1 00:00:01,620 --> 00:00:05,500 Okay. So the first reason you might be convinced 2 00:00:05,580 --> 00:00:09,500 is coming in the next five years is that a lot of existing people in the space 3 00:00:09,580 --> 00:00:13,120 believe this. And I think some of them actually 4 00:00:13,160 --> 00:00:16,329 genuinely believe this. Like, this is not entirely just hype or 5 00:00:16,360 --> 00:00:19,580 marketing. 6 00:00:26,720 --> 00:00:30,360 In between 5 and 20 years from now, there's a good chance, a 7 00:00:30,500 --> 00:00:33,540 50% chance, we'll get AI smarter than us. 8 00:00:34,880 --> 00:00:38,760 Oh, yeah. Yes. In fact, I have reduced my expectations for 9 00:00:38,900 --> 00:00:42,060 when we will get to AGI or whatever, human 10 00:00:42,120 --> 00:00:45,960 level. Uh, I used to think it would be 5 to 20 11 00:00:45,980 --> 00:00:49,560 years, and now I think it could be, you know, 2 to 10 years. 12 00:00:51,200 --> 00:00:55,080 And so as a result of that, I, I... You can find it on my blog 13 00:00:55,300 --> 00:00:59,050 from about 2009, I think is the first time I publicly talked about 14 00:00:59,080 --> 00:01:02,020 I predicted a 50% chance of, um, AGI by 15 00:01:02,160 --> 00:01:03,560 2028. 16 00:01:03,620 --> 00:01:03,790 How did- 17 00:01:03,800 --> 00:01:05,920 I still, I still believe that today. 18 00:01:07,380 --> 00:01:10,970 But still, if you, if you extrapolate this kind of progress, 19 00:01:11,020 --> 00:01:12,220 very different place 20 00:01:13,140 --> 00:01:15,680 in 5 years, let alone 10 years. 21 00:01:15,690 --> 00:01:19,140 ... has been added. If you just kind of, like, eyeball the rate at which these 22 00:01:19,180 --> 00:01:23,160 capabilities are increasing, it does make you think that we'll 23 00:01:23,200 --> 00:01:25,640 get there by 2026 or 2027. 24 00:01:25,700 --> 00:01:29,420 In the essay, you actually say a really big 25 00:01:29,520 --> 00:01:33,410 thing, which is ASI, superintelligence, is 26 00:01:33,480 --> 00:01:36,960 actually thousands of days away. 27 00:01:37,020 --> 00:01:38,550 Maybe. I mean, that's our hope, our guess. 28 00:01:38,580 --> 00:01:39,680 Yeah. 29 00:01:41,960 --> 00:01:43,800 Like, I think, I think we're quite close to digital 30 00:01:43,860 --> 00:01:46,520 superintelligence. It, it may 31 00:01:46,580 --> 00:01:50,500 happen this year, and if it doesn't happen this year, next year 32 00:01:50,560 --> 00:01:51,660 for sure. 33 00:01:54,120 --> 00:01:57,230 The second reason you might be convinced, uh, 34 00:01:58,200 --> 00:02:00,660 is just by trying the models yourself. 35 00:02:00,680 --> 00:02:03,940 Like, instead of trusting what, you know, other people say about the various AI 36 00:02:03,980 --> 00:02:07,900 models, try the models yourself. And don't just try, you know, 37 00:02:07,940 --> 00:02:09,979 one AI model. Try at least two or three. 38 00:02:10,000 --> 00:02:13,850 Try the model, you know, 2019, try a model from 2022, try a 39 00:02:13,880 --> 00:02:17,260 model from 2025. And then once you have, like, at least two or three data points 40 00:02:17,380 --> 00:02:21,160 from, you know, three different time periods, 41 00:02:21,240 --> 00:02:24,440 then you can use that to, you know, extrapolate, you know, 42 00:02:24,480 --> 00:02:28,300 next. Uh, I think a common issue 43 00:02:28,580 --> 00:02:32,200 is that a lot of people who are using AI have only been using it since 44 00:02:32,360 --> 00:02:36,200 2023 or 2024. That simply does not give you enough data 45 00:02:36,420 --> 00:02:37,280 about the past. 46 00:02:38,380 --> 00:02:41,730 Like, AI has been a thing since at least 2019. 47 00:02:41,820 --> 00:02:44,230 Uh, transformers in particular have been a 48 00:02:44,440 --> 00:02:46,680 2019. 49 00:03:49,780 --> 00:03:53,600 The third reason you might be convinced is more of a speculative 50 00:03:53,660 --> 00:03:57,390 argument. There isn't a lot of actual data backing 51 00:03:57,420 --> 00:04:01,300 something I believe. Uh, so today's, uh, AI 52 00:04:01,320 --> 00:04:05,030 models can do tasks that can be done quickly. 53 00:04:05,060 --> 00:04:07,840 Like, if you can do a task that will take, you know, one minute or 10 minutes, you 54 00:04:07,880 --> 00:04:11,720 know, the AI can do it for you. If you want to do a task 55 00:04:11,740 --> 00:04:14,640 know, one hour of your time and lots of failed attempts, you know, 56 00:04:14,700 --> 00:04:18,660 probably not do it for you. Uh, however, the tasks 57 00:04:18,680 --> 00:04:21,480 which it can do, those it can do way faster than you. 58 00:04:21,540 --> 00:04:24,740 Like, if it will take, you know, 10 minutes, the AI will do it in, 59 00:04:24,820 --> 00:04:28,800 seconds. Uh, AIs operate, at least as 60 00:04:28,840 --> 00:04:32,420 of today, between 10 to, you know, 100 times the speed at which you're 61 00:04:32,560 --> 00:04:34,180 operating. 62 00:04:35,760 --> 00:04:39,640 Also, the time horizon of the tasks 63 00:04:39,720 --> 00:04:43,560 is increasing. So, you know, AI six months ago could do tasks 64 00:04:43,620 --> 00:04:44,000 that were 65 00:04:44,900 --> 00:04:48,140 maybe one minute long, and the AI today can do tasks that are, 66 00:04:48,160 --> 00:04:50,760 long. There is a trend that is increasing in that direction. 67 00:04:50,780 --> 00:04:54,650 The technical reason behind this is, uh, increasing 68 00:04:54,740 --> 00:04:58,220 compute being given to reinforcement learning on top of transformers, 69 00:04:58,760 --> 00:05:02,160 and that's something I'll get to later, but this is a specific technical reason 70 00:05:02,200 --> 00:05:06,060 this is happening. Uh, my view of intelligence, at 71 00:05:06,100 --> 00:05:06,320 least 72 00:05:07,140 --> 00:05:10,940 to some extent, is that you get better at a thing 73 00:05:11,000 --> 00:05:14,170 by doing it a lot of times. You know, you get better at giving speeches by 74 00:05:14,180 --> 00:05:17,000 lot of speeches. You get better at writing code by writing 75 00:05:17,020 --> 00:05:19,780 code. You get better 76 00:05:21,040 --> 00:05:24,280 at, you know, producing movies by producing lots of 77 00:05:24,340 --> 00:05:27,240 And if you can run faster, you can run more 78 00:05:27,320 --> 00:05:31,120 iterations. Uh, all scientific discovery, in my 79 00:05:31,140 --> 00:05:34,860 view, works like this. Uh, every lab 80 00:05:34,880 --> 00:05:38,580 experiment you run is yet another iteration 81 00:05:38,620 --> 00:05:42,460 system. You know, it could be microscopes 82 00:05:42,500 --> 00:05:46,120 biology. It could be, uh, you know, 83 00:05:46,140 --> 00:05:49,010 telescopes that help you study stars. 84 00:05:49,460 --> 00:05:53,400 It could be any field. And if you can run more iterations, 85 00:05:53,440 --> 00:05:57,080 you can make progress faster. Uh, there are some people who 86 00:05:57,120 --> 00:06:01,040 think that the amount of AI-driven scientific progress 87 00:06:01,100 --> 00:06:04,984 we can get is limited because...The physical world 88 00:06:05,003 --> 00:06:07,883 is slow. You know, labs are slow. Labs are expensive. 89 00:06:07,904 --> 00:06:10,664 Running experiments takes time, and this is 90 00:06:10,804 --> 00:06:14,623 true. However, A, there are domains where the 91 00:06:14,664 --> 00:06:16,183 labs are cheap. Uh, 92 00:06:19,003 --> 00:06:21,344 learning how to get good speeches is very cheap. 93 00:06:21,383 --> 00:06:24,104 Learning how to get good at math and software is cheap. 94 00:06:24,204 --> 00:06:27,244 Uh, like, it does not cost a lot of money, and I 95 00:06:27,284 --> 00:06:30,962 think that is partly why we have gotten a lot of good 96 00:06:31,003 --> 00:06:34,964 capabilities from reinforcement learning in math and software. 97 00:06:34,974 --> 00:06:38,484 Like, AI is extremely good at math and software in particular as of 98 00:06:38,584 --> 00:06:39,464 2025. 99 00:06:41,183 --> 00:06:44,484 I think this also carries over to other fields of scientific discovery. 100 00:06:46,024 --> 00:06:49,683 Maybe the speed up is not that much, but there will be a 101 00:06:49,724 --> 00:06:51,943 speed up. Uh, 102 00:06:52,804 --> 00:06:55,524 you can still get super intelligence this way. 103 00:06:55,703 --> 00:06:59,424 If, you know, humanity would otherwise have taken, 104 00:06:59,464 --> 00:07:03,344 know, invent a certain technology, if AI can do 105 00:07:03,384 --> 00:07:06,443 that in ten years, that is still super intelligence. 106 00:07:06,464 --> 00:07:10,323 Like, imagine if you were sitting in 1800 and you could get 107 00:07:10,404 --> 00:07:14,264 entire 1800 to 1900, you know, technological inventions to you in 108 00:07:14,304 --> 00:07:18,114 ten years. Or imagine you were in 1900, you know, 109 00:07:18,123 --> 00:07:21,964 before World War I, you know, people used sticks and stones to literally 110 00:07:22,024 --> 00:07:25,703 fight, and by the end of the century people used, 111 00:07:25,864 --> 00:07:28,164 and, you know, space satellites 112 00:07:28,484 --> 00:07:31,904 and, uh, bomber planes and so 113 00:07:32,024 --> 00:07:33,284 on. 114 00:07:35,304 --> 00:07:39,203 If you can speed that up, you can compress the amount of time 115 00:07:39,244 --> 00:07:40,024 requires to 116 00:07:40,864 --> 00:07:44,784 invent things, that gets you super intelligence, and 117 00:07:44,943 --> 00:07:46,104 third reason. 118 00:07:48,884 --> 00:07:52,683 The fourth reason I believe we might get super intelligence in the next 119 00:07:52,744 --> 00:07:56,244 five years is, uh, model scaling. This is a 120 00:07:56,284 --> 00:07:58,724 well-known trend among machine learning researchers. 121 00:07:58,744 --> 00:08:02,703 However, if you are not very in the space, you may not be aware of it. 122 00:08:02,804 --> 00:08:06,544 Uh, most of the AI progress in the last six 123 00:08:06,623 --> 00:08:10,404 2025 has been driven by one trend and one trend only, 124 00:08:10,464 --> 00:08:13,743 which is you take a transformer, you give it more data, more 125 00:08:13,803 --> 00:08:17,703 computation, and it will produce better results. 126 00:08:17,724 --> 00:08:19,943 People don't know how this happens, but this happens. 127 00:08:19,963 --> 00:08:23,063 This predictability repeatedly happens. 128 00:08:23,183 --> 00:08:26,044 Uh, the two ingredients are compute and data. 129 00:08:26,134 --> 00:08:30,073 Uh, data is just the entire internet's data 130 00:08:30,083 --> 00:08:33,644 the model. Compute is just more computers that people have 131 00:08:33,664 --> 00:08:36,764 purchased. Uh, there's a special type of computer used. 132 00:08:36,774 --> 00:08:39,223 They're called graphics processing units. 133 00:08:39,284 --> 00:08:42,683 They are good at this operation of matrix multiplication in 134 00:08:42,723 --> 00:08:45,884 particular. But basically it's just more computers. 135 00:08:45,904 --> 00:08:46,964 We have gone from 136 00:08:48,084 --> 00:08:51,084 spending, you know, like, 137 00:08:51,104 --> 00:08:54,704 $100,000, which could be literally just one machine 138 00:08:54,764 --> 00:08:57,743 person's room, you know, $100 billion 139 00:08:58,784 --> 00:09:02,684 worth of compute to train a single AI model that is, you know, an 140 00:09:02,764 --> 00:09:06,284 entire data center spanning, you know, thousands of 141 00:09:06,324 --> 00:09:10,284 acres. This scale up in, you know, 142 00:09:10,324 --> 00:09:14,153 six orders of magnitude has just been driven because every 143 00:09:14,243 --> 00:09:18,164 time we notice we use more compute and more data, we get more AI 144 00:09:18,184 --> 00:09:21,784 performance. We can now even predict this up to three 145 00:09:21,844 --> 00:09:25,243 decimal places. There is a law called the Chinchilla Scaling 146 00:09:25,304 --> 00:09:28,024 Law, uh, I'll put that on the screen right now, 147 00:09:29,364 --> 00:09:33,243 where N is the number of parameters, like how big the model 148 00:09:33,304 --> 00:09:36,504 is, and D is how much compute we give to it. 149 00:09:36,564 --> 00:09:37,223 Assume 150 00:09:38,153 --> 00:09:42,144 we s- show the model every token in the data 151 00:09:42,204 --> 00:09:45,844 exactly once, this L gives you the training loss. 152 00:09:45,904 --> 00:09:49,104 This is a proxy for how good the model 153 00:09:49,243 --> 00:09:53,144 performs. There is one important 154 00:09:53,184 --> 00:09:56,924 nitpick here that's really important, which is that the training 155 00:09:56,983 --> 00:10:00,424 loss correlates with your actual performance that you care about, 156 00:10:00,464 --> 00:10:03,344 but it is not the performance itself. 157 00:10:03,404 --> 00:10:07,304 It is possible for training loss to reduce and for capabilities to not increase 158 00:10:07,344 --> 00:10:10,824 that much. It is possible for training loss to reduce 159 00:10:10,864 --> 00:10:14,684 capabilities to increase a lot. Predicting this 160 00:10:14,743 --> 00:10:18,564 relationship is hard between training loss and capabilities, 161 00:10:18,624 --> 00:10:21,464 but there is a relationship. More training loss does mean more 162 00:10:21,483 --> 00:10:25,464 capabilities. Most AI researchers across the field have 163 00:10:25,524 --> 00:10:29,404 failed to predict exactly what training loss leads to exactly 164 00:10:29,444 --> 00:10:33,404 what capabilities. There are few people who have predicted it, 165 00:10:33,483 --> 00:10:37,243 and these people are, for example, Ilya Sutskever at OpenAI was very 166 00:10:37,284 --> 00:10:41,104 bullish on this, and he predicts, you know, we will get 167 00:10:41,144 --> 00:10:44,844 super intelligence in the next five to ten years. 168 00:10:44,884 --> 00:10:48,664 People who have actually believed in this law have generally predicted that AI will 169 00:10:48,723 --> 00:10:51,324 come sooner than what most other people in the field 170 00:10:51,483 --> 00:10:56,024 predict. 171 00:10:58,503 --> 00:11:02,404 Uh, we still have at least two orders of 172 00:11:02,444 --> 00:11:06,284 up left. Uh, we have spent, like I said, you know, $10 billion has 173 00:11:06,324 --> 00:11:10,124 already been spent and the next set of ra- data centers we're making are being 174 00:11:10,144 --> 00:11:13,304 made so that we spend $100 billion on the training of a single 175 00:11:13,384 --> 00:11:17,304 model. We might scale that up another order of magnitude, 176 00:11:17,364 --> 00:11:19,344 trillion dollars on a single model. 177 00:11:19,364 --> 00:11:23,223 It seems likely we will not go beyond a trillion dollars on a single model 178 00:11:23,944 --> 00:11:26,844 unless we get, you know, really exceptional capabilities. 179 00:11:26,884 --> 00:11:30,824 This alone means, you know, two more orders of magnitude from 10 180 00:11:30,944 --> 00:11:32,404 trillion, 181 00:11:33,304 --> 00:11:36,223 which you can compare to, you know, the six orders of magnitude they have 182 00:11:36,304 --> 00:11:39,884 gone through. There are some people who say, you know, 183 00:11:39,944 --> 00:11:43,723 model, which was GPT-4.5 was not that 184 00:11:43,884 --> 00:11:47,723 impressive, meaning, you know, this law might be weakening and that 185 00:11:47,764 --> 00:11:51,684 might be true, but just remember we still have two more 186 00:11:53,964 --> 00:11:57,584 The fifth reason you might be convinced super intelligence might be coming in the 187 00:11:57,644 --> 00:12:00,983 next five years is reinforcement learning scaling. 188 00:12:01,024 --> 00:12:03,864 Reinforcement learning is an old technique in machine learning. 189 00:12:03,904 --> 00:12:07,764 However, it has been applied to transformers in 190 00:12:07,804 --> 00:12:11,464 Literally only one year ago we figured out how 191 00:12:11,644 --> 00:12:15,144 to apply reinforcement learning on a transformer to get 192 00:12:15,204 --> 00:12:18,843 even more capabilities.Uh, reinforcement 193 00:12:18,904 --> 00:12:22,843 learning allows the model to try something, fail, try something, fail, 194 00:12:22,883 --> 00:12:25,584 try again, and actually gets the correct answer. 195 00:12:25,684 --> 00:12:29,564 Uh, this is different from the base model 196 00:12:29,584 --> 00:12:31,973 know, cached answer which it gives out. 197 00:12:31,983 --> 00:12:35,483 Reinforcement learning allows us to spend more compute on a single 198 00:12:35,623 --> 00:12:39,454 task, uh, whereas earlier it could just give the 199 00:12:39,524 --> 00:12:43,223 answer for the task. There are multiple, you 200 00:12:43,304 --> 00:12:47,103 know, tasks which people consider difficult 201 00:12:47,164 --> 00:12:49,684 would never crack, you know, as of 2023 and 202 00:12:49,723 --> 00:12:53,483 2024. Uh, researchers said that 203 00:12:53,623 --> 00:12:55,603 certain benchmarks would not be cracked. 204 00:12:55,644 --> 00:12:56,544 People said, you know, 205 00:12:57,524 --> 00:12:59,623 Humanities Last Exam would not be cracked. 206 00:12:59,644 --> 00:13:01,983 People said, you know, ARC AGI would not be cracked. 207 00:13:02,024 --> 00:13:05,164 People said, you know, mathematics would not be cracked. 208 00:13:05,243 --> 00:13:07,404 All three of these have been cracked. 209 00:13:07,483 --> 00:13:11,304 Uh, Humanities Last Exam consists of lot of 210 00:13:11,363 --> 00:13:15,003 answers by PhDs in various fields of both science 211 00:13:15,243 --> 00:13:18,024 and social science. This includes biology, chemistry, 212 00:13:18,034 --> 00:13:22,023 archeology, uh, journalism, law and so on. 213 00:13:22,123 --> 00:13:25,723 Uh, all these PhDs agreed that, you know, if these questions get 214 00:13:25,763 --> 00:13:29,723 cracked, that means the model is as good as a PhD in those fields, 215 00:13:29,743 --> 00:13:32,584 seeing AI has already solved, like, almost half of that data 216 00:13:32,684 --> 00:13:34,763 set. 217 00:13:35,924 --> 00:13:39,644 We have started seeing a loss to predict, you 218 00:13:39,723 --> 00:13:40,243 know, 219 00:13:41,123 --> 00:13:44,623 how much computation leads to how much performance, 220 00:13:45,123 --> 00:13:49,103 like we had with model scaling where we have a very clear empirical law which 221 00:13:49,123 --> 00:13:53,084 gives you up to three decimal places accuracy predictions. 222 00:13:53,204 --> 00:13:57,003 Uh, we do not have as accurate predictions for 223 00:13:57,084 --> 00:14:00,424 scaling of reinforcement learning, because it is relatively new. 224 00:14:00,444 --> 00:14:03,523 We have less data points, we have tried this for less longer. 225 00:14:03,544 --> 00:14:07,404 However, people are already starting to make some 226 00:14:07,444 --> 00:14:08,804 are starting to see that 227 00:14:09,784 --> 00:14:13,743 every time we increase the compute by, you know, 10x, we get a 228 00:14:13,863 --> 00:14:17,743 little bit of improvement. So if you spend, you know, $10 per task, 229 00:14:17,824 --> 00:14:21,343 spend $100, then you spend $1000 a task, you get little improvement, little 230 00:14:21,384 --> 00:14:23,103 improvement, little improvement. 231 00:14:23,144 --> 00:14:26,824 We have gone all the way, at least as per public data of OpenAI, 232 00:14:26,843 --> 00:14:30,684 the way up to $500,000 per task, and we also know 233 00:14:30,723 --> 00:14:34,404 that, uh, AI has been able to get, you know, the gold 234 00:14:34,483 --> 00:14:38,134 medal on the International Mathematics Olympiad, which is like top olympiad 235 00:14:38,464 --> 00:14:42,294 for mathematics worldwide. We still have many, 236 00:14:42,324 --> 00:14:45,373 many, many more orders of magnitude to go from, 237 00:14:45,384 --> 00:14:49,363 $100,000, and in theory what we can spend 238 00:14:49,404 --> 00:14:52,284 is, which is at least a billion dollars per 239 00:14:52,304 --> 00:14:55,674 Like, if you were really solving, you know, novel physics, 240 00:14:55,684 --> 00:14:59,564 know, novel chemistry, if you were, you know, trying to solve something like 241 00:14:59,584 --> 00:15:03,084 cancer or trying to, you know, solve math theorems that have, you know, 242 00:15:03,123 --> 00:15:06,824 unsolved for, you know, the past century, like, why would you not spend a billion 243 00:15:06,884 --> 00:15:10,843 dollars on those tasks? So we have many more orders 244 00:15:10,863 --> 00:15:14,424 of magnitude to go here. We have a lot more compute we 245 00:15:14,464 --> 00:15:18,324 this, and we don't know how far this law scales, 246 00:15:18,363 --> 00:15:21,824 find out very soon. 247 00:15:24,023 --> 00:15:27,684 The sixth and final reason you might be convinced superintelligence might be 248 00:15:27,704 --> 00:15:31,304 by 2030 is that reinforcement learning has worked 249 00:15:31,343 --> 00:15:34,603 before. Like I said, this is a not new technique. 250 00:15:34,644 --> 00:15:38,343 We are applying it to transformers newly because transformers can, you know, 251 00:15:38,363 --> 00:15:42,184 understand English and generalize across so many, 252 00:15:42,223 --> 00:15:46,023 tasks. Whereas earlier, before transformers, we had, you know, 253 00:15:46,103 --> 00:15:49,623 doing specific things. Even when we just had narrow models doing 254 00:15:49,723 --> 00:15:53,603 specific things, like we had just one model 255 00:15:53,664 --> 00:15:56,863 we knew that reinforcement learning actually works pretty well. 256 00:15:56,983 --> 00:16:00,704 Uh, we have, uh, an AI model that is, you know, the world 257 00:16:00,763 --> 00:16:02,964 champion in chess that knows nothing about chess. 258 00:16:02,983 --> 00:16:06,743 Like, no chess experts had to tell that model how to get good at chess 259 00:16:07,003 --> 00:16:10,623 the best at chess. Uh, this is obviously by DeepMind. 260 00:16:10,664 --> 00:16:14,363 This is, uh, AlphaZero. We 261 00:16:14,483 --> 00:16:18,424 also have DeepMind's model that beat the best 262 00:16:18,483 --> 00:16:21,523 Go player in the world. This happened, again, before transformers were even 263 00:16:21,584 --> 00:16:25,284 invented. Like, this is very old news, but we have gotten 264 00:16:25,363 --> 00:16:27,444 superhuman performance in narrow domains. 265 00:16:27,454 --> 00:16:31,164 You know, poker has been solved. Dota has been solved. 266 00:16:31,204 --> 00:16:35,103 StarCraft has been solved. People, AI researchers moved on from 267 00:16:35,163 --> 00:16:38,804 games because they found them too easy, they found it too easy to make an AI that 268 00:16:38,843 --> 00:16:41,164 would just completely obliterate all the human players in a 269 00:16:41,223 --> 00:16:45,064 game. And that is why we are now working with general 270 00:16:45,103 --> 00:16:48,804 domains. So we know point number A on narrow 271 00:16:48,824 --> 00:16:51,723 domains, reinforcement learning gets superhuman 272 00:16:51,743 --> 00:16:55,543 Point number B, transformers can work on general domains. 273 00:16:55,584 --> 00:16:59,464 Point number C, reinforcement learning plus transformers 274 00:16:59,483 --> 00:17:03,444 domains actually works. Point number D, this thing scales up. 275 00:17:03,464 --> 00:17:05,304 Point number E, we are going to scale it 276 00:17:05,424 --> 00:17:07,743 up.