200 Wow. 236 00:07:45,200 --> 00:07:47,800 That's a pretty decisive victory. 237 00:07:47,800 --> 00:07:49,800 But I guess it kind of makes sense, right, 238 00:07:49,800 --> 00:07:51,520 considering some of the state level predictions 239 00:07:51,520 --> 00:07:52,360 we just talked about. 240 00:07:52,360 --> 00:07:53,400 Yeah, for sure. 241 00:07:53,400 --> 00:07:55,080 The model basically showed Harris 242 00:07:55,080 --> 00:07:58,040 at the slight edge in a lot of those battleground states. 243 00:07:58,040 --> 00:08:00,040 And that ultimately pushed her to victory. 244 00:08:00,040 --> 00:08:02,880 So you're telling me AI is predicting a female president 245 00:08:02,880 --> 00:08:04,000 in 2024. 246 00:08:04,000 --> 00:08:05,440 That's what the model suggests. 247 00:08:05,440 --> 00:08:07,840 Oh, that's pretty groundbreaking. 248 00:08:07,840 --> 00:08:08,560 It is. 249 00:08:08,560 --> 00:08:11,200 And I think it's a good reminder that the political landscape 250 00:08:11,200 --> 00:08:12,640 is constantly changing. 251 00:08:12,640 --> 00:08:15,440 What seems unlikely now maybe won't in a few years. 252 00:08:15,440 --> 00:08:16,480 That's true. 253 00:08:16,480 --> 00:08:19,520 We can't forget, though, that this is just one possible future, 254 00:08:19,520 --> 00:08:19,840 right? 255 00:08:19,840 --> 00:08:22,720 A lot can happen between now and 2024. 256 00:08:22,720 --> 00:08:23,480 Absolutely. 257 00:08:23,480 --> 00:08:25,360 AI predictions aren't guarantees. 258 00:08:25,360 --> 00:08:27,360 They're more like tools. 259 00:08:27,360 --> 00:08:31,080 They help us understand how things might play out 260 00:08:31,080 --> 00:08:33,680 and then hopefully make better decisions based on that. 261 00:08:33,680 --> 00:08:36,080 Speaking of informed decisions, did the researchers 262 00:08:36,080 --> 00:08:40,200 just test out Trump versus Harris? 263 00:08:40,200 --> 00:08:43,240 Or did they run this model for other potential matchups, too? 264 00:08:43,240 --> 00:08:43,760 Oh, yeah. 265 00:08:43,760 --> 00:08:47,040 They ran simulations for a bunch of different possible matchups. 266 00:08:47,040 --> 00:08:49,560 They wanted to see how the AI would handle 267 00:08:49,560 --> 00:08:51,640 different political situations. 268 00:08:51,640 --> 00:08:56,640 So I mean, did they find that the AI always favored Democrats? 269 00:08:56,640 --> 00:09:00,080 Or were there some matchups where Republicans came out on top? 270 00:09:00,080 --> 00:09:01,400 That's a really great question. 271 00:09:01,400 --> 00:09:03,400 And it actually brings up one of the biggest challenges 272 00:09:03,400 --> 00:09:04,800 with AI bias. 273 00:09:04,800 --> 00:09:05,280 Bias. 274 00:09:05,280 --> 00:09:06,240 OK, how so? 275 00:09:06,240 --> 00:09:07,120 Well, think about it. 276 00:09:07,120 --> 00:09:10,480 AI models are trained on huge amounts of data, right? 277 00:09:10,480 --> 00:09:10,960 Right. 278 00:09:10,960 --> 00:09:13,640 And that data, it can reflect the biases that 279 00:09:13,640 --> 00:09:15,120 exiled in the real world. 280 00:09:15,120 --> 00:09:17,680 So if the data the AI is trained on 281 00:09:17,680 --> 00:09:21,040 has a certain political leaning, then its predictions 282 00:09:21,040 --> 00:09:22,520 might also have that same leaning. 283 00:09:22,520 --> 00:09:23,200 Exactly. 284 00:09:23,200 --> 00:09:25,760 It's kind of like if you only watched one news channel, 285 00:09:25,760 --> 00:09:27,760 you might start to believe that their perspective is 286 00:09:27,760 --> 00:09:29,480 the only one that matters, right? 287 00:09:29,480 --> 00:09:31,720 Well, AI can fall into that same trap. 288 00:09:31,720 --> 00:09:33,240 It's like an echo chamber effect. 289 00:09:33,240 --> 00:09:33,880 Oh, I see. 290 00:09:33,880 --> 00:09:37,000 So how did the researchers deal with this bias 291 00:09:37,000 --> 00:09:38,720 issue in their model? 292 00:09:38,720 --> 00:09:41,320 They knew it was a big concern, so they took steps 293 00:09:41,320 --> 00:09:42,840 to try and minimize it. 294 00:09:42,840 --> 00:09:45,240 For one, they were super careful about the data 295 00:09:45,240 --> 00:09:46,800 the AI was trained on. 296 00:09:46,800 --> 00:09:49,560 They made sure it included lots of different perspectives, 297 00:09:49,560 --> 00:09:51,520 not just one side of the story. 298 00:09:51,520 --> 00:09:54,800 And they also tested the model on a bunch of different scenarios 299 00:09:54,800 --> 00:09:57,760 to see if it consistently favored one party. 300 00:09:57,760 --> 00:09:59,840 Did they find any bias? 301 00:09:59,840 --> 00:10:01,000 It's interesting, actually. 302 00:10:01,000 --> 00:10:03,760 The AI did seem to have a slight preference 303 00:10:03,760 --> 00:10:06,880 for Democratic candidates in most of the matchups. 304 00:10:06,880 --> 00:10:09,080 So even though they tried to be careful, 305 00:10:09,080 --> 00:10:10,920 some bias still snuck in. 306 00:10:10,920 --> 00:10:12,000 It seems that way. 307 00:10:12,000 --> 00:10:15,920 And it just goes to show how tricky this issue of AI bias is. 308 00:10:15,920 --> 00:10:17,040 There's no easy thing. 309 00:10:17,040 --> 00:10:23,160 I think even if the AI has a little bit of bias, 310 00:10:23,160 --> 00:10:25,840 it doesn't mean its predictors are totally worthless. 311 00:10:25,840 --> 00:10:27,680 Right, exactly. 312 00:10:27,680 --> 00:10:29,040 We should think of these predictions 313 00:10:29,040 --> 00:10:30,680 as more of a starting point. 314 00:10:30,680 --> 00:10:32,960 Something to get us thinking and talking 315 00:10:32,960 --> 00:10:35,680 and to help us figure out where we need to learn more. 316 00:10:35,680 --> 00:10:37,840 So don't take them as gospel, but we can still 317 00:10:37,840 --> 00:10:39,000 learn something from them. 318 00:10:39,000 --> 00:10:39,880 Exactly. 319 00:10:39,880 --> 00:10:43,880 It's about using AI as a tool to help us understand things 320 00:10:43,880 --> 00:10:47,640 better, not to replace human judgment altogether. 321 00:10:47,640 --> 00:10:48,640 Makes sense. 322 00:10:48,640 --> 00:10:50,800 Now, I want to go back to something you mentioned earlier 323 00:10:50,800 --> 00:10:55,480 about how the AI came up with a bunch of different outcomes 324 00:10:55,480 --> 00:10:56,920 for the 2024 election. 325 00:10:56,920 --> 00:10:59,400 Oh, yeah, the range of possible outcomes. 326 00:10:59,400 --> 00:11:00,840 Can you explain that a little more? 327 00:11:00,840 --> 00:11:01,640 Sure. 328 00:11:01,640 --> 00:11:05,280 The researchers didn't just run the model once and say, OK, 329 00:11:05,280 --> 00:11:05,920 we're done. 330 00:11:05,920 --> 00:11:10,000 They ran it thousands of times, each time tweaking 331 00:11:10,000 --> 00:11:11,400 the parameters a little bit. 332 00:11:11,400 --> 00:11:14,600 OK, so instead of saying this candidate will definitely win, 333 00:11:14,600 --> 00:11:17,000 it's like saying, OK, there's a chance this candidate will win, 334 00:11:17,000 --> 00:11:19,200 but here are some other things that could happen. 335 00:11:19,200 --> 00:11:20,160 Yeah, exactly. 336 00:11:20,160 --> 00:11:22,640 That's why this approach is so valuable, I think, 337 00:11:22,640 --> 00:11:24,520 because it acknowledges that there's always 338 00:11:24,520 --> 00:11:26,320 uncertainty in an election. 339 00:11:26,320 --> 00:11:27,320 Right, for sure. 340 00:11:27,320 --> 00:11:30,080 Like, there are so many things that can influence the outcome, 341 00:11:30,080 --> 00:11:32,040 you can never be 100% sure. 342 00:11:32,040 --> 00:11:34,560 Right, so much can change between now and 2024. 343 00:11:34,560 --> 00:11:35,480 Oh, absolutely. 344 00:11:35,480 --> 00:11:39,480 Some big news story, a sudden economic downturn, 345 00:11:39,480 --> 00:11:43,160 even like a celebrity endorsement can totally change things. 346 00:11:43,160 --> 00:11:46,600 It's like, these AI predictions are just snapshots in time. 347 00:11:46,600 --> 00:11:49,840 Exactly, a glimpse into the current state of the race. 348 00:11:49,840 --> 00:11:51,160 Right, not a crystal ball. 349 00:11:51,160 --> 00:11:52,040 Exactly. 350 00:11:52,040 --> 00:11:54,400 Now, you mentioned that they ran the model thousands 351 00:11:54,400 --> 00:11:56,920 of times, each time adjusting the parameters. 352 00:11:56,920 --> 00:11:57,400 Uh-huh. 353 00:11:57,400 --> 00:11:59,360 What exactly do you mean by parameters? 354 00:11:59,360 --> 00:12:00,080 Good question. 355 00:12:00,080 --> 00:12:04,120 So think of parameters as the controls on the AI model. 356 00:12:04,120 --> 00:12:05,800 By adjusting those controls, they 357 00:12:05,800 --> 00:12:09,240 can test out how different factors might affect the election. 358 00:12:09,240 --> 00:12:11,520 So they might adjust the parameters 359 00:12:11,520 --> 00:12:14,320 to see what would happen if voter turnout changed, 360 00:12:14,320 --> 00:12:16,320 or if the economy took a downturn. 361 00:12:16,320 --> 00:12:19,200 Exactly, it's like running different simulations, 362 00:12:19,200 --> 00:12:20,960 seeing how things play out in the model. 363 00:12:20,960 --> 00:12:22,440 OK, that makes sense. 364 00:12:22,440 --> 00:12:24,720 Sounds like a lot of work went into this research. 365 00:12:24,720 --> 00:12:25,920 Oh, yeah, for sure. 366 00:12:25,920 --> 00:12:27,760 But the researchers were really good about being 367 00:12:27,760 --> 00:12:29,800 transparent about their process. 368 00:12:29,800 --> 00:12:32,440 They shared all their code and data online, 369 00:12:32,440 --> 00:12:34,400 so other researchers can check their work. 370 00:12:34,400 --> 00:12:34,800 That's great. 371 00:12:34,800 --> 00:12:36,400 Transparency is always important. 372 00:12:36,400 --> 00:12:38,120 But especially when we're talking about something 373 00:12:38,120 --> 00:12:40,280 as sensitive as election prediction, right? 374 00:12:40,280 --> 00:12:41,000 Absolutely. 375 00:12:41,000 --> 00:12:43,120 It means other experts can look at it, 376 00:12:43,120 --> 00:12:44,840 point out any potential problems, 377 00:12:44,840 --> 00:12:46,520 and hopefully build on it too. 378 00:12:46,520 --> 00:12:48,360 So this research is really just the start 379 00:12:48,360 --> 00:12:51,960 of a bigger conversation about AI and elections, huh? 380 00:12:51,960 --> 00:12:52,840 I think so. 381 00:12:52,840 --> 00:12:55,480 And it's a conversation we need to be having now, you know? 382 00:12:55,480 --> 00:12:56,000 Yeah. 383 00:12:56,000 --> 00:12:59,680 As AI gets more powerful and becomes more common, 384 00:12:59,680 --> 00:13:02,280 we need to understand how it might affect democracy. 385 00:13:02,280 --> 00:13:03,000 Totally agree. 386 00:13:03,000 --> 00:13:05,040 OK, before we move on, I want to circle back 387 00:13:05,040 --> 00:13:07,040 to something we talked about earlier, 388 00:13:07,040 --> 00:13:10,480 that potential for bias in AI models. 389 00:13:10,480 --> 00:13:11,760 Oh, yeah, for sure. 390 00:13:11,760 --> 00:13:12,840 It's super important. 391 00:13:12,840 --> 00:13:14,240 We can't just ignore it, right? 392 00:13:14,240 --> 00:13:17,120 No, we have to actively try to find it and fix it. 393 00:13:17,120 --> 00:13:19,160 So what are some things researchers are doing to try 394 00:13:19,160 --> 00:13:22,080 and combat that bias? 395 00:13:22,080 --> 00:13:24,320 One thing they're doing is being a lot more careful 396 00:13:24,320 --> 00:13:26,480 about the data they use to train the models, 397 00:13:26,480 --> 00:13:28,800 making sure it's representative and doesn't just 398 00:13:28,800 --> 00:13:30,640 repeat existing biases. 399 00:13:30,640 --> 00:13:32,800 So instead of just grabbing data from the internet, 400 00:13:32,800 --> 00:13:35,920 which can be full of all sorts of, you know, bias stuff, 401 00:13:35,920 --> 00:13:36,480 Exactly. 402 00:13:36,480 --> 00:13:38,360 They're looking for ways to get data that's 403 00:13:38,360 --> 00:13:39,920 more balanced and accurate. 404 00:13:39,920 --> 00:13:41,040 Yes. 405 00:13:41,040 --> 00:13:43,600 And they're also working on developing new algorithms, 406 00:13:43,600 --> 00:13:46,320 you know, ones that are more resistant to bias. 407 00:13:46,320 --> 00:13:46,840 OK. 408 00:13:46,840 --> 00:13:48,840 There's a lot of cool research happening there, 409 00:13:48,840 --> 00:13:51,160 and we're starting to see some really promising results. 410 00:13:51,160 --> 00:13:52,040 That's good to hear. 411 00:13:52,040 --> 00:13:52,560 Yeah. 412 00:13:52,560 --> 00:13:55,480 But it sounds like even with those efforts, 413 00:13:55,480 --> 00:13:58,760 dealing with bias in AI is a constant challenge, right? 414 00:13:58,760 --> 00:13:59,200 It is. 415 00:13:59,200 --> 00:14:00,640 It's not a one-time fix. 416 00:14:00,640 --> 00:14:03,120 It takes constant vigilance, you know, 417 00:14:03,120 --> 00:14:05,760 being open about how things work and always questioning 418 00:14:05,760 --> 00:14:06,680 our assumptions. 419 00:14:06,680 --> 00:14:07,000 OK. 420 00:14:07,000 --> 00:14:08,120 That makes sense. 421 00:14:08,120 --> 00:14:09,480 We've covered a lot of ground today. 422 00:14:09,480 --> 00:14:12,360 We've talked about how this new prediction method works, 423 00:14:12,360 --> 00:14:15,400 what the AI thinks about the 2024 election, 424 00:14:15,400 --> 00:14:17,800 and the whole complicated issue of bias. 425 00:14:17,800 --> 00:14:18,840 Right. 426 00:14:18,840 --> 00:14:21,280 Before we wrap up, I'm curious to hear, 427 00:14:21,280 --> 00:14:23,040 like, what are your big takeaways from all this? 428 00:14:23,040 --> 00:14:24,880 What should listeners be thinking about? 429 00:14:24,880 --> 00:14:27,320 I think the biggest takeaway is that AI 430 00:14:27,320 --> 00:14:29,920 is getting really good at analyzing data 431 00:14:29,920 --> 00:14:32,320 and making predictions, even about something 432 00:14:32,320 --> 00:14:34,400 as complex as an election. 433 00:14:34,400 --> 00:14:35,920 But it's not a magic solution. 434 00:14:35,920 --> 00:14:36,200 Yep. 435 00:14:36,200 --> 00:14:38,600 It's not like we can just blindly trust the AI 436 00:14:38,600 --> 00:14:42,080 without understanding, you know, its limitations. 437 00:14:42,080 --> 00:14:42,680 Exactly. 438 00:14:42,680 --> 00:14:44,720 And all those potential biases we talked about. 439 00:14:44,720 --> 00:14:45,440 Exactly. 440 00:14:45,440 --> 00:14:49,120 We got to be skeptical, think critically, and remember, 441 00:14:49,120 --> 00:14:50,800 AI is a tool. 442 00:14:50,800 --> 00:14:51,080 Right. 443 00:14:51,080 --> 00:14:53,120 We need to use it responsibly, not 444 00:14:53,120 --> 00:14:54,440 let it make decisions for us. 445 00:14:54,440 --> 00:14:55,200 Totally. 446 00:14:55,200 --> 00:14:57,080 This research is really fascinating. 447 00:14:57,080 --> 00:14:59,960 And it gives us a glimpse into what election forecasting might 448 00:14:59,960 --> 00:15:01,640 look like in the future. 449 00:15:01,640 --> 00:15:03,600 But it's not a perfect system. 450 00:15:03,600 --> 00:15:06,200 We need to use it in a smart and ethical way. 451 00:15:06,200 --> 00:15:07,000 Exactly. 452 00:15:07,000 --> 00:15:10,040 And as we develop even more powerful AI, 453 00:15:10,040 --> 00:15:12,160 it's going to be even more important to think about things 454 00:15:12,160 --> 00:15:15,280 like fairness, transparency, and accountability. 455 00:15:15,280 --> 00:15:16,320 Absolutely. 456 00:15:16,320 --> 00:15:18,240 Well, before we go, the researchers also 457 00:15:18,240 --> 00:15:20,920 looked at how the size of the AI model affected 458 00:15:20,920 --> 00:15:22,760 how accurate the predictions were. 459 00:15:22,760 --> 00:15:23,640 Oh, right. 460 00:15:23,640 --> 00:15:25,160 That's another interesting part of this. 461 00:15:25,160 --> 00:15:26,640 Can you tell me a bit more about that? 462 00:15:26,640 --> 00:15:26,920 Yeah. 463 00:15:26,920 --> 00:15:30,000 So they experimented with different sizes of LLMs. 464 00:15:30,000 --> 00:15:32,440 Some were small, some were massive, you know, 465 00:15:32,440 --> 00:15:34,040 with billions of parameters. 466 00:15:34,040 --> 00:15:35,600 Remind me, what are parameters again? 467 00:15:35,600 --> 00:15:38,480 Oh, think of them as the building blocks of the AI. 468 00:15:38,480 --> 00:15:38,880 OK. 469 00:15:38,880 --> 00:15:41,000 The more parameters a model has, 470 00:15:41,000 --> 00:15:43,280 the more complex stuff it can learn, 471 00:15:43,280 --> 00:15:45,960 and the more detailed its predictions can be. 472 00:15:45,960 --> 00:15:49,280 So basically, the bigger the AI, the better the predictions. 473 00:15:49,280 --> 00:15:51,000 In a lot of cases, yeah. 474 00:15:51,000 --> 00:15:53,720 They found that the bigger models were usually more accurate. 475 00:15:53,720 --> 00:15:56,120 They could see those subtle trends and patterns 476 00:15:56,120 --> 00:15:57,680 that the smaller models missed. 477 00:15:57,680 --> 00:15:58,160 OK. 478 00:15:58,160 --> 00:15:58,760 That makes sense. 479 00:15:58,760 --> 00:16:00,960 Kind of like having a bigger, more powerful telescope. 480 00:16:00,960 --> 00:16:02,200 Exactly. 481 00:16:02,200 --> 00:16:04,440 And as we keep making bigger and better AI, 482 00:16:04,440 --> 00:16:06,440 who knows what they'll be able to predict? 483 00:16:06,440 --> 00:16:08,320 Maybe one day they'll be able to predict elections 484 00:16:08,320 --> 00:16:10,600 with perfect accuracy. 485 00:16:10,600 --> 00:16:12,480 It's definitely possible. 486 00:16:12,480 --> 00:16:15,560 But even then, we'll still need to be careful about those biases 487 00:16:15,560 --> 00:16:17,240 and use AI responsibly. 488 00:16:17,240 --> 00:16:18,960 Good point. 489 00:16:18,960 --> 00:16:21,120 Well, I think we've covered just about everything. 490 00:16:21,120 --> 00:16:22,400 Yeah, I think so too. 491 00:16:22,400 --> 00:16:24,120 Any last thoughts before we wrap up? 492 00:16:24,120 --> 00:16:25,960 Just that this research is really exciting. 493 00:16:25,960 --> 00:16:27,840 It shows us what's possible with AI 494 00:16:27,840 --> 00:16:30,520 and how it can help us understand the world. 495 00:16:30,520 --> 00:16:32,680 But it's a fast-moving field, so we 496 00:16:32,680 --> 00:16:35,320 need to stay informed and keep learning. 497 00:16:35,320 --> 00:16:36,360 I couldn't agree more. 498 00:16:36,360 --> 00:16:37,320 Thanks for joining me today. 499 00:16:37,320 --> 00:16:38,560 It's been a really great discussion. 500 00:16:38,560 --> 00:16:39,160 My pleasure. 501 00:16:39,160 --> 00:16:41,000 It's always fun to dive into these topics. 502 00:16:41,000 --> 00:16:41,560 OK, folks. 503 00:16:41,560 --> 00:16:43,240 That's all the time we have for today. 504 00:16:43,240 --> 00:16:46,840 We hope you enjoyed this deep dive into AI and election 505 00:16:46,840 --> 00:16:47,800 prediction. 506 00:16:47,800 --> 00:16:51,040 Stay tuned for more exciting episodes. 507 00:16:51,040 --> 00:16:54,520 So we've talked a lot about how AI might predict elections. 508 00:16:54,520 --> 00:16:57,440 But you mentioned something earlier, Silicon Sampling, 509 00:16:57,440 --> 00:17:01,240 that the researchers decided not to use for this. 510 00:17:01,240 --> 00:17:03,000 Why'd they scrap that approach? 511 00:17:03,000 --> 00:17:06,040 So at first, they thought about using Silicon Sampling 512 00:17:06,040 --> 00:17:09,640 to create these simulated voters, almost like building 513 00:17:09,640 --> 00:17:13,320 a little world of AI citizens to cast votes. 514 00:17:13,320 --> 00:17:15,960 But they realized pretty quickly that this method was just 515 00:17:15,960 --> 00:17:17,960 it was too likely to result in stereotypes. 516 00:17:17,960 --> 00:17:21,360 So instead of capturing how complex real voters are, 517 00:17:21,360 --> 00:17:24,440 the AI was just making these broad generalizations. 518 00:17:24,440 --> 00:17:25,480 Exactly. 519 00:17:25,480 --> 00:17:27,720 Like, imagine it predicting that everyone 520 00:17:27,720 --> 00:17:30,040 from a certain area with a certain religion 521 00:17:30,040 --> 00:17:31,720 would vote the same way. 522 00:17:31,720 --> 00:17:33,440 That's obviously way too simple, right? 523 00:17:33,440 --> 00:17:36,120 It doesn't show how different people's opinions can 524 00:17:36,120 --> 00:17:37,120 be within any group. 525 00:17:37,120 --> 00:17:40,000 Right, like saying all cat owners vote the same way. 526 00:17:40,000 --> 00:17:42,080 Because I don't think it doesn't make sense. 527 00:17:42,080 --> 00:17:44,480 People's votes depend on so many things. 528 00:17:44,480 --> 00:17:47,200 Other beliefs, experiences, values. 529 00:17:47,200 --> 00:17:49,240 It's way more than just demographics. 530 00:17:49,240 --> 00:17:50,120 Exactly. 531 00:17:50,120 --> 00:17:52,320 And the researchers, they understood that. 532 00:17:52,320 --> 00:17:55,760 They saw that simulating individual voters like that 533 00:17:55,760 --> 00:17:57,840 was just too risky. 534 00:17:57,840 --> 00:18:01,720 It could lead to predictions that were biased and misleading. 535 00:18:01,720 --> 00:18:04,760 So they decided to focus on this distribution-based 536 00:18:04,760 --> 00:18:06,160 prediction instead. 537 00:18:06,160 --> 00:18:08,480 Right, which is a much better approach, I think. 538 00:18:08,480 --> 00:18:11,200 Yeah, it seems a lot more robust. 539 00:18:11,200 --> 00:18:11,800 It is. 540 00:18:11,800 --> 00:18:13,360 Because instead of getting stuck trying 541 00:18:13,360 --> 00:18:15,320 to simulate every single voter, they're 542 00:18:15,320 --> 00:18:17,480 looking at the bigger picture. 543 00:18:17,480 --> 00:18:19,680 They analyze the patterns in the data, 544 00:18:19,680 --> 00:18:23,000 which is much less likely to fall into those stereotypes. 545 00:18:23,000 --> 00:18:25,640 It's like instead of trying to track every single raindrop 546 00:18:25,640 --> 00:18:27,960 in a storm, you're looking at the direction 547 00:18:27,960 --> 00:18:30,960 the storm is moving, how strong it is, 548 00:18:30,960 --> 00:18:32,320 the overall picture. 549 00:18:32,320 --> 00:18:33,600 That's a great analogy. 550 00:18:33,600 --> 00:18:36,200 And it kind of highlights something important about AI, 551 00:18:36,200 --> 00:18:36,480 I think. 552 00:18:36,480 --> 00:18:37,560 That's a app. 553 00:18:37,560 --> 00:18:40,480 We can't just blindly trust it. 554 00:18:40,480 --> 00:18:42,040 We've got to question our assumptions, 555 00:18:42,040 --> 00:18:44,800 make sure we're not just repeating the biases that 556 00:18:44,800 --> 00:18:45,520 are already out there. 557 00:18:45,520 --> 00:18:46,480 Absolutely. 558 00:18:46,480 --> 00:18:48,080 Especially when it comes to something 559 00:18:48,080 --> 00:18:50,480 as important as elections. 560 00:18:50,480 --> 00:18:53,600 We can't just let the AI make the decisions for us. 561 00:18:53,600 --> 00:18:55,120 You have to be thoughtful about it. 562 00:18:55,120 --> 00:18:55,600 Exactly. 563 00:18:55,600 --> 00:18:58,960 Understand the limitations and be critical of the results. 564 00:18:58,960 --> 00:19:01,560 So this research, it wasn't just about finding a new way 565 00:19:01,560 --> 00:19:02,920 to predict elections. 566 00:19:02,920 --> 00:19:06,160 It was also about using AI responsibly and ethically. 567 00:19:06,160 --> 00:19:06,760 Exactly. 568 00:19:06,760 --> 00:19:08,880 It's a great example of how researchers are starting 569 00:19:08,880 --> 00:19:11,000 to think about these things. 570 00:19:11,000 --> 00:19:14,080 And how to make sure AI is developed in a way that's fair. 571 00:19:14,080 --> 00:19:14,760 Yeah. 572 00:19:14,760 --> 00:19:18,000 It's a powerful tool, but it's still a new tool, right? 573 00:19:18,000 --> 00:19:18,520 It is. 574 00:19:18,520 --> 00:19:20,360 We're still figuring out how to use it properly. 575 00:19:20,360 --> 00:19:22,920 And the more powerful AI gets, the more important 576 00:19:22,920 --> 00:19:25,080 it's going to be to keep having these conversations 577 00:19:25,080 --> 00:19:26,800 about how to use it ethically. 578 00:19:26,800 --> 00:19:30,280 OK, so as we wrap up here, what are your main takeaways 579 00:19:30,280 --> 00:19:31,360 from this research? 580 00:19:31,360 --> 00:19:35,840 What do you want listeners to walk away thinking about? 581 00:19:35,840 --> 00:19:38,640 Well, I think the big thing is that AI is becoming really 582 00:19:38,640 --> 00:19:41,720 good at analyzing information and making predictions, 583 00:19:41,720 --> 00:19:44,960 even about something as complicated as elections. 584 00:19:44,960 --> 00:19:48,520 But like we've been saying, it's not a magic crystal ball. 585 00:19:48,520 --> 00:19:52,120 We have to be careful about those limitations and biases. 586 00:19:52,120 --> 00:19:52,600 Absolutely. 587 00:19:52,600 --> 00:19:53,320 Be critical. 588 00:19:53,320 --> 00:19:54,840 Ask questions. 589 00:19:54,840 --> 00:19:56,880 This research is a fascinating look 590 00:19:56,880 --> 00:20:01,760 at what might be possible, but it's not a perfect solution. 591 00:20:01,760 --> 00:20:04,480 We need to be thoughtful and responsible with this kind 592 00:20:04,480 --> 00:20:05,360 of technology. 593 00:20:05,360 --> 00:20:06,880 Well said. 594 00:20:06,880 --> 00:20:10,200 And as we keep developing even more powerful AI, 595 00:20:10,200 --> 00:20:11,800 fairness and transparency are going 596 00:20:11,800 --> 00:20:13,200 to be more important than ever. 597 00:20:13,200 --> 00:20:14,280 Totally agree. 598 00:20:14,280 --> 00:20:16,480 Well, that's all the time we have for today's Deep Dives 599 00:20:16,480 --> 00:20:18,640 into AI and election prediction. 600 00:20:18,640 --> 00:20:20,480 It's been great talking about this with you. 601 00:20:20,480 --> 00:20:21,280 Thanks for joining me. 602 00:20:21,280 --> 00:20:23,320 We hope you learned something new and interesting today. 603 00:20:23,320 --> 00:20:25,320 And remember to stay curious, stay informed, 604 00:20:25,320 --> 00:20:37,280 and keep those conversations about the future of AI going.