1 00:00:00,000 --> 00:00:01,920 All right, get ready. 2 00:00:01,920 --> 00:00:05,120 Because today, we're tackling some serious math. 3 00:00:05,120 --> 00:00:06,640 Oh yeah, it's gonna be a fun one. 4 00:00:06,640 --> 00:00:09,040 You wanted to know if AI can solve 5 00:00:09,040 --> 00:00:10,720 really hard math problems. 6 00:00:10,720 --> 00:00:11,760 Yeah. 7 00:00:11,760 --> 00:00:13,080 And to answer that, 8 00:00:13,080 --> 00:00:15,320 we're doing a deep dive into frontier math. 9 00:00:15,320 --> 00:00:16,160 Okay. 10 00:00:16,160 --> 00:00:20,240 A brand new benchmark designed to test AI's limits. 11 00:00:20,240 --> 00:00:23,920 So like, how good is AI really at doing math? 12 00:00:23,920 --> 00:00:26,800 Yeah, think of it like the SATs for AI, 13 00:00:26,800 --> 00:00:27,880 but way, way harder. 14 00:00:27,880 --> 00:00:29,400 Like way harder. 15 00:00:29,400 --> 00:00:30,720 So what's really interesting here is that 16 00:00:30,720 --> 00:00:32,800 we're not just talking about like your typical 17 00:00:32,800 --> 00:00:34,000 high school algebra. 18 00:00:34,000 --> 00:00:36,200 No, no, no, this is the big leagues. 19 00:00:36,200 --> 00:00:38,400 Frontier math is all about pushing AI 20 00:00:38,400 --> 00:00:40,680 beyond its comfort zone with problems 21 00:00:40,680 --> 00:00:43,400 that even experienced mathematicians find challenging. 22 00:00:43,400 --> 00:00:46,840 Yeah, stuff that even makes like seasoned math pros 23 00:00:46,840 --> 00:00:47,840 sweat a little. 24 00:00:47,840 --> 00:00:49,840 So it's kind of like giving a calculator 25 00:00:49,840 --> 00:00:51,400 a PhD level exam. 26 00:00:51,400 --> 00:00:52,520 That's a great way to put it. 27 00:00:52,520 --> 00:00:53,440 Is that a good analogy? 28 00:00:53,440 --> 00:00:54,280 It really is. 29 00:00:54,280 --> 00:00:55,120 Okay. 30 00:00:55,120 --> 00:00:57,320 Because the thing is, current AI benchmarks, 31 00:00:57,320 --> 00:00:58,720 they're getting kind of easy. 32 00:00:58,720 --> 00:00:59,360 Okay. 33 00:00:59,360 --> 00:01:03,560 AI is hitting near perfect scores on tests like GSMAK 34 00:01:03,560 --> 00:01:05,920 or math, which means we need a new challenge, right? 35 00:01:05,920 --> 00:01:07,320 We need to push the limits. 36 00:01:07,320 --> 00:01:08,160 Yeah, yeah. 37 00:01:08,160 --> 00:01:10,560 So like AI's aced all the beginner levels 38 00:01:10,560 --> 00:01:12,320 and now we need the boss fight. 39 00:01:12,320 --> 00:01:13,440 Right, the boss fight, yeah. 40 00:01:13,440 --> 00:01:14,280 Exactly. 41 00:01:14,280 --> 00:01:15,120 Okay, that makes sense. 42 00:01:15,120 --> 00:01:17,280 And plus there's this whole thing about contamination. 43 00:01:17,280 --> 00:01:18,120 Contamination. 44 00:01:18,120 --> 00:01:21,520 Yeah, it means like AI might have seen similar problems 45 00:01:21,520 --> 00:01:23,560 when it was being trained, right? 46 00:01:23,560 --> 00:01:26,560 So it's not really solving them from scratch. 47 00:01:27,520 --> 00:01:28,440 Frontier math, 48 00:01:28,440 --> 00:01:31,280 they try to avoid this by using all new problems. 49 00:01:31,280 --> 00:01:32,120 Yeah. 50 00:01:32,120 --> 00:01:33,840 Stuff that AI has never encountered before. 51 00:01:33,840 --> 00:01:36,440 So it's like making sure the AI isn't just like 52 00:01:36,440 --> 00:01:38,200 spitting out memorized answers. 53 00:01:38,200 --> 00:01:39,720 You got it. 54 00:01:39,720 --> 00:01:42,440 It's got to think deeply and creatively. 55 00:01:42,440 --> 00:01:44,680 Just like a human mathematician would. 56 00:01:44,680 --> 00:01:46,880 Okay, so how did they come up with these 57 00:01:46,880 --> 00:01:48,720 brain melting math problems? 58 00:01:48,720 --> 00:01:49,560 Yeah. 59 00:01:49,560 --> 00:01:52,080 Did they like walk a bunch of mathematicians 60 00:01:52,080 --> 00:01:54,120 in a room with no pizza until they came up 61 00:01:54,120 --> 00:01:56,360 with like truly evil problems? 62 00:01:56,360 --> 00:01:58,280 It wasn't quite that dramatic. 63 00:01:58,280 --> 00:02:01,840 But they did assemble this team of over 60 mathematicians. 64 00:02:01,840 --> 00:02:04,560 And this ranged from like grad students 65 00:02:04,560 --> 00:02:07,200 all the way up to some of the biggest names in math, 66 00:02:07,200 --> 00:02:08,040 you know? 67 00:02:08,040 --> 00:02:10,200 People who've won something called the Fields Medal. 68 00:02:10,200 --> 00:02:12,920 The Fields Medal, is that like the Nobel Prize of math? 69 00:02:12,920 --> 00:02:14,400 You could say that, yeah. 70 00:02:14,400 --> 00:02:16,960 It's one of the highest honors a mathematician can receive. 71 00:02:16,960 --> 00:02:19,920 So they've got like the best of the best. 72 00:02:19,920 --> 00:02:20,760 Absolutely. 73 00:02:20,760 --> 00:02:21,600 Coming up with these problems. 74 00:02:21,600 --> 00:02:23,120 And having that range of expertise, 75 00:02:23,120 --> 00:02:24,160 that was important, right? 76 00:02:24,160 --> 00:02:27,080 Because it made sure that frontier math covers 77 00:02:27,080 --> 00:02:29,680 like a huge spectrum of math. 78 00:02:29,680 --> 00:02:31,280 We're talking number theory, 79 00:02:31,280 --> 00:02:35,560 combinatorics, algebraic geometry, you name it. 80 00:02:35,560 --> 00:02:40,560 Wow, okay, so we've got these insanely hard problems 81 00:02:41,080 --> 00:02:44,400 created by like the smartest mathematicians in the world. 82 00:02:44,400 --> 00:02:46,120 But can you give us a taste? 83 00:02:46,120 --> 00:02:48,160 Like what do these problems actually look like? 84 00:02:48,160 --> 00:02:51,280 Absolutely, let's look at some examples from the paper. 85 00:02:51,280 --> 00:02:54,760 One problem, which is rated as high difficulty, 86 00:02:54,760 --> 00:02:57,800 involves figuring out how often a certain pattern 87 00:02:57,800 --> 00:02:59,440 shows up with prime numbers. 88 00:02:59,440 --> 00:03:00,280 Oh, okay. 89 00:03:00,280 --> 00:03:02,080 It sounds simple, but it actually connects 90 00:03:02,080 --> 00:03:04,800 to some of the biggest unsolved mysteries in math. 91 00:03:04,800 --> 00:03:07,440 Like something called Arton's conjecture. 92 00:03:07,440 --> 00:03:10,360 Wait, so even humans haven't figured these problems out? 93 00:03:10,360 --> 00:03:11,200 That's right. 94 00:03:11,200 --> 00:03:13,840 And that's part of what makes frontier math so cool, right? 95 00:03:13,840 --> 00:03:14,680 Yeah. 96 00:03:14,680 --> 00:03:16,440 It's not just about testing AI. 97 00:03:16,440 --> 00:03:18,960 It's about pushing the boundaries of like 98 00:03:18,960 --> 00:03:20,560 what we know as humans. 99 00:03:20,560 --> 00:03:21,520 That's pretty awesome. 100 00:03:21,520 --> 00:03:22,360 It is. 101 00:03:22,360 --> 00:03:24,000 Okay, so that's next level. 102 00:03:24,000 --> 00:03:25,400 What about another example? 103 00:03:25,400 --> 00:03:26,920 All right, another one. 104 00:03:26,920 --> 00:03:29,520 This is rated high medium difficulty. 105 00:03:29,520 --> 00:03:30,360 Okay. 106 00:03:30,360 --> 00:03:32,960 It asks AI to design a specific type 107 00:03:32,960 --> 00:03:34,640 of polynomial equation. 108 00:03:34,640 --> 00:03:35,480 Okay. 109 00:03:35,480 --> 00:03:38,720 And the solution, you have to visualize these like 110 00:03:38,720 --> 00:03:43,720 twisting and looping paths in a really abstract way. 111 00:03:43,800 --> 00:03:44,640 Whoa. 112 00:03:44,640 --> 00:03:46,680 It's not the kind of math you'd find in a textbook. 113 00:03:46,680 --> 00:03:48,920 No, so it's not just about crunching numbers. 114 00:03:48,920 --> 00:03:49,760 No. 115 00:03:49,760 --> 00:03:51,840 It's about like real creative thinking. 116 00:03:51,840 --> 00:03:52,840 Exactly. 117 00:03:52,840 --> 00:03:53,840 And there's more. 118 00:03:53,840 --> 00:03:54,680 Okay. 119 00:03:54,680 --> 00:03:57,200 Another problem uses something called piatic numbers. 120 00:03:57,200 --> 00:03:58,720 Piatic numbers, huh? 121 00:03:58,720 --> 00:04:01,000 Which is a totally different way of thinking about numbers 122 00:04:01,000 --> 00:04:01,840 than what we're used to. 123 00:04:01,840 --> 00:04:04,080 It's like, uh, whole new world of math. 124 00:04:04,080 --> 00:04:06,400 Okay, my brain is starting to hurt a little bit. 125 00:04:06,400 --> 00:04:07,880 But this is cool. 126 00:04:07,880 --> 00:04:11,560 So the big question is, how did AI actually do on these 127 00:04:11,560 --> 00:04:12,760 frontier math challenges? 128 00:04:12,760 --> 00:04:16,280 Well, let's just say it wasn't exactly a stellar performance. 129 00:04:16,280 --> 00:04:18,600 Oh no, did they completely crash and burn? 130 00:04:18,600 --> 00:04:21,360 None of the AI models they tested could solve more than 131 00:04:21,360 --> 00:04:23,320 2% of the problems. 132 00:04:23,320 --> 00:04:24,320 Wow. 133 00:04:24,320 --> 00:04:25,600 It was a real wake-up call. 134 00:04:25,600 --> 00:04:28,560 It shows that AI still has a long way to go before it can 135 00:04:28,560 --> 00:04:32,440 match, you know, real human mathematicians at this kind 136 00:04:32,440 --> 00:04:33,720 of hardcore math. 137 00:04:33,720 --> 00:04:35,240 Yeah, wow, that's a reality check. 138 00:04:35,240 --> 00:04:38,520 Yeah, it wasn't quite the AI takeover some people 139 00:04:38,520 --> 00:04:39,640 were maybe predicting. 140 00:04:39,640 --> 00:04:41,120 So a total wipeout. 141 00:04:41,120 --> 00:04:43,680 But I'm curious, which AI models were even brave enough 142 00:04:43,680 --> 00:04:45,560 to step into the ring on this one? 143 00:04:45,560 --> 00:04:49,240 Oh, they tested some of the heavy hitters like GPT-4, 144 00:04:49,240 --> 00:04:53,040 Claude, couple different versions of open AI's, 145 00:04:53,040 --> 00:04:58,040 01 model, Grock 2, even Google DeepMind's Gemini. 146 00:04:58,560 --> 00:05:00,480 Whoa, so the all-star team of AI? 147 00:05:00,480 --> 00:05:04,200 Yeah, these are like the most advanced AI systems we've got. 148 00:05:04,200 --> 00:05:06,400 And they still got like completely stumped. 149 00:05:06,400 --> 00:05:09,120 Pretty much, yeah, frontier math was a whole other ball game. 150 00:05:09,120 --> 00:05:13,680 Okay, so no AI came close to solving like a decent chunk 151 00:05:13,680 --> 00:05:14,520 of the problems. 152 00:05:14,520 --> 00:05:17,800 But I'm curious, were there any like interesting patterns 153 00:05:17,800 --> 00:05:18,880 in how they approached it? 154 00:05:18,880 --> 00:05:21,840 Yeah, actually there were some interesting differences. 155 00:05:21,840 --> 00:05:25,320 Some AI like 01 Preview and Gemini, 156 00:05:25,320 --> 00:05:28,160 they kind of just jumped right into answering 157 00:05:28,160 --> 00:05:31,480 without a whole lot of like exploration or experimenting. 158 00:05:31,480 --> 00:05:33,480 Yeah, they were like skimming the textbook 159 00:05:33,480 --> 00:05:35,880 and thinking they were ready for the exam 160 00:05:35,880 --> 00:05:37,760 without really understanding the material. 161 00:05:37,760 --> 00:05:39,760 Yeah, they seem to rely a lot 162 00:05:39,760 --> 00:05:43,720 on pattern recognition, quick calculations, 163 00:05:43,720 --> 00:05:47,920 but maybe not so much on that deep conceptual understanding. 164 00:05:47,920 --> 00:05:49,080 Okay, that makes sense. 165 00:05:49,080 --> 00:05:50,020 What about the others? 166 00:05:50,020 --> 00:05:52,200 Did any of them take a different approach? 167 00:05:52,200 --> 00:05:54,320 Yeah, what about the other AI models? 168 00:05:54,320 --> 00:05:57,840 Some models like Grock 2 seem to be more methodical. 169 00:05:57,840 --> 00:06:00,080 They ran more code experiments. 170 00:06:00,080 --> 00:06:00,920 Interesting. 171 00:06:00,920 --> 00:06:03,920 Tried different approaches like reflecting on the results 172 00:06:03,920 --> 00:06:05,480 before they gave a final answer. 173 00:06:05,480 --> 00:06:07,960 It's more like the student who's carefully working 174 00:06:07,960 --> 00:06:09,960 through the practice problems. 175 00:06:09,960 --> 00:06:10,800 You got it. 176 00:06:10,800 --> 00:06:11,880 Double checking their work. 177 00:06:11,880 --> 00:06:12,720 Exactly. 178 00:06:12,720 --> 00:06:15,720 Okay, so that's reflected in how much they actually used 179 00:06:15,720 --> 00:06:17,120 like computationally. 180 00:06:17,120 --> 00:06:17,960 It is, yeah. 181 00:06:17,960 --> 00:06:21,960 Interesting, like is one approach better than the other? 182 00:06:21,960 --> 00:06:22,920 That's a great question. 183 00:06:22,920 --> 00:06:26,560 And honestly, the research doesn't say for sure. 184 00:06:26,560 --> 00:06:29,200 It just points out this difference in strategy. 185 00:06:29,200 --> 00:06:32,280 Like this is something to pay attention to. 186 00:06:32,280 --> 00:06:33,120 Yeah. 187 00:06:33,120 --> 00:06:34,560 It's something to keep researching. 188 00:06:34,560 --> 00:06:36,760 It makes me wonder though, 189 00:06:36,760 --> 00:06:40,560 is it like how we're training these AIs, 190 00:06:40,560 --> 00:06:43,760 are we accidentally rewarding speed 191 00:06:43,760 --> 00:06:45,400 over that deeper thinking? 192 00:06:45,400 --> 00:06:46,360 That's a really good point. 193 00:06:46,360 --> 00:06:48,160 Like is that built into how we're making them? 194 00:06:48,160 --> 00:06:50,440 It leads to an even bigger question, right? 195 00:06:50,440 --> 00:06:53,720 Like how do we evaluate AI in the first place? 196 00:06:53,720 --> 00:06:54,680 Oh, interesting. 197 00:06:54,680 --> 00:06:57,440 Because as these systems get more complex, 198 00:06:57,440 --> 00:07:00,300 maybe just looking at did they get the right answer, 199 00:07:00,300 --> 00:07:01,680 that's not enough anymore. 200 00:07:01,680 --> 00:07:02,520 Right, right. 201 00:07:02,520 --> 00:07:05,040 We gotta look at the process, the reasoning, the strategy. 202 00:07:05,040 --> 00:07:07,800 So it's not just like did they get an A on the test? 203 00:07:07,800 --> 00:07:09,320 It's like how are they learning? 204 00:07:09,320 --> 00:07:10,160 How are they thinking? 205 00:07:10,160 --> 00:07:10,980 Exactly. 206 00:07:10,980 --> 00:07:13,280 Are they learning to think like really deeply 207 00:07:13,280 --> 00:07:15,500 and creatively like a mathematician? 208 00:07:15,500 --> 00:07:16,880 Or are they just getting really good 209 00:07:16,880 --> 00:07:20,240 at finding shortcuts and mimicking patterns 210 00:07:20,240 --> 00:07:22,200 without truly understanding? 211 00:07:22,200 --> 00:07:23,560 Right, without the deeper knowledge. 212 00:07:23,560 --> 00:07:25,600 And those are big questions, right? 213 00:07:25,600 --> 00:07:29,400 And frontier math, it really forces us to think about that. 214 00:07:29,400 --> 00:07:31,400 So it's not just a test for AI. 215 00:07:31,400 --> 00:07:32,240 No. 216 00:07:32,240 --> 00:07:33,040 It's a test for us, 217 00:07:33,040 --> 00:07:34,840 for how we're even approaching this whole thing. 218 00:07:34,840 --> 00:07:35,680 I'd say so, yeah, 219 00:07:35,680 --> 00:07:37,760 it's a challenge for the whole AI field. 220 00:07:37,760 --> 00:07:40,240 You know, we've been talking about the AI struggles, 221 00:07:40,240 --> 00:07:42,080 but I'm kinda curious, 222 00:07:42,080 --> 00:07:45,280 did the mathematicians who designed these problems, 223 00:07:45,280 --> 00:07:46,800 did they find them hard too? 224 00:07:46,800 --> 00:07:48,160 Oh, absolutely. 225 00:07:48,160 --> 00:07:51,760 They interviewed several fields medalists. 226 00:07:51,760 --> 00:07:54,440 Oh, remember, those are the top mathematicians in the world. 227 00:07:54,440 --> 00:07:55,480 Yeah, the best of the best. 228 00:07:55,480 --> 00:07:56,560 And even a top coach 229 00:07:56,560 --> 00:07:59,240 for the International Mathematical Olympiad. 230 00:07:59,240 --> 00:08:00,480 Wow, okay, so this is- 231 00:08:00,480 --> 00:08:02,480 Like the Olympics of math competitions 232 00:08:02,480 --> 00:08:03,880 for high school students. 233 00:08:03,880 --> 00:08:04,720 Next level. 234 00:08:04,720 --> 00:08:05,540 Yeah. 235 00:08:05,540 --> 00:08:06,880 What did they think about this? 236 00:08:06,880 --> 00:08:10,680 They all agreed the problems were really hard, 237 00:08:10,680 --> 00:08:11,840 even for them. 238 00:08:11,840 --> 00:08:13,280 Okay, good, I feel a little better 239 00:08:13,280 --> 00:08:14,720 about my own math skills now. 240 00:08:14,720 --> 00:08:16,960 Huh, that's reassuring. 241 00:08:16,960 --> 00:08:18,360 It wasn't just the concepts, 242 00:08:18,360 --> 00:08:21,240 it was the creative thinking, the problem solving. 243 00:08:21,240 --> 00:08:23,120 Like that was the tricky part. 244 00:08:23,120 --> 00:08:27,480 They did point out that the way frontier math is set up, 245 00:08:27,480 --> 00:08:29,680 you know, giving numerical answers 246 00:08:29,680 --> 00:08:32,280 instead of formal proofs, 247 00:08:32,280 --> 00:08:35,480 that's not really typical for math research. 248 00:08:35,480 --> 00:08:38,000 But it still needs that deep understanding, 249 00:08:38,000 --> 00:08:40,920 that ability to connect different ideas. 250 00:08:40,920 --> 00:08:42,280 So it's still real math, 251 00:08:42,280 --> 00:08:43,880 it's just a different way of showing it. 252 00:08:43,880 --> 00:08:45,160 Yeah, exactly. 253 00:08:45,160 --> 00:08:47,560 So then they asked these mathematicians, 254 00:08:47,560 --> 00:08:51,960 when do you think AI will be able to crack frontier math? 255 00:08:51,960 --> 00:08:52,800 Okay. 256 00:08:52,800 --> 00:08:54,720 And their estimates were, 257 00:08:54,720 --> 00:08:56,800 well, let's just say they weren't holding their breath. 258 00:08:56,800 --> 00:08:59,200 It's a years, decades. 259 00:08:59,200 --> 00:09:00,960 Yeah, that seemed to be the consensus. 260 00:09:00,960 --> 00:09:02,400 Wow, that's a long time. 261 00:09:02,400 --> 00:09:04,960 It is, but here's where things get interesting. 262 00:09:04,960 --> 00:09:05,800 Okay. 263 00:09:05,800 --> 00:09:07,160 They were much more optimistic 264 00:09:07,160 --> 00:09:09,040 about human AI collaboration. 265 00:09:09,040 --> 00:09:11,560 Oh, tell me more about that, what did they picture? 266 00:09:11,560 --> 00:09:15,000 They were thinking, you know, AI could be a really powerful 267 00:09:15,000 --> 00:09:16,520 research assistant. 268 00:09:16,520 --> 00:09:19,200 Helping mathematicians explore ideas, 269 00:09:19,200 --> 00:09:21,800 test things out, verify calculations. 270 00:09:21,800 --> 00:09:23,400 But like a souped up calculator. 271 00:09:23,400 --> 00:09:24,240 Yeah. 272 00:09:24,240 --> 00:09:26,080 But that can also like brainstorm with you. 273 00:09:26,080 --> 00:09:26,920 Exactly. 274 00:09:26,920 --> 00:09:29,400 And they thought that kind of team humans 275 00:09:29,400 --> 00:09:31,560 and AI working together, 276 00:09:31,560 --> 00:09:35,560 they could make progress on frontier math way sooner 277 00:09:35,560 --> 00:09:36,400 Oh, interesting. 278 00:09:36,400 --> 00:09:38,520 Than AI could on its own. 279 00:09:38,520 --> 00:09:39,960 So we don't have to wait for AI 280 00:09:39,960 --> 00:09:41,920 to become some kind of math genius. 281 00:09:41,920 --> 00:09:42,760 Right. 282 00:09:42,760 --> 00:09:43,760 Before it can actually be useful, 283 00:09:43,760 --> 00:09:45,240 we can start working with it now. 284 00:09:45,240 --> 00:09:46,120 Exactly. 285 00:09:46,120 --> 00:09:46,960 That's cool. 286 00:09:46,960 --> 00:09:49,440 It is, but for this to actually work, 287 00:09:49,440 --> 00:09:51,720 we need the right tools, the right way for them 288 00:09:51,720 --> 00:09:52,920 to talk to each other. 289 00:09:52,920 --> 00:09:53,760 Yeah. 290 00:09:53,760 --> 00:09:56,360 AI needs to understand human instructions. 291 00:09:56,360 --> 00:09:59,160 And humans need to be able to guide the AI's thinking. 292 00:09:59,160 --> 00:10:01,160 Like learning to speak each other's language. 293 00:10:01,160 --> 00:10:02,720 I like that, yeah, building a bridge 294 00:10:02,720 --> 00:10:03,880 between those two minds. 295 00:10:03,880 --> 00:10:04,720 Yeah. 296 00:10:04,720 --> 00:10:05,560 And that's a challenge in itself. 297 00:10:05,560 --> 00:10:06,400 It is. 298 00:10:06,400 --> 00:10:08,080 But I like the idea of like teamwork. 299 00:10:08,080 --> 00:10:09,120 Yeah. 300 00:10:09,120 --> 00:10:10,200 Like powerful stuff. 301 00:10:10,200 --> 00:10:12,840 So I'm sold on this team idea. 302 00:10:12,840 --> 00:10:13,680 Yeah. 303 00:10:13,680 --> 00:10:16,000 But what are some of the like practical hurdles 304 00:10:16,000 --> 00:10:18,280 we got it clear to actually make this happen? 305 00:10:18,280 --> 00:10:20,520 Well, for one thing, AI models right now, 306 00:10:20,520 --> 00:10:22,160 they're still pretty limited. 307 00:10:22,160 --> 00:10:23,000 Okay. 308 00:10:23,000 --> 00:10:25,040 When it comes to this really hardcore math, 309 00:10:25,040 --> 00:10:26,640 even with human guidance. 310 00:10:26,640 --> 00:10:27,480 Right. 311 00:10:27,480 --> 00:10:29,520 It might be tough for them to make a lot of progress. 312 00:10:29,520 --> 00:10:31,960 So it's like having a brilliant, 313 00:10:31,960 --> 00:10:33,720 but inexperienced assistant. 314 00:10:33,720 --> 00:10:34,560 Exactly. 315 00:10:34,560 --> 00:10:36,920 Tons of potential, but still needs a lot of training. 316 00:10:36,920 --> 00:10:39,080 And that training, like we talked about before, 317 00:10:39,080 --> 00:10:41,800 it's hard because there's not enough data. 318 00:10:41,800 --> 00:10:42,640 Right. 319 00:10:42,640 --> 00:10:44,400 In these really specific areas of math. 320 00:10:44,400 --> 00:10:46,480 So we need more data, is that what you're saying? 321 00:10:46,480 --> 00:10:47,320 We do. 322 00:10:47,320 --> 00:10:51,480 And we also need better ways to create synthetic data. 323 00:10:51,480 --> 00:10:52,320 Okay. 324 00:10:52,320 --> 00:10:53,920 That actually captures how complex 325 00:10:53,920 --> 00:10:55,520 these math problems are. 326 00:10:55,520 --> 00:10:56,360 Okay. 327 00:10:56,360 --> 00:10:57,280 So data is one thing, 328 00:10:57,280 --> 00:10:59,120 but what about the computing power, right? 329 00:10:59,120 --> 00:10:59,960 Yeah. 330 00:10:59,960 --> 00:11:00,800 We mentioned earlier, 331 00:11:00,800 --> 00:11:04,920 it could take days on Google servers to solve one problem. 332 00:11:04,920 --> 00:11:07,400 It's not really practical if you're trying to do research. 333 00:11:07,400 --> 00:11:08,240 You're right. 334 00:11:08,240 --> 00:11:10,000 We need more efficient algorithms, 335 00:11:10,000 --> 00:11:11,880 better hardware that can handle 336 00:11:11,880 --> 00:11:13,960 these really intense calculations 337 00:11:13,960 --> 00:11:16,200 without breaking the bank. 338 00:11:16,200 --> 00:11:17,040 Right. 339 00:11:17,040 --> 00:11:17,880 Without taking forever. 340 00:11:17,880 --> 00:11:18,720 Exactly. 341 00:11:18,720 --> 00:11:19,960 So it's a lot of different challenges all at the same time. 342 00:11:19,960 --> 00:11:21,680 It is, but the good news is 343 00:11:21,680 --> 00:11:24,400 there's a ton of research happening in all these areas. 344 00:11:24,400 --> 00:11:25,240 Okay. That's good. 345 00:11:25,240 --> 00:11:26,080 There's some hope then. 346 00:11:26,080 --> 00:11:26,920 Yeah, I think so. 347 00:11:26,920 --> 00:11:28,680 It seems like frontier math has been 348 00:11:28,680 --> 00:11:31,960 like a shakeup for the MCU world. 349 00:11:31,960 --> 00:11:32,800 It has. 350 00:11:32,800 --> 00:11:36,600 It's forcing everyone to rethink how they're doing things. 351 00:11:36,600 --> 00:11:39,040 It's a wake-up call and a call to action. 352 00:11:39,040 --> 00:11:41,800 It reminds us that true AI intelligence, 353 00:11:41,800 --> 00:11:43,520 we're not there yet, 354 00:11:43,520 --> 00:11:46,000 but it shows us the way forward. 355 00:11:46,000 --> 00:11:49,000 How to get to that future where humans and AI 356 00:11:49,000 --> 00:11:53,320 are working together to solve these amazing math problems. 357 00:11:53,320 --> 00:11:55,360 Cool. This has been a really interesting journey. 358 00:11:55,360 --> 00:11:57,520 We started with a pretty simple question. 359 00:11:57,520 --> 00:11:59,520 Can AI solve really hard math? 360 00:11:59,520 --> 00:12:00,360 Yeah. 361 00:12:00,360 --> 00:12:04,400 And we've ended up exploring so much more. 362 00:12:04,400 --> 00:12:05,240 We have. 363 00:12:05,240 --> 00:12:06,760 We have a lot of sense of AI, human AI teams. 364 00:12:06,760 --> 00:12:08,600 What even is AI intelligence? 365 00:12:08,600 --> 00:12:09,440 Yeah. 366 00:12:09,440 --> 00:12:10,280 It's been a lot. 367 00:12:10,280 --> 00:12:12,360 And we need more data, more computing power. 368 00:12:12,360 --> 00:12:13,360 It's a big challenge. 369 00:12:13,360 --> 00:12:14,800 Yeah, it really is. 370 00:12:14,800 --> 00:12:18,000 But what's amazing to me is the sense of possibility. 371 00:12:18,000 --> 00:12:18,840 Yeah. 372 00:12:18,840 --> 00:12:19,680 AI is still so, 373 00:12:19,680 --> 00:12:22,400 we're just starting to understand what it can do. 374 00:12:22,400 --> 00:12:23,240 Right. 375 00:12:23,240 --> 00:12:25,280 Who knows what crazy discoveries are waiting for us 376 00:12:25,280 --> 00:12:27,160 as AI gets smarter. 377 00:12:27,160 --> 00:12:31,240 Maybe AI will not only solve the problems in frontier math, 378 00:12:31,240 --> 00:12:34,200 but also help us discover new math concepts 379 00:12:34,200 --> 00:12:35,560 that we haven't even thought of yet. 380 00:12:35,560 --> 00:12:36,880 I wouldn't bet against it. 381 00:12:36,880 --> 00:12:37,720 That would be amazing. 382 00:12:37,720 --> 00:12:40,200 It would be a sign that AI isn't just copying us. 383 00:12:40,200 --> 00:12:41,040 Right. 384 00:12:41,040 --> 00:12:43,280 It's coming up with its own understanding of math. 385 00:12:43,280 --> 00:12:46,520 Okay, well, sadly, that's all the time we have 386 00:12:46,520 --> 00:12:47,720 for today's deep dive. 387 00:12:47,720 --> 00:12:48,800 Oh! 388 00:12:48,800 --> 00:12:49,840 Thanks for coming along with us 389 00:12:49,840 --> 00:12:51,920 on this journey into AI and math. 390 00:12:51,920 --> 00:12:53,000 It's been fun. 391 00:12:53,000 --> 00:12:54,080 It really has. 392 00:12:54,080 --> 00:12:56,520 Until next time, keep those minds curious 393 00:12:56,520 --> 00:12:58,680 and keep on diving deep. 394 00:12:58,680 --> 00:13:01,440 Welcome back for the final part of our deep dive 395 00:13:01,440 --> 00:13:02,880 into frontier math. 396 00:13:02,880 --> 00:13:04,080 It's been a wild ride. 397 00:13:04,080 --> 00:13:04,920 It has. 398 00:13:04,920 --> 00:13:07,800 We started off realizing that AI isn't quite ready 399 00:13:07,800 --> 00:13:09,440 to take over the math world, 400 00:13:09,440 --> 00:13:12,480 but then we saw the potential of these human AI teams 401 00:13:12,480 --> 00:13:14,240 pushing the limits of what we know. 402 00:13:14,240 --> 00:13:15,400 And we also talked about, 403 00:13:15,400 --> 00:13:17,480 what even is AI intelligence? 404 00:13:17,480 --> 00:13:18,320 Yeah. 405 00:13:18,320 --> 00:13:20,040 It's not just about getting the right answer anymore. 406 00:13:20,040 --> 00:13:22,600 No, it's about the process, the reasoning, 407 00:13:22,600 --> 00:13:24,480 the creativity behind it all. 408 00:13:24,480 --> 00:13:26,360 So as we wrap things up, I'm curious. 409 00:13:26,360 --> 00:13:27,200 Yeah. 410 00:13:27,200 --> 00:13:29,520 What do you think are the bigger implications of all this? 411 00:13:29,520 --> 00:13:30,640 It's a good question. 412 00:13:30,640 --> 00:13:34,160 What does frontier math tell us about AI as a whole 413 00:13:34,160 --> 00:13:35,720 and where it's going? 414 00:13:35,720 --> 00:13:37,680 Well, I think it's both a reality check 415 00:13:37,680 --> 00:13:39,360 and a roadmap, you know what I mean? 416 00:13:39,360 --> 00:13:41,800 Like on the one hand, it shows us the limits 417 00:13:41,800 --> 00:13:43,600 of what AI can do right now, 418 00:13:43,600 --> 00:13:45,680 especially with that abstract reasoning, 419 00:13:45,680 --> 00:13:48,880 creative problem solving, which is important for math, 420 00:13:48,880 --> 00:13:50,880 but also for so many other fields too. 421 00:13:50,880 --> 00:13:53,000 So it's telling us that we haven't reached 422 00:13:53,000 --> 00:13:56,760 that like true artificial general intelligence yet. 423 00:13:56,760 --> 00:13:58,760 Right, that AI that can think like a human 424 00:13:58,760 --> 00:14:00,240 across all these different areas. 425 00:14:00,240 --> 00:14:01,320 Okay, that makes sense. 426 00:14:01,320 --> 00:14:04,520 But on the other hand, frontier math gives us clues 427 00:14:04,520 --> 00:14:06,760 about where AI needs to get better. 428 00:14:06,760 --> 00:14:08,000 Okay, I like that. 429 00:14:08,000 --> 00:14:10,800 By seeing how AI struggled with these problems, 430 00:14:10,800 --> 00:14:13,440 researchers can figure out the weak spots, right? 431 00:14:13,440 --> 00:14:15,800 And focus on making new algorithms, 432 00:14:15,800 --> 00:14:17,800 new ways of training these systems. 433 00:14:17,800 --> 00:14:22,400 So it's like a diagnostic tool for the whole field of AI. 434 00:14:22,400 --> 00:14:25,400 Okay, and this isn't just about math problems then, right? 435 00:14:25,400 --> 00:14:26,480 No, it's bigger than that. 436 00:14:26,480 --> 00:14:28,120 Because the skills we're talking about, 437 00:14:28,120 --> 00:14:31,760 abstract reasoning, logic, creativity, 438 00:14:31,760 --> 00:14:33,560 those are important for everything. 439 00:14:33,560 --> 00:14:36,160 Absolutely, if AI can master that stuff, 440 00:14:36,160 --> 00:14:38,120 it can change so many things. 441 00:14:38,120 --> 00:14:42,320 So the lessons from frontier math could ripple out 442 00:14:42,320 --> 00:14:44,440 and affect all kinds of AI research. 443 00:14:44,440 --> 00:14:45,400 I think so, yeah. 444 00:14:45,400 --> 00:14:48,960 From science and engineering to maybe even art. 445 00:14:48,960 --> 00:14:52,160 It's pushing the boundaries of what AI can do. 446 00:14:52,160 --> 00:14:54,160 And it's making everyone aim higher. 447 00:14:54,160 --> 00:14:55,000 I love that. 448 00:14:55,000 --> 00:14:57,440 It's not just about making AI good at math, 449 00:14:57,440 --> 00:15:00,920 it's about using math to create a whole new kind of AI. 450 00:15:00,920 --> 00:15:01,760 That's so cool. 451 00:15:01,760 --> 00:15:04,960 One that's more powerful, more flexible, more like us. 452 00:15:04,960 --> 00:15:07,200 Well, this deep dive has been mind blowing. 453 00:15:07,200 --> 00:15:08,080 I've enjoyed it. 454 00:15:08,080 --> 00:15:10,000 We went from a simple question, 455 00:15:10,000 --> 00:15:12,520 can AI solve hard math? 456 00:15:12,520 --> 00:15:13,360 Yeah. 457 00:15:13,360 --> 00:15:15,000 To talking about the future of AI. 458 00:15:15,000 --> 00:15:16,280 It's all connected. 459 00:15:16,280 --> 00:15:17,120 It really is. 460 00:15:17,120 --> 00:15:19,040 It's been awesome sharing this journey with you 461 00:15:19,040 --> 00:15:20,080 and with everyone listening. 462 00:15:20,080 --> 00:15:20,920 It has. 463 00:15:20,920 --> 00:15:22,600 I hope it's made everyone a little more curious 464 00:15:22,600 --> 00:15:23,440 about this stuff. 465 00:15:23,440 --> 00:15:24,280 Me too. 466 00:15:24,280 --> 00:15:26,040 And I think that's a perfect note to end on. 467 00:15:26,040 --> 00:15:26,880 Yeah. 468 00:15:26,880 --> 00:15:29,840 Thanks for joining us on this exploration of AI and math. 469 00:15:29,840 --> 00:15:30,720 It's been a pleasure. 470 00:15:30,720 --> 00:15:31,680 It really has. 471 00:15:31,680 --> 00:15:34,720 Until next time, keep those minds questioning, 472 00:15:34,720 --> 00:15:37,600 keep exploring, and keep diving deep 473 00:15:37,600 --> 00:15:39,600 into this amazing world of knowledge. 474 00:15:39,600 --> 00:15:58,600 Couldn't have said it better myself.