1 00:00:00,386 --> 00:00:01,653 the people who burn Tokens the fastest 2 00:00:01,946 --> 00:00:02,932 why are they worth the most 3 00:00:03,786 --> 00:00:06,713 a strange ranking has been trending at big Silicon Valley firms lately 4 00:00:07,105 --> 00:00:08,212 it's not about who writes better code 5 00:00:08,465 --> 00:00:10,153 not about who ships products faster 6 00:00:10,305 --> 00:00:12,413 instead it compares something that sounds absurd 7 00:00:12,906 --> 00:00:15,352 who burns through AI Tokens the fastest 8 00:00:15,826 --> 00:00:16,533 the faster you burn them 9 00:00:16,905 --> 00:00:19,353 the more the engineer is seen as a master 10 00:00:19,832 --> 00:00:20,140 hey there 11 00:00:20,392 --> 00:00:20,860 hi everyone 12 00:00:21,152 --> 00:00:22,160 I'm Wang Lijie 13 00:00:22,473 --> 00:00:24,160 let me first explain what a Token is 14 00:00:24,392 --> 00:00:26,439 you can think of it as fuel for AI 15 00:00:26,672 --> 00:00:28,119 every time you put AI to work 16 00:00:28,192 --> 00:00:30,360 from reading the instructions to generating the result 17 00:00:30,473 --> 00:00:33,080 the compute it uses is measured in Tokens 18 00:00:33,352 --> 00:00:34,340 the more Tokens you burn 19 00:00:34,633 --> 00:00:35,992 the more work AI does 20 00:00:35,992 --> 00:00:37,559 and the more you pay 21 00:00:38,031 --> 00:00:38,857 a follower asked me 22 00:00:39,190 --> 00:00:40,617 isn't this totally counterintuitive 23 00:00:41,110 --> 00:00:43,357 don't companies chase lower costs and higher efficiency 24 00:00:43,430 --> 00:00:45,018 isn't burning fewer Tokens better 25 00:00:45,391 --> 00:00:46,338 how is it the other way around 26 00:00:46,631 --> 00:00:48,458 burning them faster somehow means you're better 27 00:00:49,071 --> 00:00:52,157 this question hits the core logic of the AI era 28 00:00:53,490 --> 00:00:54,476 picture a scene 29 00:00:54,929 --> 00:00:56,817 there are two excavators on a construction site 30 00:00:56,929 --> 00:00:59,097 one burns two hundred liters of diesel a day 31 00:00:59,170 --> 00:01:00,856 the other burns only twenty liters 32 00:01:01,130 --> 00:01:01,877 looking at fuel use alone 33 00:01:02,170 --> 00:01:04,017 you'd think the fuel-saving one is the better deal 34 00:01:04,289 --> 00:01:07,336 but if the one burning two hundred liters dug a thousand cubic meters of earth 35 00:01:07,449 --> 00:01:09,476 while the one burning twenty liters only dug fifty 36 00:01:09,970 --> 00:01:11,336 then the answer flips 37 00:01:11,529 --> 00:01:12,636 the diesel isn't waste 38 00:01:12,889 --> 00:01:14,336 it's the shadow of output 39 00:01:14,880 --> 00:01:16,767 Tokens are the diesel of AI 40 00:01:17,240 --> 00:01:19,127 burning a lot doesn't mean wasting 41 00:01:19,240 --> 00:01:23,967 it just means you can run the machine at full speed keeping it fully loaded 42 00:01:24,360 --> 00:01:25,187 and most people 43 00:01:25,481 --> 00:01:26,928 simply don't have that ability 44 00:01:27,240 --> 00:01:28,288 let's take a look 45 00:01:28,321 --> 00:01:31,368 from an ordinary programmer to a top AI engineer 46 00:01:31,440 --> 00:01:33,047 how many hurdles lie in between 47 00:01:33,889 --> 00:01:34,475 the first hurdle 48 00:01:34,769 --> 00:01:36,095 is called the basic level 49 00:01:36,288 --> 00:01:39,856 an ordinary engineer gets an AI coding tool and starts writing code 50 00:01:40,008 --> 00:01:42,035 this is the now-popular vibe coding 51 00:01:42,448 --> 00:01:43,975 you don't write it line by line yourself 52 00:01:44,049 --> 00:01:46,327 you tell AI in plain words what feature you want 53 00:01:46,329 --> 00:01:47,296 and let it do the work 54 00:01:47,856 --> 00:01:48,683 sounds great 55 00:01:49,056 --> 00:01:49,963 but in practice 56 00:01:50,576 --> 00:01:52,704 what AI writes breaks the moment you test it 57 00:01:53,017 --> 00:01:55,023 you throw the error back and let it fix it 58 00:01:55,177 --> 00:01:57,583 you test again after the fix and new bugs pop up 59 00:01:57,696 --> 00:01:58,736 after three or four rounds 60 00:01:58,737 --> 00:02:01,103 AI might even break what it had already fixed 61 00:02:01,177 --> 00:02:02,624 so you have to start over 62 00:02:02,776 --> 00:02:04,856 this do-it-and-redo-it pattern 63 00:02:04,856 --> 00:02:07,055 over a month on a subscription plan 64 00:02:07,056 --> 00:02:09,064 burns roughly a hundred dollars of Tokens 65 00:02:09,474 --> 00:02:10,340 a hundred dollars isn't much 66 00:02:10,673 --> 00:02:11,780 but what does it buy 67 00:02:12,233 --> 00:02:14,201 an extremely inefficient worker 68 00:02:14,354 --> 00:02:16,753 ninety percent of the time AI isn't advancing the project 69 00:02:16,754 --> 00:02:18,320 it's all fixing its own mistakes 70 00:02:18,689 --> 00:02:19,277 the second hurdle 71 00:02:19,570 --> 00:02:21,136 is called workflow optimization 72 00:02:21,409 --> 00:02:23,717 smart engineers quickly find the trick 73 00:02:24,170 --> 00:02:25,777 the problem isn't that AI is dumb 74 00:02:25,850 --> 00:02:28,576 it's that the way you assign tasks is too crude 75 00:02:28,770 --> 00:02:30,596 just saying build me a website 76 00:02:31,130 --> 00:02:32,928 AI does it based on its own understanding 77 00:02:32,929 --> 00:02:34,856 and most likely it's not what you wanted 78 00:02:34,890 --> 00:02:37,136 then comes endless revising 79 00:02:37,290 --> 00:02:40,289 but if you first spend an hour laying out the development process 80 00:02:40,290 --> 00:02:43,976 and clearly defining the roles and collaboration rules of multiple agents 81 00:02:44,010 --> 00:02:46,256 then split the website into five modules 82 00:02:46,290 --> 00:02:50,648 and write out each module's function inputs outputs and boundary conditions clearly 83 00:02:50,649 --> 00:02:51,808 and then hand it to AI 84 00:02:51,809 --> 00:02:54,576 the odds of getting it right the first time go way up 85 00:02:55,148 --> 00:02:56,636 this is called workflow design 86 00:02:56,949 --> 00:02:57,815 when the input is precise 87 00:02:57,909 --> 00:02:59,236 AI reworks less 88 00:02:59,469 --> 00:03:00,136 but note this 89 00:03:00,509 --> 00:03:02,188 even though each task needs less rework 90 00:03:02,189 --> 00:03:04,636 you actually push forward more projects in a day 91 00:03:04,789 --> 00:03:06,796 total Token use goes up instead of down 92 00:03:07,590 --> 00:03:09,917 at the same time you start using better models 93 00:03:09,990 --> 00:03:11,178 expensive models make fewer mistakes 94 00:03:11,471 --> 00:03:12,877 and get it right the first time more often 95 00:03:13,231 --> 00:03:13,898 all things considered 96 00:03:14,150 --> 00:03:14,817 you save time 97 00:03:15,110 --> 00:03:16,278 but spend more money 98 00:03:16,671 --> 00:03:17,669 at this stage 99 00:03:17,671 --> 00:03:20,078 it's about two to four hundred dollars a month 100 00:03:20,962 --> 00:03:22,409 there's a subtle shift here 101 00:03:22,642 --> 00:03:23,349 at the first level 102 00:03:23,682 --> 00:03:25,608 the engineer inputs a ton of instructions 103 00:03:25,642 --> 00:03:27,488 AI outputs a pile of half-finished work 104 00:03:27,522 --> 00:03:29,689 the human inputs even more instructions to correct it 105 00:03:29,721 --> 00:03:31,768 high input low useful output 106 00:03:31,922 --> 00:03:32,629 at the second level 107 00:03:32,962 --> 00:03:34,801 the engineer inputs less and less 108 00:03:34,802 --> 00:03:36,969 yet AI's output gets more and more precise 109 00:03:37,202 --> 00:03:39,409 a good workflow is like a precision mold 110 00:03:39,481 --> 00:03:40,689 you put the raw material in 111 00:03:40,721 --> 00:03:42,328 and out comes a standard part 112 00:03:42,522 --> 00:03:45,969 the engineer goes from a handcrafter to a mold designer 113 00:03:46,242 --> 00:03:49,608 a mold designer's value isn't in how many products they made by hand 114 00:03:49,642 --> 00:03:53,129 but in how many products the mold can produce automatically 115 00:03:53,960 --> 00:03:55,187 what really opens up the gap 116 00:03:55,441 --> 00:03:56,387 is the third hurdle 117 00:03:56,840 --> 00:03:57,727 high concurrency 118 00:03:58,240 --> 00:03:59,967 here's an analogy that makes it clear 119 00:04:00,240 --> 00:04:01,028 in the first two levels 120 00:04:01,321 --> 00:04:02,368 you're a boss 121 00:04:02,601 --> 00:04:04,208 with only one employee 122 00:04:04,481 --> 00:04:05,227 you hand it the work 123 00:04:05,680 --> 00:04:06,928 it finishes and you check 124 00:04:07,120 --> 00:04:08,967 then you assign the next one 125 00:04:08,1000 --> 00:04:10,847 you can only do one thing at a time 126 00:04:11,322 --> 00:04:12,848 the third level is different 127 00:04:13,001 --> 00:04:15,089 you're upgraded to the company's CEO 128 00:04:15,162 --> 00:04:16,601 you don't have one person under you 129 00:04:16,602 --> 00:04:17,628 but a whole team 130 00:04:18,081 --> 00:04:19,161 some write code 131 00:04:19,162 --> 00:04:20,281 some do design 132 00:04:20,281 --> 00:04:21,108 some write docs 133 00:04:21,361 --> 00:04:22,529 some do testing 134 00:04:22,762 --> 00:04:23,548 and even better 135 00:04:23,762 --> 00:04:25,408 these employees can clone themselves 136 00:04:25,481 --> 00:04:27,920 like Sun Wukong pulling out a hair and blowing on it 137 00:04:27,921 --> 00:04:29,168 conjuring fifty copies of himself 138 00:04:30,130 --> 00:04:31,976 take my actual workflow as an example 139 00:04:32,169 --> 00:04:34,416 making one video needs thirty scenes 140 00:04:34,570 --> 00:04:37,077 the traditional way is the main AI handles them one by one 141 00:04:37,450 --> 00:04:38,397 first generate the image 142 00:04:38,729 --> 00:04:39,888 then generate the audio 143 00:04:39,890 --> 00:04:40,676 then compose the video 144 00:04:41,289 --> 00:04:43,296 you finish one before starting the next 145 00:04:43,330 --> 00:04:46,376 processing thirty scenes in a queue takes several hours 146 00:04:46,570 --> 00:04:47,436 how do you do high concurrency 147 00:04:48,049 --> 00:04:50,769 design a role dedicated to handling a single scene 148 00:04:50,770 --> 00:04:54,217 let it call the image audio and video composition functions itself 149 00:04:54,409 --> 00:04:57,097 then split this role into thirty clones 150 00:04:57,250 --> 00:04:59,248 each takes one scene number 151 00:04:59,250 --> 00:05:00,337 and they all start at once 152 00:05:00,609 --> 00:05:01,676 what used to take hours 153 00:05:01,929 --> 00:05:03,416 gets done in minutes 154 00:05:04,353 --> 00:05:05,760 this is high concurrency 155 00:05:06,113 --> 00:05:10,880 right now one main AI can dispatch up to fifty clones working in parallel 156 00:05:11,272 --> 00:05:11,940 did you notice 157 00:05:12,512 --> 00:05:15,320 by this point it's no longer one person writing code 158 00:05:15,353 --> 00:05:17,800 it's one person managing an AI army 159 00:05:18,392 --> 00:05:19,880 but this still isn't the limit 160 00:05:20,141 --> 00:05:20,848 the fourth hurdle 161 00:05:21,141 --> 00:05:22,788 is called full-chain squeezing 162 00:05:23,062 --> 00:05:25,749 you're no longer a CEO running one product line 163 00:05:25,822 --> 00:05:28,468 but a chairman overseeing ten CEOs at once 164 00:05:28,822 --> 00:05:31,548 each CEO runs their own project with a team 165 00:05:31,581 --> 00:05:33,269 ten lines moving forward at once 166 00:05:33,382 --> 00:05:36,709 and each line has tons of clones collaborating with high concurrency 167 00:05:37,097 --> 00:05:38,984 the most extreme is overnight mode 168 00:05:39,337 --> 00:05:41,696 before bed you give AI a big complex task 169 00:05:41,698 --> 00:05:45,936 laying out the rules the judgment criteria and the exception handling plans clearly 170 00:05:45,938 --> 00:05:47,224 and let it run all night 171 00:05:47,578 --> 00:05:48,496 you wake up in the morning 172 00:05:48,498 --> 00:05:50,737 a project that would take a team a week 173 00:05:50,737 --> 00:05:51,984 is already done 174 00:05:52,578 --> 00:05:54,085 sounds too good to be true 175 00:05:54,458 --> 00:05:54,765 right 176 00:05:55,498 --> 00:05:57,145 but there's a fatal threshold here 177 00:05:57,498 --> 00:05:59,825 if one part of the workflow isn't designed well 178 00:05:59,857 --> 00:06:01,825 or one rule isn't spelled out 179 00:06:01,977 --> 00:06:05,984 at three in the morning AI hits a point it can't decide on and stops to wait for you 180 00:06:06,297 --> 00:06:09,337 you wake up the next morning to find it stopped working at three 181 00:06:09,337 --> 00:06:10,785 wasting the whole night for nothing 182 00:06:11,842 --> 00:06:13,330 I've been through this 183 00:06:13,482 --> 00:06:16,482 once I gave AI a complex video production task 184 00:06:16,482 --> 00:06:17,510 I laid out all the steps 185 00:06:17,683 --> 00:06:18,629 but missed just one 186 00:06:18,922 --> 00:06:21,609 what to do when a certain image fails to generate 187 00:06:21,683 --> 00:06:24,970 and at two in the morning AI generated a bad image 188 00:06:25,042 --> 00:06:26,962 not knowing whether to skip it or retry 189 00:06:26,962 --> 00:06:28,330 it just stopped there waiting for me 190 00:06:28,443 --> 00:06:29,770 I looked the next morning 191 00:06:29,802 --> 00:06:31,681 and only six of thirty scenes were done 192 00:06:31,683 --> 00:06:33,729 the rest of the time was spent waiting for one word from me 193 00:06:34,505 --> 00:06:38,592 to keep AI running nonstop on a complex task for fifteen or twenty hours 194 00:06:38,666 --> 00:06:41,592 without getting stuck or making directional mistakes 195 00:06:41,666 --> 00:06:43,553 deciding on its own when it should 196 00:06:43,585 --> 00:06:46,152 and following the rules when it shouldn't act alone 197 00:06:46,265 --> 00:06:49,152 this demands extreme precision in the workflow 198 00:06:49,505 --> 00:06:51,412 it's like designing a real company 199 00:06:51,825 --> 00:06:53,972 each role's responsibilities the reporting flow 200 00:06:54,265 --> 00:06:56,784 the decision authority the escalation mechanism 201 00:06:56,785 --> 00:07:01,272 all have to be written into the virtual company's charter manuals and even its constitution 202 00:07:01,426 --> 00:07:02,505 miss one rule 203 00:07:02,505 --> 00:07:05,073 and AI might crash right there in the middle of the night 204 00:07:05,606 --> 00:07:06,813 at this level 205 00:07:06,847 --> 00:07:10,253 the engineer's rhythm is completely different from a traditional programmer's 206 00:07:10,407 --> 00:07:11,974 he's not sitting there typing 207 00:07:12,087 --> 00:07:14,573 he's running several AI product lines at once 208 00:07:14,727 --> 00:07:16,214 his time is cut into fragments 209 00:07:16,407 --> 00:07:17,573 while AI is running 210 00:07:17,647 --> 00:07:19,474 he's thinking about how to arrange the next task 211 00:07:19,926 --> 00:07:21,405 whichever line's AI stops 212 00:07:21,407 --> 00:07:24,366 he immediately jumps over to give new instructions or clear the blocker 213 00:07:24,366 --> 00:07:25,414 and keep it running 214 00:07:25,686 --> 00:07:27,773 he barely does any hands-on execution himself 215 00:07:27,926 --> 00:07:29,674 all his energy goes into one thing 216 00:07:30,006 --> 00:07:31,253 keeping AI from stopping 217 00:07:32,242 --> 00:07:33,150 an engineer like this 218 00:07:33,523 --> 00:07:34,629 how much do they burn in a month 219 00:07:35,283 --> 00:07:36,410 take myself as an example 220 00:07:36,763 --> 00:07:38,730 I use the Claude Code subscription plan 221 00:07:38,882 --> 00:07:40,209 two hundred dollars a month 222 00:07:40,362 --> 00:07:41,230 the amount of Tokens it buys 223 00:07:41,482 --> 00:07:44,050 roughly equals four thousand dollars of pay-as-you-go usage 224 00:07:44,322 --> 00:07:47,949 I burn through two of these accounts a month just to fill my working hours 225 00:07:48,403 --> 00:07:49,562 one Claude Code 226 00:07:49,562 --> 00:07:50,269 one Codex 227 00:07:50,562 --> 00:07:51,610 dividing the work 228 00:07:51,922 --> 00:07:53,442 converted to pay-as-you-go 229 00:07:53,442 --> 00:07:56,209 my monthly AI compute runs about eight thousand dollars 230 00:07:56,830 --> 00:07:57,957 you might be curious 231 00:07:58,070 --> 00:07:59,457 the Claude Code Max subscription 232 00:07:59,471 --> 00:08:00,498 the 20X after it 233 00:08:01,070 --> 00:08:01,957 what does it mean 234 00:08:02,270 --> 00:08:02,777 let me tell you 235 00:08:03,111 --> 00:08:05,878 once the monthly quota runs out and you switch to pay-as-you-go 236 00:08:05,991 --> 00:08:09,517 the best models burn about ten to fifteen dollars an hour 237 00:08:09,630 --> 00:08:12,669 ten hours a day is a hundred to a hundred fifty dollars 238 00:08:12,671 --> 00:08:14,957 three to over four thousand dollars a month 239 00:08:15,231 --> 00:08:17,717 so the quota the two hundred dollar plan gives you 240 00:08:17,750 --> 00:08:20,358 is about equal to four thousand dollars of pay-as-you-go 241 00:08:20,510 --> 00:08:21,878 that's the twenty times 242 00:08:22,190 --> 00:08:24,277 this is already a decent level in Silicon Valley 243 00:08:24,471 --> 00:08:26,918 but the engineers who use AI best 244 00:08:26,950 --> 00:08:28,918 need two to four of these accounts 245 00:08:29,031 --> 00:08:32,438 adding up to eight to sixteen thousand dollars of AI compute a month 246 00:08:32,854 --> 00:08:34,654 now looking back at that ranking 247 00:08:34,654 --> 00:08:36,262 it's not strange at all 248 00:08:36,575 --> 00:08:37,442 burning Tokens fast 249 00:08:37,695 --> 00:08:38,681 means three things 250 00:08:39,254 --> 00:08:41,161 one your workflow is well designed 251 00:08:41,454 --> 00:08:42,561 so AI won't get stuck 252 00:08:43,175 --> 00:08:44,982 two you understand high concurrency 253 00:08:45,055 --> 00:08:47,321 and can orchestrate dozens of AI clones at once 254 00:08:48,014 --> 00:08:51,781 three you can keep AI running at full speed while you sleep 255 00:08:52,094 --> 00:08:54,441 these three things are really saying the same thing 256 00:08:54,935 --> 00:08:58,022 you can squeeze AI labor to the absolute limit 257 00:08:58,398 --> 00:08:58,985 notice something 258 00:08:59,398 --> 00:09:02,445 this is completely different from how engineers used to be judged 259 00:09:02,798 --> 00:09:04,445 engineers used to be measured 260 00:09:04,518 --> 00:09:05,985 by how much code they could write themselves 261 00:09:06,557 --> 00:09:09,304 now it's by how much AI labor they can orchestrate 262 00:09:09,678 --> 00:09:11,804 and to what extreme they can squeeze it 263 00:09:12,117 --> 00:09:13,024 what does this resemble 264 00:09:13,557 --> 00:09:15,125 it resembles management ability 265 00:09:15,437 --> 00:09:17,205 the most classic leap in a career 266 00:09:17,278 --> 00:09:19,625 is going from an individual contributor to a manager 267 00:09:19,998 --> 00:09:22,164 it's no longer about how much work you can do yourself 268 00:09:22,197 --> 00:09:25,125 but how big a team you can lead to deliver together 269 00:09:25,880 --> 00:09:28,159 now the same leap is happening again 270 00:09:28,160 --> 00:09:29,648 except what you manage isn't people 271 00:09:29,680 --> 00:09:30,487 it's AI 272 00:09:30,800 --> 00:09:32,148 there's not a single living person on the team 273 00:09:32,400 --> 00:09:33,648 they're all agents 274 00:09:33,841 --> 00:09:36,227 but the core abilities a manager needs haven't changed 275 00:09:36,680 --> 00:09:41,487 breaking down tasks setting processes drawing boundaries handling exceptions and allocating resources 276 00:09:41,680 --> 00:09:43,207 a top AI engineer 277 00:09:43,241 --> 00:09:45,847 is essentially a top AI manager 278 00:09:46,160 --> 00:09:47,568 so that ranking 279 00:09:47,601 --> 00:09:49,648 isn't about who can spend more 280 00:09:49,721 --> 00:09:52,007 but who's better at being AI's boss 281 00:09:52,754 --> 00:09:55,121 but this story has an even deeper meaning 282 00:09:55,353 --> 00:09:58,793 when an engineer uses eight thousand dollars of AI compute a month 283 00:09:58,794 --> 00:10:01,520 to produce what a ten-person team used to make in a month 284 00:10:01,634 --> 00:10:03,260 do the math on those ten people's cost 285 00:10:03,713 --> 00:10:06,760 a Silicon Valley engineer averages around two hundred thousand dollars a year 286 00:10:06,794 --> 00:10:09,560 plus benefits insurance and office space 287 00:10:09,674 --> 00:10:12,800 ten people easily cost over two hundred thousand dollars a month 288 00:10:12,993 --> 00:10:13,741 eight thousand dollars 289 00:10:14,034 --> 00:10:15,321 against two hundred thousand dollars 290 00:10:15,353 --> 00:10:17,392 this twenty-five to one leverage ratio 291 00:10:17,394 --> 00:10:19,920 is what the ranking is really about 292 00:10:20,193 --> 00:10:21,953 this is no longer a tech topic 293 00:10:21,953 --> 00:10:24,400 it's an economic issue an industry-structure issue 294 00:10:24,554 --> 00:10:26,680 and a signal about the future job market 295 00:10:27,473 --> 00:10:30,040 whoever can orchestrate AI labor most efficiently 296 00:10:30,113 --> 00:10:33,440 can create the most output at the lowest cost 297 00:10:34,017 --> 00:10:35,905 we used to call this automation 298 00:10:36,098 --> 00:10:39,244 but old automation replaced repetitive physical labor 299 00:10:39,617 --> 00:10:42,185 assembly line workers tollbooth cashiers 300 00:10:42,417 --> 00:10:43,704 today it's different 301 00:10:43,738 --> 00:10:45,645 automation is starting to replace cognitive labor 302 00:10:46,058 --> 00:10:49,324 writing code doing design writing copy doing research 303 00:10:49,698 --> 00:10:52,217 these brain jobs once seen as iron rice bowls 304 00:10:52,218 --> 00:10:53,584 are being taken over by AI 305 00:10:53,738 --> 00:10:56,645 ever wonder why this wave of automation hits so hard 306 00:10:57,218 --> 00:11:01,425 the last generation of automation needed factories robots and production lines 307 00:11:01,537 --> 00:11:02,857 the upfront investment was huge 308 00:11:02,858 --> 00:11:04,464 only big companies could afford it 309 00:11:04,617 --> 00:11:05,324 this generation is different 310 00:11:05,698 --> 00:11:08,896 one person one laptop one subscription account 311 00:11:08,897 --> 00:11:11,344 is one or even several production lines 312 00:11:12,137 --> 00:11:13,704 there's no minimum investment threshold 313 00:11:13,777 --> 00:11:15,505 no economies-of-scale limit 314 00:11:15,578 --> 00:11:17,045 the spread isn't linear 315 00:11:17,218 --> 00:11:18,584 it's exponential 316 00:11:19,214 --> 00:11:21,941 the key to driving this cognitive automation machine 317 00:11:22,053 --> 00:11:23,653 isn't the ability to write code 318 00:11:23,653 --> 00:11:28,461 it's the ability to design systems define processes and manage teams of agents 319 00:11:28,653 --> 00:11:30,100 if you're a programmer 320 00:11:30,214 --> 00:11:33,500 the skill to practice now isn't learning a new programming language 321 00:11:33,573 --> 00:11:36,661 it's learning how to be a good CEO of an AI team 322 00:11:36,854 --> 00:11:38,281 if you're not a programmer 323 00:11:38,734 --> 00:11:40,220 you should pay even more attention to this 324 00:11:40,454 --> 00:11:43,140 this trend won't stop at programming 325 00:11:43,254 --> 00:11:45,340 any job that requires cognitive ability 326 00:11:45,454 --> 00:11:47,201 will face the same question in the future 327 00:11:47,893 --> 00:11:49,720 who can orchestrate AI more efficiently 328 00:11:50,173 --> 00:11:51,581 is worth more 329 00:11:51,813 --> 00:11:53,201 and those who can't orchestrate AI 330 00:11:53,734 --> 00:11:56,140 are like someone holding a shovel next to an excavator 331 00:11:56,933 --> 00:11:57,520 you can dig too 332 00:11:58,134 --> 00:11:59,620 but you and that machine 333 00:11:59,734 --> 00:12:01,380 aren't even in the same dimension 334 00:12:01,840 --> 00:12:02,728 what do you think 335 00:12:02,760 --> 00:12:05,947 what's the next industry to compete on Token-burning speed 336 00:12:06,441 --> 00:12:08,427 and what level are you at right now 337 00:12:08,1000 --> 00:12:10,088 let me know in the comments 338 00:12:10,281 --> 00:12:10,988 I'm Wang Lijie 339 00:12:11,281 --> 00:12:12,195 see you next time