1 00:00:05,239 --> 00:00:07,259 [music] 2 00:00:09,825 --> 00:00:11,845 [music] 3 00:00:15,839 --> 00:00:18,160 Hello, Pyon. 4 00:00:18,160 --> 00:00:20,960 Welcome. Welcome to the first breakout 5 00:00:20,960 --> 00:00:23,920 session in Ballroom 3. Uh if that's not 6 00:00:23,920 --> 00:00:25,359 where you thought you were, you might 7 00:00:25,359 --> 00:00:28,080 have to change. Um but here in ballroom 8 00:00:28,080 --> 00:00:32,320 3 we are going to hear from Beno and 9 00:00:32,320 --> 00:00:35,120 Daisy about a product about 10 00:00:35,120 --> 00:00:38,480 productivity. Please welcome Daisy and 11 00:00:38,480 --> 00:00:41,265 Benor. 12 00:00:41,265 --> 00:00:43,285 [applause] 13 00:00:45,850 --> 00:00:47,360 [sighs] 14 00:00:47,360 --> 00:00:49,360 All righty. 15 00:00:49,360 --> 00:00:51,600 How we doing? 16 00:00:51,600 --> 00:00:54,640 Oh. Oh. Oh [ __ ] Sorry. Um 17 00:00:54,640 --> 00:00:57,199 I did take it. I did take it. 18 00:00:57,199 --> 00:01:00,719 All righty. Um, so, um, 19 00:01:00,719 --> 00:01:01,760 I'm just going to apologize a little bit 20 00:01:01,760 --> 00:01:03,359 in advance my voice. I'm a bit croaky, 21 00:01:03,359 --> 00:01:06,000 but moving on. So, um, so there there's 22 00:01:06,000 --> 00:01:08,240 two of us here, right? Two podiums, and 23 00:01:08,240 --> 00:01:09,520 this is, you know, an important topic 24 00:01:09,520 --> 00:01:10,640 that we're doing. This is a debate, 25 00:01:10,640 --> 00:01:10,960 right? 26 00:01:10,960 --> 00:01:12,560 Well, not really. This is a prepared 27 00:01:12,560 --> 00:01:13,760 talk we've written together. So, you're 28 00:01:13,760 --> 00:01:15,360 not going to see us taking consistent 29 00:01:15,360 --> 00:01:16,799 opposing positions from each other, but 30 00:01:16,799 --> 00:01:18,479 you'll see us teeing each other a lot 31 00:01:18,479 --> 00:01:20,799 up, teeing each other up a lot. Uh, so 32 00:01:20,799 --> 00:01:22,479 the format of this is less like a debate 33 00:01:22,479 --> 00:01:24,240 and more like a Socratic dialogue played 34 00:01:24,240 --> 00:01:25,520 out live on stage. 35 00:01:25,520 --> 00:01:27,360 So Plato, 36 00:01:27,360 --> 00:01:29,040 well, more in the style of transgender 37 00:01:29,040 --> 00:01:30,880 fies blogging about information security 38 00:01:30,880 --> 00:01:32,880 who are in turn doing it in the style of 39 00:01:32,880 --> 00:01:34,079 Plato. 40 00:01:34,079 --> 00:01:35,920 I don't think I have the lived 41 00:01:35,920 --> 00:01:38,251 experience to make that comparison. 42 00:01:38,251 --> 00:01:38,640 [laughter] 43 00:01:38,640 --> 00:01:40,560 Your cheering has broken our auto quue. 44 00:01:40,560 --> 00:01:42,640 I couldn't possibly comment. 45 00:01:42,640 --> 00:01:45,200 Well, moving on. So, um, this is a talk 46 00:01:45,200 --> 00:01:47,439 about productivity by which we mean what 47 00:01:47,439 --> 00:01:49,680 do we mean by that? the efficiency of 48 00:01:49,680 --> 00:01:51,119 production of goods or services 49 00:01:51,119 --> 00:01:52,399 expressed by some measure. 50 00:01:52,399 --> 00:01:55,119 Thank you, Wikipedia. Um, but really 51 00:01:55,119 --> 00:01:56,399 we're talking about how you can achieve 52 00:01:56,399 --> 00:01:58,159 gains in it or even better how to 53 00:01:58,159 --> 00:02:00,159 achieve acceptable sustainable levels of 54 00:02:00,159 --> 00:02:00,560 it. 55 00:02:00,560 --> 00:02:02,479 So, there's been a bunch of attempts to 56 00:02:02,479 --> 00:02:04,079 improve productivity in the tech sector 57 00:02:04,079 --> 00:02:05,280 over time and it's interesting to look 58 00:02:05,280 --> 00:02:07,040 at how they succeeded or didn't. Hey, 59 00:02:07,040 --> 00:02:11,239 Beno, you're [snorts] old, right? 60 00:02:11,760 --> 00:02:13,599 I couldn't possibly comment. 61 00:02:13,599 --> 00:02:15,440 I [laughter] I just want to be clear, he 62 00:02:15,440 --> 00:02:18,160 wrote that line for me. 63 00:02:18,160 --> 00:02:20,000 Moving on. [laughter] 64 00:02:20,000 --> 00:02:22,640 So Daisy, what do you know about agile? 65 00:02:22,640 --> 00:02:24,160 That's why you never like plan anything, 66 00:02:24,160 --> 00:02:25,840 right? It's cards stuck to walls and 67 00:02:25,840 --> 00:02:27,040 sprints. 68 00:02:27,040 --> 00:02:28,879 Well, they're more artifacts of various 69 00:02:28,879 --> 00:02:30,879 systems that grew out of the idea. I 70 00:02:30,879 --> 00:02:32,800 mean, I am old enough to remember the 71 00:02:32,800 --> 00:02:35,120 stuff that agile was a reaction against. 72 00:02:35,120 --> 00:02:36,480 So, you know, those who understand what 73 00:02:36,480 --> 00:02:38,800 a waterfall model is, it expects that 74 00:02:38,800 --> 00:02:40,480 you'd construct a rigorous set of 75 00:02:40,480 --> 00:02:42,160 requirements. And only once you'd done 76 00:02:42,160 --> 00:02:43,760 that could you move on to working on 77 00:02:43,760 --> 00:02:45,599 your design. And only once you'd done 78 00:02:45,599 --> 00:02:48,000 that could you actually write code. U if 79 00:02:48,000 --> 00:02:49,120 you came across anything that would 80 00:02:49,120 --> 00:02:50,400 require a change in design or 81 00:02:50,400 --> 00:02:52,480 requirements then you'd have to stop go 82 00:02:52,480 --> 00:02:54,239 back fix the requirements andor the 83 00:02:54,239 --> 00:02:55,760 design and get back to coding again. 84 00:02:55,760 --> 00:02:58,800 I'm picturing so many meetings 85 00:02:58,800 --> 00:03:01,360 pretty much. Um the thing is this 86 00:03:01,360 --> 00:03:03,280 approach largely mirrors other forms of 87 00:03:03,280 --> 00:03:04,800 engineering and ignores the main 88 00:03:04,800 --> 00:03:06,480 difference between them and writing 89 00:03:06,480 --> 00:03:08,720 software which is that software isn't a 90 00:03:08,720 --> 00:03:10,720 physical thing. it can change far more 91 00:03:10,720 --> 00:03:13,200 easily once it exists unlike pretty much 92 00:03:13,200 --> 00:03:14,640 you know any piece of physical 93 00:03:14,640 --> 00:03:17,280 engineering. So a bunch of approaches 94 00:03:17,280 --> 00:03:19,280 started being developed and implemented 95 00:03:19,280 --> 00:03:21,120 based on this understanding and a bunch 96 00:03:21,120 --> 00:03:22,319 of the people working on this got 97 00:03:22,319 --> 00:03:24,720 together and put out the agile manifesto 98 00:03:24,720 --> 00:03:27,200 which reads thus we are uncovering 99 00:03:27,200 --> 00:03:28,959 better ways of developing software by 100 00:03:28,959 --> 00:03:30,720 doing it and helping others do it. 101 00:03:30,720 --> 00:03:32,159 Through this work, we have come to value 102 00:03:32,159 --> 00:03:34,480 into value individuals interactions over 103 00:03:34,480 --> 00:03:36,799 processes and tools, working software 104 00:03:36,799 --> 00:03:39,120 over comprehensive documentation, 105 00:03:39,120 --> 00:03:40,959 customer collaboration over contract 106 00:03:40,959 --> 00:03:43,360 negotiation, and responding to change 107 00:03:43,360 --> 00:03:45,760 over following a plan. That is, while 108 00:03:45,760 --> 00:03:47,120 there is value in the items on the 109 00:03:47,120 --> 00:03:48,720 right, we value the items on the left 110 00:03:48,720 --> 00:03:50,319 more. Did someone over here boo the 111 00:03:50,319 --> 00:03:53,840 agile manifesto? Did I hear that? 112 00:03:53,840 --> 00:03:56,000 Sorry, should have pissed. [laughter] 113 00:03:56,000 --> 00:03:57,840 My actual line is, but that says nothing 114 00:03:57,840 --> 00:04:00,239 about sticking cards to walls. Well, 115 00:04:00,239 --> 00:04:01,920 yeah, because the sticking cards to 116 00:04:01,920 --> 00:04:04,879 walls folks, whoops. Um, [laughter] 117 00:04:04,879 --> 00:04:06,879 were were part of this but weren't all 118 00:04:06,879 --> 00:04:09,120 of it. In fact, the whole first point is 119 00:04:09,120 --> 00:04:11,439 about not being a slave to process. I 120 00:04:11,439 --> 00:04:12,879 feel like it's more about emphasizing 121 00:04:12,879 --> 00:04:14,560 communication and adaptability over 122 00:04:14,560 --> 00:04:15,280 anything else. 123 00:04:15,280 --> 00:04:16,880 And effective communication is only 124 00:04:16,880 --> 00:04:18,239 possible between equals. 125 00:04:18,239 --> 00:04:19,919 Exactly. This means that the people 126 00:04:19,919 --> 00:04:21,600 delivering the product not just the 127 00:04:21,600 --> 00:04:23,600 coders but the designers the testers all 128 00:04:23,600 --> 00:04:24,960 the workers on the project should have 129 00:04:24,960 --> 00:04:26,960 an equal voice over the way the product 130 00:04:26,960 --> 00:04:28,720 is delivered to the people who are 131 00:04:28,720 --> 00:04:29,840 commissioning the delivery of the 132 00:04:29,840 --> 00:04:31,919 product. Uh the overall deal was that 133 00:04:31,919 --> 00:04:33,280 management was allowed to change what 134 00:04:33,280 --> 00:04:35,280 was needed whenever they wanted that the 135 00:04:35,280 --> 00:04:36,320 workers should be able to tell 136 00:04:36,320 --> 00:04:37,840 management that based on their knowledge 137 00:04:37,840 --> 00:04:39,759 and understanding of what was there that 138 00:04:39,759 --> 00:04:41,360 the delivery schedule or the feature set 139 00:04:41,360 --> 00:04:43,520 needed to change as a result. Resolving 140 00:04:43,520 --> 00:04:45,360 that might require some discussion, but 141 00:04:45,360 --> 00:04:47,040 it was a discussion between equals, not 142 00:04:47,040 --> 00:04:48,880 necessarily one side issuing dictats to 143 00:04:48,880 --> 00:04:49,919 the other. 144 00:04:49,919 --> 00:04:52,160 Sounds great. So, why does this not 145 00:04:52,160 --> 00:04:54,400 sound like literally any form of agile 146 00:04:54,400 --> 00:04:56,080 I've ever experienced? 147 00:04:56,080 --> 00:04:57,680 If only there was some foreshadowing for 148 00:04:57,680 --> 00:05:00,160 this point. Well, [laughter] 149 00:05:00,160 --> 00:05:02,560 this is best summed up by the halfass 150 00:05:02,560 --> 00:05:04,960 agile manifesto, which is a website you 151 00:05:04,960 --> 00:05:08,080 can find. I love it. And it reads, "We 152 00:05:08,080 --> 00:05:09,680 have heard about new ways of developing 153 00:05:09,680 --> 00:05:11,039 software by paying consultants and 154 00:05:11,039 --> 00:05:13,680 reading gardener reports. [laughter] 155 00:05:13,680 --> 00:05:15,680 Through this, we have been told to value 156 00:05:15,680 --> 00:05:17,199 individuals and interactions over 157 00:05:17,199 --> 00:05:18,800 processes and tools. And we have 158 00:05:18,800 --> 00:05:20,479 mandatory processes and tools to control 159 00:05:20,479 --> 00:05:22,400 how these individuals, we prefer the 160 00:05:22,400 --> 00:05:24,800 term resources, interact. [laughter] 161 00:05:24,800 --> 00:05:27,120 Uh working software over comprehensive 162 00:05:27,120 --> 00:05:28,880 documentation as long as that software 163 00:05:28,880 --> 00:05:31,280 is comprehensively documented. uh 164 00:05:31,280 --> 00:05:32,880 customer collaboration over contract 165 00:05:32,880 --> 00:05:34,560 negotiation within the boundaries of 166 00:05:34,560 --> 00:05:35,919 strict contracts of course and subject 167 00:05:35,919 --> 00:05:37,919 to rigorous change control and 168 00:05:37,919 --> 00:05:39,759 responding to change over following a 169 00:05:39,759 --> 00:05:41,840 plan provided a detailed plan is in 170 00:05:41,840 --> 00:05:43,280 place to respond to the change and it is 171 00:05:43,280 --> 00:05:45,759 followed precisely that is while the 172 00:05:45,759 --> 00:05:47,360 items on the left sound nice in theory 173 00:05:47,360 --> 00:05:48,800 we're an enterprise company and there's 174 00:05:48,800 --> 00:05:50,400 no way we're letting go of the items on 175 00:05:50,400 --> 00:05:51,280 the right 176 00:05:51,280 --> 00:05:53,520 right yeah that sounds more like the 177 00:05:53,520 --> 00:05:55,199 agile I'm familiar with 178 00:05:55,199 --> 00:05:56,960 yeah uh faced with a movement that 179 00:05:56,960 --> 00:05:58,720 emphasized worker participation in 180 00:05:58,720 --> 00:06:00,560 defining how software was delivered 181 00:06:00,560 --> 00:06:02,240 management took the bits they liked and 182 00:06:02,240 --> 00:06:03,919 rejected the bits that actually made it 183 00:06:03,919 --> 00:06:06,000 work. The most productive teams I've 184 00:06:06,000 --> 00:06:08,240 been part of as a manager or leader or 185 00:06:08,240 --> 00:06:10,400 as a member have allowed a lot of 186 00:06:10,400 --> 00:06:12,240 communication and discussion as to how 187 00:06:12,240 --> 00:06:14,479 all that process stuff works and also 188 00:06:14,479 --> 00:06:15,840 have been willing to flex what's 189 00:06:15,840 --> 00:06:17,840 delivered when based on the feedback 190 00:06:17,840 --> 00:06:20,319 that uh the people delivering the work. 191 00:06:20,319 --> 00:06:21,759 It was really interesting listening to 192 00:06:21,759 --> 00:06:23,360 Liz's keynote yesterday and looking at 193 00:06:23,360 --> 00:06:24,560 the way Honeycomb's adoption of 194 00:06:24,560 --> 00:06:26,479 generative AI was structured since it 195 00:06:26,479 --> 00:06:28,479 looked a lot like what I'd expect from 196 00:06:28,479 --> 00:06:30,160 an organization that fits with what I'd 197 00:06:30,160 --> 00:06:32,720 expect from the agile manifesto. The way 198 00:06:32,720 --> 00:06:34,560 in which adoption happened there seemed 199 00:06:34,560 --> 00:06:36,479 very collaborative and workerled even if 200 00:06:36,479 --> 00:06:38,880 the objective was set by management. I 201 00:06:38,880 --> 00:06:40,800 also asked a question of Liz afterwards 202 00:06:40,800 --> 00:06:42,240 that had been sitting with me through a 203 00:06:42,240 --> 00:06:44,560 lot of her talk. Did all the workflow 204 00:06:44,560 --> 00:06:46,080 improvements that they developed to 205 00:06:46,080 --> 00:06:47,840 support the use of agents and generative 206 00:06:47,840 --> 00:06:50,639 AI also make workers who weren't using 207 00:06:50,639 --> 00:06:52,160 that tooling able to deliver more 208 00:06:52,160 --> 00:06:55,919 efficiently? And the answer was yes. 209 00:06:55,919 --> 00:06:58,319 But now let's get on to the generative 210 00:06:58,319 --> 00:07:00,319 tools themselves. So Daisy, they're 211 00:07:00,319 --> 00:07:01,840 going to make me what, 10 times more 212 00:07:01,840 --> 00:07:03,360 productive, right? If I hold them right. 213 00:07:03,360 --> 00:07:06,720 Well, I've heard that said, uh, but it 214 00:07:06,720 --> 00:07:08,160 seems to be pretty widely agreed at this 215 00:07:08,160 --> 00:07:09,759 point that that sort of productivity 216 00:07:09,759 --> 00:07:11,840 boost isn't really realistic. Nobody's 217 00:07:11,840 --> 00:07:13,120 getting a Fortnite's worth of work done 218 00:07:13,120 --> 00:07:14,880 in a day. As for how much of a 219 00:07:14,880 --> 00:07:16,479 productivity boost or burden generative 220 00:07:16,479 --> 00:07:18,639 AI can provide, everyone in this room 221 00:07:18,639 --> 00:07:20,560 pretty much already has a pretty strong 222 00:07:20,560 --> 00:07:21,840 opinion on how useful they are. And 223 00:07:21,840 --> 00:07:23,759 those opinions go in wildly different 224 00:07:23,759 --> 00:07:26,160 directions. There's a rather small body 225 00:07:26,160 --> 00:07:28,479 of my script not scrolling properly. 226 00:07:28,479 --> 00:07:30,560 Academic literature mostly funded by AI 227 00:07:30,560 --> 00:07:32,160 labs themselves with mixed and 228 00:07:32,160 --> 00:07:34,319 inconsistent conclusions. The best we 229 00:07:34,319 --> 00:07:36,000 can really say at the moment is it's 230 00:07:36,000 --> 00:07:37,759 highly dependent on the task and the 231 00:07:37,759 --> 00:07:39,919 person doing it and a whole lot of other 232 00:07:39,919 --> 00:07:43,039 variables. So in this talk about AI and 233 00:07:43,039 --> 00:07:44,479 productivity, we're not actually going 234 00:07:44,479 --> 00:07:46,400 to talk about whether and how much AI 235 00:07:46,400 --> 00:07:47,440 makes you more productive. 236 00:07:47,440 --> 00:07:49,039 Not really. Sorry. I kind of want to 237 00:07:49,039 --> 00:07:51,199 step back and look at the benefits and 238 00:07:51,199 --> 00:07:53,039 the costs and where they fall. 239 00:07:53,039 --> 00:07:56,160 So the energy usage and climate 240 00:07:56,160 --> 00:07:58,080 um enormous data centers and how they 241 00:07:58,080 --> 00:07:59,280 impact communities. 242 00:07:59,280 --> 00:08:01,520 No. And this next line is wrong. It's 243 00:08:01,520 --> 00:08:02,000 about 244 00:08:02,000 --> 00:08:03,599 training using cheap labor in developing 245 00:08:03,599 --> 00:08:04,800 countries. Um 246 00:08:04,800 --> 00:08:06,720 sorry, I forgot to edit edit the script. 247 00:08:06,720 --> 00:08:08,000 Uh the line should have been something 248 00:08:08,000 --> 00:08:10,160 else. Uh the fact the fact that Cam 249 00:08:10,160 --> 00:08:13,280 Wilson says that everyone hates us now. 250 00:08:13,280 --> 00:08:15,520 No. Although the first three things that 251 00:08:15,520 --> 00:08:18,000 you mentioned are probably a big part of 252 00:08:18,000 --> 00:08:20,000 why we're more unpopular than we've ever 253 00:08:20,000 --> 00:08:22,390 been, which again Cam kind of went to 254 00:08:22,390 --> 00:08:24,720 [snorts] do in a lot of detail. Uh those 255 00:08:24,720 --> 00:08:26,160 topics are all really really important. 256 00:08:26,160 --> 00:08:27,599 We've only got half an hour and they all 257 00:08:27,599 --> 00:08:29,120 discussed at length in a lot of places. 258 00:08:29,120 --> 00:08:31,120 This is more about the relationship 259 00:08:31,120 --> 00:08:32,959 between us as workers, the places we 260 00:08:32,959 --> 00:08:35,680 work and the work that we do. And so one 261 00:08:35,680 --> 00:08:37,360 of the costs that impacts all three of 262 00:08:37,360 --> 00:08:39,279 those things but falls mostly on us is 263 00:08:39,279 --> 00:08:40,240 deskkilling. 264 00:08:40,240 --> 00:08:41,919 But why worry about that though? I mean 265 00:08:41,919 --> 00:08:43,519 I I forget things all the time. I need 266 00:08:43,519 --> 00:08:45,519 to go look up documentation and it's 267 00:08:45,519 --> 00:08:46,959 easy to pick it back up when I need it 268 00:08:46,959 --> 00:08:49,360 later. If I can get an AI to manage my 269 00:08:49,360 --> 00:08:50,800 Git repository for me, it's not really a 270 00:08:50,800 --> 00:08:51,920 problem if I forget how to do that 271 00:08:51,920 --> 00:08:53,360 myself. I can just go look it up later, 272 00:08:53,360 --> 00:08:53,600 right? 273 00:08:53,600 --> 00:08:55,279 I mean that's very true for surface 274 00:08:55,279 --> 00:08:56,800 level skills. It's a common occurrence 275 00:08:56,800 --> 00:08:58,880 to have those sorts of things fall in 276 00:08:58,880 --> 00:09:01,760 and out of our cognitive toolbox uh as 277 00:09:01,760 --> 00:09:03,200 it were depending on how often we need 278 00:09:03,200 --> 00:09:04,399 them. But what about the more 279 00:09:04,399 --> 00:09:06,480 foundational skills? Systems design, 280 00:09:06,480 --> 00:09:08,959 critical thinking. What happens to the 281 00:09:08,959 --> 00:09:10,800 creativity and curiosity that makes us 282 00:09:10,800 --> 00:09:12,959 good at our jobs? Pilots have to do 283 00:09:12,959 --> 00:09:14,560 manual takeoffs and landings and 284 00:09:14,560 --> 00:09:16,160 simulator training on a regular basis. 285 00:09:16,160 --> 00:09:18,080 It's it's doing it. It's fine. In order 286 00:09:18,080 --> 00:09:19,440 to [laughter] 287 00:09:19,440 --> 00:09:21,200 in order to remain able to do their jobs 288 00:09:21,200 --> 00:09:23,120 precisely because a lot of what they do 289 00:09:23,120 --> 00:09:25,120 dayto-day is automated, but they still 290 00:09:25,120 --> 00:09:26,560 need to be able to take over manually in 291 00:09:26,560 --> 00:09:27,760 an emergency. 292 00:09:27,760 --> 00:09:29,680 So do a side project like developing a 293 00:09:29,680 --> 00:09:31,821 teleprompter app. 294 00:09:31,821 --> 00:09:33,841 [laughter] 295 00:09:34,160 --> 00:09:37,200 True story. True story. I have one that 296 00:09:37,200 --> 00:09:39,200 I've written. We I was wanted to use it 297 00:09:39,200 --> 00:09:41,360 for this but didn't have the time to 298 00:09:41,360 --> 00:09:43,600 uplift it to for what we needed and 299 00:09:43,600 --> 00:09:44,800 using offtheshelf software and it's 300 00:09:44,800 --> 00:09:46,000 going badly. Um 301 00:09:46,000 --> 00:09:47,519 I've seen plenty of people who are using 302 00:09:47,519 --> 00:09:49,360 AI extensively in day jobs who spun up 303 00:09:49,360 --> 00:09:51,200 side projects where they're not using AI 304 00:09:51,200 --> 00:09:53,519 precisely to keep their skills sharp. 305 00:09:53,519 --> 00:09:55,040 And that's fine if an ongoing 306 00:09:55,040 --> 00:09:56,560 time-intensive side project is something 307 00:09:56,560 --> 00:09:58,160 you have the time and capacity for. And 308 00:09:58,160 --> 00:10:00,640 that biases towards younger middle to 309 00:10:00,640 --> 00:10:02,240 upper class white male developers. 310 00:10:02,240 --> 00:10:04,240 Simulator training is part of a pilot's 311 00:10:04,240 --> 00:10:05,839 job and something they get paid for. 312 00:10:05,839 --> 00:10:07,760 I mean, Google used to do the whole 20% 313 00:10:07,760 --> 00:10:08,320 time thing. 314 00:10:08,320 --> 00:10:10,640 Used to. 315 00:10:10,640 --> 00:10:11,839 True. 316 00:10:11,839 --> 00:10:13,600 Speaking of Google engineers, do you 317 00:10:13,600 --> 00:10:15,600 remember the 10x developer phenomenon of 318 00:10:15,600 --> 00:10:16,640 the 2010s? 319 00:10:16,640 --> 00:10:17,600 Look, I [snorts] think we're both old 320 00:10:17,600 --> 00:10:18,720 enough that we can't really get away 321 00:10:18,720 --> 00:10:20,160 with pretending that we don't, but I 322 00:10:20,160 --> 00:10:21,519 guess one of us is going to have to 323 00:10:21,519 --> 00:10:22,320 explain it anyway. 324 00:10:22,320 --> 00:10:25,760 Go on then. So [laughter] 10x developers 325 00:10:25,760 --> 00:10:27,680 or rockstar developers or ninja 326 00:10:27,680 --> 00:10:29,440 developers where this class of super 327 00:10:29,440 --> 00:10:31,279 developers who were so good at their 328 00:10:31,279 --> 00:10:32,880 jobs that they were 10 times as 329 00:10:32,880 --> 00:10:34,560 productive as your average developer you 330 00:10:34,560 --> 00:10:35,760 know just people who were built 331 00:10:35,760 --> 00:10:37,200 different perhaps even better than the 332 00:10:37,200 --> 00:10:39,360 gods and were really really good at 333 00:10:39,360 --> 00:10:39,920 their job. 334 00:10:39,920 --> 00:10:41,120 Well I mean that was kind of the 335 00:10:41,120 --> 00:10:42,959 mythology of the 10x developer. The 336 00:10:42,959 --> 00:10:44,880 other perspective on 10x developers is 337 00:10:44,880 --> 00:10:46,720 that they hurriedly throw sloppy code 338 00:10:46,720 --> 00:10:48,480 together and they leave in their wake 339 00:10:48,480 --> 00:10:50,560 huge amounts of rework technical debt 340 00:10:50,560 --> 00:10:52,880 and maintenance burden. The velocity 341 00:10:52,880 --> 00:10:54,800 that they're lorded for is achieved by 342 00:10:54,800 --> 00:10:56,640 carelessly externalizing the costs of 343 00:10:56,640 --> 00:10:58,320 moving that fast, usually onto their 344 00:10:58,320 --> 00:10:59,360 co-workers. 345 00:10:59,360 --> 00:11:01,040 So, the work isn't as good as it's made 346 00:11:01,040 --> 00:11:02,720 out to be, and that's why it's so fast. 347 00:11:02,720 --> 00:11:04,240 Sort of. The cost of the high velocity 348 00:11:04,240 --> 00:11:06,079 is a lot of high interest technical debt 349 00:11:06,079 --> 00:11:07,279 that the people around them end up 350 00:11:07,279 --> 00:11:08,880 having to pay off. That creates 351 00:11:08,880 --> 00:11:10,640 resentment from the 10X's co-workers, 352 00:11:10,640 --> 00:11:12,880 and it fractures team cohesion. So, the 353 00:11:12,880 --> 00:11:14,480 short-term gain in efficiency has a 354 00:11:14,480 --> 00:11:16,560 longerterm technical cost and a social 355 00:11:16,560 --> 00:11:19,279 cost. And sometimes LLM assisted coding 356 00:11:19,279 --> 00:11:21,360 can look a lot like that too because a 357 00:11:21,360 --> 00:11:23,839 lot of the work becomes verifying the 358 00:11:23,839 --> 00:11:26,000 output of the LLM and LLMs are very good 359 00:11:26,000 --> 00:11:28,240 at engineering uh not engineering very 360 00:11:28,240 --> 00:11:29,360 good at generating code that 361 00:11:29,360 --> 00:11:31,360 superficially looks acceptable but has 362 00:11:31,360 --> 00:11:34,160 uh subtle bugs under the surface. So the 363 00:11:34,160 --> 00:11:35,839 work can be that that work of 364 00:11:35,839 --> 00:11:37,279 verification can be very hard and 365 00:11:37,279 --> 00:11:39,120 timeconuming to do well and very easy to 366 00:11:39,120 --> 00:11:41,600 skip because it's very easy to think 367 00:11:41,600 --> 00:11:43,279 that the sole output of writing code is 368 00:11:43,279 --> 00:11:45,200 the code but it isn't. When you write 369 00:11:45,200 --> 00:11:47,279 code yourself, um, you're also 370 00:11:47,279 --> 00:11:48,880 simultaneously creating a human 371 00:11:48,880 --> 00:11:51,839 understanding of what that code does, 372 00:11:51,839 --> 00:11:53,600 how it works, and why it is the way that 373 00:11:53,600 --> 00:11:55,360 it is. Which is why if you've ever 374 00:11:55,360 --> 00:11:56,800 accidentally deleted code and had to 375 00:11:56,800 --> 00:11:58,480 rewrite it, it's so much quicker the 376 00:11:58,480 --> 00:12:01,040 second time. But when an LLM generates 377 00:12:01,040 --> 00:12:03,279 code for you, you skip entirely over the 378 00:12:03,279 --> 00:12:04,000 second part 379 00:12:04,000 --> 00:12:05,680 and then it becomes the reviewer's job. 380 00:12:05,680 --> 00:12:07,279 Exactly. And so there's kind of this 381 00:12:07,279 --> 00:12:08,880 implicit social contract with code 382 00:12:08,880 --> 00:12:11,360 review uh that we've always had without 383 00:12:11,360 --> 00:12:13,200 realizing it where we expect that the 384 00:12:13,200 --> 00:12:14,720 author has done the work of creating 385 00:12:14,720 --> 00:12:16,800 that understanding and can therefore use 386 00:12:16,800 --> 00:12:17,920 that in that back and forth. And the 387 00:12:17,920 --> 00:12:20,560 LLMs make it really easy to break that 388 00:12:20,560 --> 00:12:22,560 social contract. I mean the developer 389 00:12:22,560 --> 00:12:23,760 writing the code is kind of damned if 390 00:12:23,760 --> 00:12:24,959 they do, damned if they don't though. 391 00:12:24,959 --> 00:12:26,079 They've been given this tool that's 392 00:12:26,079 --> 00:12:28,079 supposed to make them five 10 times as 393 00:12:28,079 --> 00:12:30,320 as fast and they're being pressured to 394 00:12:30,320 --> 00:12:30,959 use it. 395 00:12:30,959 --> 00:12:32,320 And this is where we kind of do touch on 396 00:12:32,320 --> 00:12:33,600 the question of how much faster they 397 00:12:33,600 --> 00:12:35,839 make you. There was a 2025 study where 398 00:12:35,839 --> 00:12:37,760 developers perceived themselves being 399 00:12:37,760 --> 00:12:39,920 about 20% faster with AI, but they were 400 00:12:39,920 --> 00:12:41,519 actually about 20% slower. 401 00:12:41,519 --> 00:12:43,040 Isn't that study flawed, though? I mean, 402 00:12:43,040 --> 00:12:44,880 it was a really small sample size and 403 00:12:44,880 --> 00:12:46,320 was like 18 months old and was only 404 00:12:46,320 --> 00:12:47,600 testing one particular model. 405 00:12:47,600 --> 00:12:49,360 I mean, sure, the percentages don't 406 00:12:49,360 --> 00:12:51,200 really mean anything by themselves, but 407 00:12:51,200 --> 00:12:52,800 what's interesting is the difference. It 408 00:12:52,800 --> 00:12:54,399 highlights how it's really hard to 409 00:12:54,399 --> 00:12:56,560 actually perceive how efficient you are 410 00:12:56,560 --> 00:12:58,320 and whether LLM are speeding you up in 411 00:12:58,320 --> 00:12:59,519 the first place. Something that I think 412 00:12:59,519 --> 00:13:01,920 is true today. And then there's the 413 00:13:01,920 --> 00:13:04,959 executive level. uh a late 2025 survey 414 00:13:04,959 --> 00:13:07,760 of white collar workers found that 415 00:13:07,760 --> 00:13:09,839 seuite executives per perceptions of 416 00:13:09,839 --> 00:13:11,519 their own use of AI are massively 417 00:13:11,519 --> 00:13:13,040 different to non-man management workers 418 00:13:13,040 --> 00:13:15,200 and that is a slide cue for you better. 419 00:13:15,200 --> 00:13:15,519 There you go. 420 00:13:15,519 --> 00:13:17,799 Thank you. Uh 421 00:13:17,799 --> 00:13:18,320 [snorts] 422 00:13:18,320 --> 00:13:21,200 um so there's two layers here. There's 423 00:13:21,200 --> 00:13:22,639 the difference between actual reality 424 00:13:22,639 --> 00:13:24,160 and developer perception and then the 425 00:13:24,160 --> 00:13:25,120 difference between the developer 426 00:13:25,120 --> 00:13:26,880 perception and leadership perception. 427 00:13:26,880 --> 00:13:30,240 Exactly. So, so far it seems like 2026 428 00:13:30,240 --> 00:13:31,920 is the year of eternal crunch where so 429 00:13:31,920 --> 00:13:33,360 many individual contributors and team 430 00:13:33,360 --> 00:13:34,880 leads across so many companies are 431 00:13:34,880 --> 00:13:37,200 rolling their eyes and sarcastically 432 00:13:37,200 --> 00:13:38,480 saying to each other, "We're being 433 00:13:38,480 --> 00:13:40,560 expected to do more with less, but it's 434 00:13:40,560 --> 00:13:43,040 okay cuz AI will solve it, right?" And 435 00:13:43,040 --> 00:13:45,519 then we get keeping get we keep getting 436 00:13:45,519 --> 00:13:47,120 told that the next generation of models 437 00:13:47,120 --> 00:13:48,560 will solve all the problems of the other 438 00:13:48,560 --> 00:13:50,320 ones and the perfect generative AI tool 439 00:13:50,320 --> 00:13:52,560 is just around the corner. 440 00:13:52,560 --> 00:13:56,800 So, speaking of the future, um I've been 441 00:13:56,800 --> 00:13:58,000 around a lot of systems and you know 442 00:13:58,000 --> 00:13:58,959 when you've been around a lot of 443 00:13:58,959 --> 00:14:00,160 systems, you get pretty good at 444 00:14:00,160 --> 00:14:02,079 understanding how they work, but that 445 00:14:02,079 --> 00:14:04,720 doesn't make me an economist. Um that 446 00:14:04,720 --> 00:14:06,399 said, I think it's worth taking a look 447 00:14:06,399 --> 00:14:08,160 at the current state of the the big 448 00:14:08,160 --> 00:14:09,839 players in the AI space and what that 449 00:14:09,839 --> 00:14:12,240 might mean for us as workers. 450 00:14:12,240 --> 00:14:13,440 You mean about how it's all going to 451 00:14:13,440 --> 00:14:14,240 crash? 452 00:14:14,240 --> 00:14:18,160 Well, maybe and maybe not. I mean, the 453 00:14:18,160 --> 00:14:20,160 models themselves, the LLMs and a bunch 454 00:14:20,160 --> 00:14:21,920 of the ways that we use them aren't 455 00:14:21,920 --> 00:14:24,480 going away. That said, they're not the 456 00:14:24,480 --> 00:14:26,000 entirety of the thing. A lot of the 457 00:14:26,000 --> 00:14:27,440 model use that we see at the moment is 458 00:14:27,440 --> 00:14:29,120 through some kind of service provider. 459 00:14:29,120 --> 00:14:30,880 So, the service providers are going to 460 00:14:30,880 --> 00:14:34,240 crash. Maybe when you break it down, 461 00:14:34,240 --> 00:14:36,240 running inference with an LLM is a large 462 00:14:36,240 --> 00:14:37,519 amount of floatingoint matrix 463 00:14:37,519 --> 00:14:39,440 multiplication operations. That's a lot 464 00:14:39,440 --> 00:14:40,959 of compute and a lot of bandwidth to 465 00:14:40,959 --> 00:14:42,720 move all that data around. And then when 466 00:14:42,720 --> 00:14:44,240 you add in the fact that you need to do 467 00:14:44,240 --> 00:14:46,160 an inference run for each output token 468 00:14:46,160 --> 00:14:48,320 in an LM conversation, you're also 469 00:14:48,320 --> 00:14:50,240 adding in a lot of memory to hold state. 470 00:14:50,240 --> 00:14:51,519 All of this adds up to a lot of 471 00:14:51,519 --> 00:14:52,880 computers with a lot of specialized 472 00:14:52,880 --> 00:14:54,959 hardware, which is why you can't buy RAM 473 00:14:54,959 --> 00:14:57,199 anymore. But it's also a lot of capital 474 00:14:57,199 --> 00:14:58,959 expenditure on assets that it costs a 475 00:14:58,959 --> 00:15:00,240 lot more to run. 476 00:15:00,240 --> 00:15:03,040 So they are going to crash. Again, I'm 477 00:15:03,040 --> 00:15:05,279 not an amic economics person, but the 478 00:15:05,279 --> 00:15:07,920 fact that the largest players uh OpenAI, 479 00:15:07,920 --> 00:15:10,160 Anthropic, a few others haven't really 480 00:15:10,160 --> 00:15:12,720 declared profits and seem to require 481 00:15:12,720 --> 00:15:14,320 regular injections of investment capital 482 00:15:14,320 --> 00:15:16,480 to continue to function. Doesn't sound 483 00:15:16,480 --> 00:15:18,000 like a good place to be. 484 00:15:18,000 --> 00:15:20,160 But hang on, I thought Entropic reported 485 00:15:20,160 --> 00:15:21,040 a profit. 486 00:15:21,040 --> 00:15:23,199 They did. They reported a profit for the 487 00:15:23,199 --> 00:15:25,839 second quarter of 2026, but it wasn't 488 00:15:25,839 --> 00:15:27,839 using standard accounting methods. And 489 00:15:27,839 --> 00:15:29,519 it includes a period of time where they 490 00:15:29,519 --> 00:15:31,839 got to use Elon Musk's compute for very 491 00:15:31,839 --> 00:15:33,920 cheap because Elon Musk wasn't using it, 492 00:15:33,920 --> 00:15:36,320 which should be embarrassing for him. On 493 00:15:36,320 --> 00:15:37,920 top of that, they're foreshadowing that 494 00:15:37,920 --> 00:15:39,839 this this might not extend to being 495 00:15:39,839 --> 00:15:41,440 profitable for the entire year. 496 00:15:41,440 --> 00:15:43,440 But aren't they profitable on inference 497 00:15:43,440 --> 00:15:44,240 at least though? 498 00:15:44,240 --> 00:15:46,800 I mean, maybe. I assume so for some 499 00:15:46,800 --> 00:15:48,240 models, especially older ones, but 500 00:15:48,240 --> 00:15:50,480 training isn't free. Uh the challenge 501 00:15:50,480 --> 00:15:51,839 for all of these companies is to work 502 00:15:51,839 --> 00:15:53,680 out how to train a model at a cost that 503 00:15:53,680 --> 00:15:55,600 they can amatize over a reasonable time 504 00:15:55,600 --> 00:15:57,279 frame with the money made through 505 00:15:57,279 --> 00:15:59,279 inference. For a while there it was 506 00:15:59,279 --> 00:16:00,480 questionable whether they were going to 507 00:16:00,480 --> 00:16:02,079 be able to do that before they ran out 508 00:16:02,079 --> 00:16:06,079 of runway. Now maybe but there's still I 509 00:16:06,079 --> 00:16:07,680 think a high likelihood of at least one 510 00:16:07,680 --> 00:16:10,079 of the big providers flames out. The 511 00:16:10,079 --> 00:16:11,199 question is what that does to the rest 512 00:16:11,199 --> 00:16:11,920 of the industry. 513 00:16:11,920 --> 00:16:14,800 What do you mean? Well, in 2000, for 514 00:16:14,800 --> 00:16:17,271 those of us with beards my color, um we 515 00:16:17,271 --> 00:16:19,839 [laughter] we had the.com bust. There 516 00:16:19,839 --> 00:16:21,199 was a whole bunch of investor capital 517 00:16:21,199 --> 00:16:22,240 that decided they weren't going to get 518 00:16:22,240 --> 00:16:23,519 the returns out of tech company 519 00:16:23,519 --> 00:16:25,920 investment that they wanted. And after a 520 00:16:25,920 --> 00:16:28,480 period of say uh overexuberant 521 00:16:28,480 --> 00:16:30,720 investment in internet related companies 522 00:16:30,720 --> 00:16:32,160 and the outcome of that was that it 523 00:16:32,160 --> 00:16:33,360 became really hard to get venture 524 00:16:33,360 --> 00:16:35,120 capital and that led to a lot of 525 00:16:35,120 --> 00:16:37,759 companies failing, los less hiring, jobs 526 00:16:37,759 --> 00:16:39,759 crashed, all that kind of stuff. In a 527 00:16:39,759 --> 00:16:42,399 different direction, we had 2008, which 528 00:16:42,399 --> 00:16:44,079 was where the amount of debt and other 529 00:16:44,079 --> 00:16:45,920 financial instruments linked to that 530 00:16:45,920 --> 00:16:48,399 became too large and too confused, which 531 00:16:48,399 --> 00:16:51,279 effectively made money stop moving. Um, 532 00:16:51,279 --> 00:16:53,440 this was the global financial crisis. 533 00:16:53,440 --> 00:16:55,199 And the thing is that a lot of data 534 00:16:55,199 --> 00:16:56,800 center buildouts that you hear about are 535 00:16:56,800 --> 00:17:00,320 financed with debt. This is normal, but 536 00:17:00,320 --> 00:17:04,400 the scale of it is so big that um if the 537 00:17:04,400 --> 00:17:06,160 expected returns don't show up, a bunch 538 00:17:06,160 --> 00:17:08,160 of them could turn into really bad debt. 539 00:17:08,160 --> 00:17:10,000 And if that goes too far and too wide, 540 00:17:10,000 --> 00:17:12,799 you could get a 2008 scenario. Again, 541 00:17:12,799 --> 00:17:16,079 this is still hypothetical, right? 542 00:17:16,079 --> 00:17:18,480 Absolutely. This is not the guaranteed 543 00:17:18,480 --> 00:17:20,079 thing. But I bring it up because we need 544 00:17:20,079 --> 00:17:22,160 to be ready to look after each other if 545 00:17:22,160 --> 00:17:24,319 anything like this happens. We've 546 00:17:24,319 --> 00:17:25,839 already seen that the job market's 547 00:17:25,839 --> 00:17:28,319 gotten a bit weird lately. If things get 548 00:17:28,319 --> 00:17:29,760 messy, we're going to have to a lot of 549 00:17:29,760 --> 00:17:31,600 people in our community who need help. 550 00:17:31,600 --> 00:17:33,120 Whether that's just in finding a new gig 551 00:17:33,120 --> 00:17:35,039 or in other ways, we need to take care 552 00:17:35,039 --> 00:17:36,559 of us. 553 00:17:36,559 --> 00:17:38,640 The thing is though, the thing about a 554 00:17:38,640 --> 00:17:40,960 bubble is you can't really be 100% sure 555 00:17:40,960 --> 00:17:42,720 you're in one unless your name is 556 00:17:42,720 --> 00:17:44,400 Editron, I guess. [snorts] So, it's 557 00:17:44,400 --> 00:17:46,559 worth [laughter] 558 00:17:46,559 --> 00:17:48,240 it's worth thinking about an alternate 559 00:17:48,240 --> 00:17:50,400 future. One where most of the code at 560 00:17:50,400 --> 00:17:52,240 most of the places we work is written by 561 00:17:52,240 --> 00:17:54,240 generative AI and our job description 562 00:17:54,240 --> 00:17:55,679 becomes professional reviewers, the 563 00:17:55,679 --> 00:17:57,679 human in the loop, if you will. Uh, 564 00:17:57,679 --> 00:17:59,039 those that are following along, feel 565 00:17:59,039 --> 00:18:01,679 free to take a drink. Um, have you ever 566 00:18:01,679 --> 00:18:05,840 worked in uh retail or fast food? Um, 567 00:18:05,840 --> 00:18:07,440 no. Has that got anything to do 568 00:18:07,440 --> 00:18:08,480 anything? 569 00:18:08,480 --> 00:18:10,160 Let's say you're a line cook at a uh 570 00:18:10,160 --> 00:18:12,559 popular fast food burger joint. Uh, on 571 00:18:12,559 --> 00:18:14,000 the one hand, you have efficiency rules. 572 00:18:14,000 --> 00:18:15,679 The amount of time every car spends 573 00:18:15,679 --> 00:18:17,200 waiting in the drive-thru queue is 574 00:18:17,200 --> 00:18:18,720 measured down to the second, and you're 575 00:18:18,720 --> 00:18:20,080 pushed and pushed and pushed to squeeze 576 00:18:20,080 --> 00:18:22,320 that KPI down as much as you can. And 577 00:18:22,320 --> 00:18:23,919 then you have the quality rules. There's 578 00:18:23,919 --> 00:18:25,200 rules about the actual quality of the 579 00:18:25,200 --> 00:18:27,679 product, work safety rules, food safety 580 00:18:27,679 --> 00:18:28,960 rules, and so on. 581 00:18:28,960 --> 00:18:30,480 So, you've got to find a balance between 582 00:18:30,480 --> 00:18:31,200 those. 583 00:18:31,200 --> 00:18:32,880 That's just it, though. You can't find a 584 00:18:32,880 --> 00:18:34,480 balance. There's no way to perfectly 585 00:18:34,480 --> 00:18:36,240 follow all of those rules and procedures 586 00:18:36,240 --> 00:18:38,000 around quality and also meet the speed 587 00:18:38,000 --> 00:18:40,400 requirements. And so you cut corners and 588 00:18:40,400 --> 00:18:42,400 hope that it works out okay. And if it 589 00:18:42,400 --> 00:18:44,320 doesn't, you get reprimanded or fired 590 00:18:44,320 --> 00:18:46,320 because you didn't follow the rules. It 591 00:18:46,320 --> 00:18:47,760 doesn't matter that nobody ever follows 592 00:18:47,760 --> 00:18:48,799 the rules. It doesn't matter that the 593 00:18:48,799 --> 00:18:50,320 rules are literally impossible to 594 00:18:50,320 --> 00:18:52,000 follow. The rules aren't there to be 595 00:18:52,000 --> 00:18:53,679 followed. They're there to assign blame. 596 00:18:53,679 --> 00:18:55,600 And turnover turnover is high anyway. So 597 00:18:55,600 --> 00:18:57,039 getting rid of one resource and 598 00:18:57,039 --> 00:18:59,120 replacing them with another is pretty 599 00:18:59,120 --> 00:19:00,000 low cost. 600 00:19:00,000 --> 00:19:01,679 I mean, that's a long way from where we 601 00:19:01,679 --> 00:19:03,360 are right now, though. We're skilled 602 00:19:03,360 --> 00:19:03,919 professionals. 603 00:19:03,919 --> 00:19:06,480 Yeah, true 604 00:19:06,480 --> 00:19:08,880 to some extent. True. Uh our industry 605 00:19:08,880 --> 00:19:10,160 won't become like that overnight. 606 00:19:10,160 --> 00:19:11,760 Perhaps not at all. But it's not hard to 607 00:19:11,760 --> 00:19:13,360 imagine the most powerful, the most 608 00:19:13,360 --> 00:19:15,039 ruthless, the most capitalistic leaders 609 00:19:15,039 --> 00:19:16,640 in our industry wanting to push it in 610 00:19:16,640 --> 00:19:18,640 that direction. For them, it's kind of 611 00:19:18,640 --> 00:19:20,640 the ideal state for a workforce to be 612 00:19:20,640 --> 00:19:22,480 in. The minimum amount of money possible 613 00:19:22,480 --> 00:19:24,400 goes in. The maximum amount of work 614 00:19:24,400 --> 00:19:26,480 output comes out. The individual workers 615 00:19:26,480 --> 00:19:29,120 are totally funible parts of the system 616 00:19:29,120 --> 00:19:30,799 that can be swapped out whenever it's 617 00:19:30,799 --> 00:19:32,400 convenient to keep the overall system 618 00:19:32,400 --> 00:19:33,600 running. Bet you didn't think you still 619 00:19:33,600 --> 00:19:35,083 had to think about the word funible. 620 00:19:35,083 --> 00:19:36,160 [snorts] 621 00:19:36,160 --> 00:19:37,840 The uh infamous Silicon Valley 622 00:19:37,840 --> 00:19:39,679 billboards about how AI won't complain 623 00:19:39,679 --> 00:19:41,600 if you overwork or harass them kind of 624 00:19:41,600 --> 00:19:43,440 illustrates this this kind of thinking 625 00:19:43,440 --> 00:19:45,120 that I'm talking about. 626 00:19:45,120 --> 00:19:47,360 So the worry is that we become less of 627 00:19:47,360 --> 00:19:49,200 the creators of the systems we build and 628 00:19:49,200 --> 00:19:51,919 more of a human in the loop. we become 629 00:19:51,919 --> 00:19:54,559 more we become more replaceable 630 00:19:54,559 --> 00:19:56,640 sort of the risk is we become moral 631 00:19:56,640 --> 00:19:59,760 frumple zones which is to say 632 00:19:59,760 --> 00:20:02,240 someone's read the paper um a human 633 00:20:02,240 --> 00:20:03,840 whose role is to absorb a lot of the 634 00:20:03,840 --> 00:20:05,679 accountability of the system but has 635 00:20:05,679 --> 00:20:07,679 comparatively little autonomy that term 636 00:20:07,679 --> 00:20:10,320 comes from a 2019 paper uh which talks 637 00:20:10,320 --> 00:20:13,280 about human in the loop situations uh 638 00:20:13,280 --> 00:20:15,200 like a fatal Uber driverless car crash 639 00:20:15,200 --> 00:20:16,880 where control was thrust into the safety 640 00:20:16,880 --> 00:20:18,799 driver's hands at the last moment or the 641 00:20:18,799 --> 00:20:20,080 Air France disaster 642 00:20:20,080 --> 00:20:22,160 which again most of us aren't making 643 00:20:22,160 --> 00:20:23,919 time-sensitive safety critical decisions 644 00:20:23,919 --> 00:20:25,360 while flying on the planes in the air. 645 00:20:25,360 --> 00:20:25,760 Right. 646 00:20:25,760 --> 00:20:27,360 True. These are the most extreme 647 00:20:27,360 --> 00:20:29,039 examples of this phenomenon because 648 00:20:29,039 --> 00:20:30,880 they're the most illustrative. They 649 00:20:30,880 --> 00:20:32,559 illustrate an end point and the human in 650 00:20:32,559 --> 00:20:35,120 the loop model of AI generated software 651 00:20:35,120 --> 00:20:36,880 risks pushing us a fair way in that 652 00:20:36,880 --> 00:20:38,000 direction. And Beno's forgotten about 653 00:20:38,000 --> 00:20:40,720 the bit cuz I said before 654 00:20:40,720 --> 00:20:43,120 I have failed all of you. I'm sorry. So 655 00:20:43,120 --> 00:20:44,480 there's a risk that our purpose becomes 656 00:20:44,480 --> 00:20:46,400 less about preventing flawed generative 657 00:20:46,400 --> 00:20:48,880 AI code and more becoming an ablative 658 00:20:48,880 --> 00:20:50,880 blame shield of sorts when it does go 659 00:20:50,880 --> 00:20:51,520 wrong. 660 00:20:51,520 --> 00:20:53,520 Exactly. And to make things worse, the 661 00:20:53,520 --> 00:20:56,320 way that we learn what good code is is 662 00:20:56,320 --> 00:20:57,440 by writing it 663 00:20:57,440 --> 00:20:59,679 or more often by writing bad code. 664 00:20:59,679 --> 00:21:01,760 Well, yeah, that too. Code review is one 665 00:21:01,760 --> 00:21:03,679 of those things that uses metaphorical 666 00:21:03,679 --> 00:21:05,120 muscles that it doesn't necessarily 667 00:21:05,120 --> 00:21:07,120 build by itself. So if all we do is 668 00:21:07,120 --> 00:21:08,559 review code and never write it, those 669 00:21:08,559 --> 00:21:10,400 skills atrophy. cognitive offloading, 670 00:21:10,400 --> 00:21:12,320 decision fatigue, deskilling, all of 671 00:21:12,320 --> 00:21:14,159 these things 672 00:21:14,159 --> 00:21:15,679 erode the understanding and the 673 00:21:15,679 --> 00:21:17,440 discernment that we use to evaluate 674 00:21:17,440 --> 00:21:19,120 whether code is effective. And on top of 675 00:21:19,120 --> 00:21:21,200 that, there's this pressure to go fast 676 00:21:21,200 --> 00:21:22,880 and the fact that saying no to the AI 677 00:21:22,880 --> 00:21:25,440 slows things down disincentivizes 678 00:21:25,440 --> 00:21:27,919 effective supervision even more. And 679 00:21:27,919 --> 00:21:29,679 then there's also the question of what 680 00:21:29,679 --> 00:21:31,360 happens to your brain when you spend 681 00:21:31,360 --> 00:21:33,200 significant amount of time interacting 682 00:21:33,200 --> 00:21:35,440 with something that isn't a human but is 683 00:21:35,440 --> 00:21:38,480 designed to make you react emotionally 684 00:21:38,480 --> 00:21:40,559 as though it is, which is a really weird 685 00:21:40,559 --> 00:21:43,440 thing for a programming tool to do. 686 00:21:43,440 --> 00:21:45,039 Here's a question, Ben. Do you know the 687 00:21:45,039 --> 00:21:48,049 thing about baby boomers and lead? What? 688 00:21:48,049 --> 00:21:50,070 [laughter] 689 00:21:50,080 --> 00:21:52,960 So, uh, there's a few other millennials 690 00:21:52,960 --> 00:21:54,320 in the crowd. Uh, I'm in the younger end 691 00:21:54,320 --> 00:21:55,919 of the millennials, just a couple years 692 00:21:55,919 --> 00:21:57,679 old to be Gen Z. So my peers and I have 693 00:21:57,679 --> 00:21:59,440 been seeing opinion pieces about how 694 00:21:59,440 --> 00:22:01,600 we're the downfall of society roughly 695 00:22:01,600 --> 00:22:03,280 since I was in school. And there's one 696 00:22:03,280 --> 00:22:04,880 piece of millennial folklore that tries 697 00:22:04,880 --> 00:22:06,240 to explain the intensity of the 698 00:22:06,240 --> 00:22:07,600 hostility that's been aimed at us since 699 00:22:07,600 --> 00:22:10,400 we were kids. And it's lead 700 00:22:10,400 --> 00:22:12,240 exposure from leaded petrol, lead based 701 00:22:12,240 --> 00:22:14,640 paint and so on. Lead exposure is linked 702 00:22:14,640 --> 00:22:16,159 with social aggression and there's no 703 00:22:16,159 --> 00:22:18,000 safe level of lead. And so the 704 00:22:18,000 --> 00:22:19,679 millennial folklore, I call it folklore 705 00:22:19,679 --> 00:22:22,159 because it's widely believed but kind of 706 00:22:22,159 --> 00:22:24,400 difficult to verify. Is the boomers are 707 00:22:24,400 --> 00:22:25,600 just hostile towards us because they 708 00:22:25,600 --> 00:22:27,919 were exposed to letters kids. 709 00:22:27,919 --> 00:22:30,080 So where's this going? AI is the 710 00:22:30,080 --> 00:22:32,000 lettered petrol of the 21st century. 711 00:22:32,000 --> 00:22:34,400 Well, it could be. I mean, we know about 712 00:22:34,400 --> 00:22:36,240 AI psychosis, which is where people 713 00:22:36,240 --> 00:22:38,240 believe that generative AI chat bots 714 00:22:38,240 --> 00:22:40,240 have revealed conspiracies are deities 715 00:22:40,240 --> 00:22:42,080 or just sentient or in a romantic 716 00:22:42,080 --> 00:22:43,840 relationship with the user. And that can 717 00:22:43,840 --> 00:22:45,200 have really disastrous and sometimes 718 00:22:45,200 --> 00:22:47,440 fatal effects. That's a really intense 719 00:22:47,440 --> 00:22:49,600 acute effect that's happening to a small 720 00:22:49,600 --> 00:22:50,880 number of people. 721 00:22:50,880 --> 00:22:52,880 So that bit is less like lead exposure 722 00:22:52,880 --> 00:22:54,640 and more like acute lead poisoning. 723 00:22:54,640 --> 00:22:56,480 Yeah. And then there's this paper which 724 00:22:56,480 --> 00:22:58,080 found an immediate term link between 725 00:22:58,080 --> 00:23:00,320 exposure to AI chatbots and decreased 726 00:23:00,320 --> 00:23:02,159 willingness to take responsibility and 727 00:23:02,159 --> 00:23:04,240 resolve interpersonal conflict but 728 00:23:04,240 --> 00:23:05,679 doesn't say anything about long-term 729 00:23:05,679 --> 00:23:06,640 effects. 730 00:23:06,640 --> 00:23:08,559 So it's not that AI is the leaded petrol 731 00:23:08,559 --> 00:23:10,559 of the 21st century. It's just more that 732 00:23:10,559 --> 00:23:12,000 we don't know if it will be. 733 00:23:12,000 --> 00:23:13,840 We don't. When it comes to much subtler 734 00:23:13,840 --> 00:23:15,280 and longerterm harms, we have early 735 00:23:15,280 --> 00:23:16,400 hints that they're probably not 736 00:23:16,400 --> 00:23:18,080 non-existent. We don't know how severe, 737 00:23:18,080 --> 00:23:20,400 how widespread, who is at higher risk, 738 00:23:20,400 --> 00:23:22,080 what it looks like, or how it shapes 739 00:23:22,080 --> 00:23:23,679 society. And just like with lettered 740 00:23:23,679 --> 00:23:26,320 petrol, we may never know for sure. 741 00:23:26,320 --> 00:23:29,039 Well, that's depressing. 742 00:23:29,039 --> 00:23:31,919 It is. And that is what made this talk 743 00:23:31,919 --> 00:23:33,520 really difficult to write for both of 744 00:23:33,520 --> 00:23:35,120 us. I really hope the futures we've 745 00:23:35,120 --> 00:23:36,480 talked about don't actually come to 746 00:23:36,480 --> 00:23:38,880 pass. Predicting the future is hard. 747 00:23:38,880 --> 00:23:40,240 It's really possible we're entirely 748 00:23:40,240 --> 00:23:42,880 wrong and I kind of hope that we are. 749 00:23:42,880 --> 00:23:46,240 And I mean going back to the um the 750 00:23:46,240 --> 00:23:48,880 Honeycomb stuff from yesterday, Charity 751 00:23:48,880 --> 00:23:51,520 Majors, the one of the the the other one 752 00:23:51,520 --> 00:23:53,360 of the other founders of Honeycomb, 753 00:23:53,360 --> 00:23:56,431 don't kill me, Liz. Um um also 754 00:23:56,431 --> 00:23:57,600 [clears throat] published several blog 755 00:23:57,600 --> 00:23:59,679 posts about how they use AI tools in 756 00:23:59,679 --> 00:24:01,600 Honeycomb. And one of her key points is 757 00:24:01,600 --> 00:24:04,080 that non-deterministic systems like AI 758 00:24:04,080 --> 00:24:05,919 require more engineering discipline 759 00:24:05,919 --> 00:24:08,559 rather than less. If we look at AI as a 760 00:24:08,559 --> 00:24:10,720 way to apply scale, it's critical that 761 00:24:10,720 --> 00:24:12,799 we're very deliberate with what we scale 762 00:24:12,799 --> 00:24:14,559 with it and that we do the necessary 763 00:24:14,559 --> 00:24:16,159 verification that we've actually scaled 764 00:24:16,159 --> 00:24:18,320 what we set out to scale. Cuz if we 765 00:24:18,320 --> 00:24:20,000 don't, then we run the real risk of just 766 00:24:20,000 --> 00:24:21,600 scaling the dysfunction that's already 767 00:24:21,600 --> 00:24:23,679 present in our workplaces. This one kind 768 00:24:23,679 --> 00:24:25,279 of hints at one of my biggest 769 00:24:25,279 --> 00:24:28,320 frustrations with LM at all. it's as a 770 00:24:28,320 --> 00:24:31,840 tool. It's so easy to use them badly and 771 00:24:31,840 --> 00:24:33,600 using them well is not just difficult to 772 00:24:33,600 --> 00:24:35,360 do, but it's even kind of difficult to 773 00:24:35,360 --> 00:24:37,279 pin down exactly what that looks like. 774 00:24:37,279 --> 00:24:39,440 And they're also expensive to use, not 775 00:24:39,440 --> 00:24:42,400 just in token costs, but in terms of all 776 00:24:42,400 --> 00:24:44,240 the costs that get externalized on the 777 00:24:44,240 --> 00:24:45,760 workers using them, a lot of which we 778 00:24:45,760 --> 00:24:47,279 don't really understand yet. And then 779 00:24:47,279 --> 00:24:48,720 there's the stuff we said was out of 780 00:24:48,720 --> 00:24:50,000 scope earlier, which is the costs that 781 00:24:50,000 --> 00:24:52,720 get externalized onto everyone else. 782 00:24:52,720 --> 00:24:55,279 So, uh, CEO of Microsoft recently 783 00:24:55,279 --> 00:24:57,279 remarked that the the tech sector risks 784 00:24:57,279 --> 00:24:59,200 losing the social license to consume 785 00:24:59,200 --> 00:25:01,840 resources if generative AI doesn't start 786 00:25:01,840 --> 00:25:04,240 showing some kind of commensurate social 787 00:25:04,240 --> 00:25:06,000 benefit. I mean, I honestly reckon 788 00:25:06,000 --> 00:25:07,600 Oldmate's a bit behind the times on this 789 00:25:07,600 --> 00:25:10,559 one. Uh, even before he said this, 790 00:25:10,559 --> 00:25:12,159 approval polls about generative AI and 791 00:25:12,159 --> 00:25:14,880 general society um have been in the 792 00:25:14,880 --> 00:25:17,120 gutter. We saw some of those in Cam's 793 00:25:17,120 --> 00:25:18,960 slides this morning. Some of the ones 794 00:25:18,960 --> 00:25:20,159 that I've seen were actually a bit worse 795 00:25:20,159 --> 00:25:22,240 than that. Um, and then you got things 796 00:25:22,240 --> 00:25:24,080 like on the hotel lobby TV yesterday 797 00:25:24,080 --> 00:25:26,000 morning just before the opening, the 798 00:25:26,000 --> 00:25:27,840 premier of this very state was exelling 799 00:25:27,840 --> 00:25:30,559 the virtues of not just AI, but the 800 00:25:30,559 --> 00:25:33,440 colon gas that will power it. And uh, 801 00:25:33,440 --> 00:25:34,960 I'm just going to bring back this I'm 802 00:25:34,960 --> 00:25:37,039 going to scroll my slides in the right 803 00:25:37,039 --> 00:25:39,360 direction and bring back this slide. And 804 00:25:39,360 --> 00:25:41,520 like in Cam's keynote just now, one of 805 00:25:41,520 --> 00:25:43,520 the the feelings I had in the back of my 806 00:25:43,520 --> 00:25:45,520 mind is just how much discomfort there 807 00:25:45,520 --> 00:25:47,760 is in the feeling of being one of the 808 00:25:47,760 --> 00:25:50,559 soldiers of the apocalypse almost. And 809 00:25:50,559 --> 00:25:52,080 the fact is on top of everything else 810 00:25:52,080 --> 00:25:54,080 we've talked about, some tech workers 811 00:25:54,080 --> 00:25:56,240 are going to have social environmental 812 00:25:56,240 --> 00:25:57,760 consciousnesses about, you know, 813 00:25:57,760 --> 00:26:00,000 generative AI and how it works. And that 814 00:26:00,000 --> 00:26:01,440 combines all the stuff we've talked 815 00:26:01,440 --> 00:26:02,799 about and means that the erosion of the 816 00:26:02,799 --> 00:26:05,360 social contract is happening within tech 817 00:26:05,360 --> 00:26:08,080 as well as without and just maybe not to 818 00:26:08,080 --> 00:26:11,440 the same extent. Um, in 2023, only 2% of 819 00:26:11,440 --> 00:26:13,120 developers had negative feelings towards 820 00:26:13,120 --> 00:26:16,080 AI overall. In 2025, that number is 20%. 821 00:26:16,080 --> 00:26:17,679 Um, we don't know what it's going to be 822 00:26:17,679 --> 00:26:19,840 this year. It really feels like the 823 00:26:19,840 --> 00:26:21,520 place that we're at as an industry right 824 00:26:21,520 --> 00:26:23,360 now is one where a lot of tension is 825 00:26:23,360 --> 00:26:24,960 building up. The tension between the 826 00:26:24,960 --> 00:26:27,840 tech industry and the society is 827 00:26:27,840 --> 00:26:29,840 increasing at a rapid pace, but the 828 00:26:29,840 --> 00:26:31,279 tension between tech leadership and tech 829 00:26:31,279 --> 00:26:33,039 workers is increasing too. The consent 830 00:26:33,039 --> 00:26:34,480 of the governed, if you will. 831 00:26:34,480 --> 00:26:36,320 So, where does that tension go? 832 00:26:36,320 --> 00:26:38,400 I don't know. 833 00:26:38,400 --> 00:26:40,320 I wish I had a solution to end this on, 834 00:26:40,320 --> 00:26:42,559 but I just don't. The only thing I do 835 00:26:42,559 --> 00:26:44,799 feel confident about is it doesn't go 836 00:26:44,799 --> 00:26:46,880 away by itself if we ignore it. Maybe 837 00:26:46,880 --> 00:26:48,480 next year we'll get to present a talk on 838 00:26:48,480 --> 00:26:50,320 what the solution is. Maybe one of the 839 00:26:50,320 --> 00:26:51,440 folks in the audience will get to 840 00:26:51,440 --> 00:26:53,760 present that talk instead. 841 00:26:53,760 --> 00:26:56,400 But when it comes to how you do work, uh 842 00:26:56,400 --> 00:26:57,919 the most important thing is that you 843 00:26:57,919 --> 00:26:59,360 need to be a part of the discussion as 844 00:26:59,360 --> 00:27:01,279 to how that work gets done. The fact 845 00:27:01,279 --> 00:27:03,840 that it's a dialogue is the canary in 846 00:27:03,840 --> 00:27:05,840 this particular coal mine. If that 847 00:27:05,840 --> 00:27:07,679 dialogue dies and you're only allowed to 848 00:27:07,679 --> 00:27:09,360 do things the way that you're told, then 849 00:27:09,360 --> 00:27:10,960 it's a sign that things have gone very 850 00:27:10,960 --> 00:27:12,240 very wrong. 851 00:27:12,240 --> 00:27:13,840 We take care of us. We just have to 852 00:27:13,840 --> 00:27:15,600 figure out not just how but also who 853 00:27:15,600 --> 00:27:17,279 gets to be part of we. 854 00:27:17,279 --> 00:27:20,793 And with that, thank you very much. 855 00:27:20,793 --> 00:27:22,813 [applause] 856 00:27:26,033 --> 00:27:26,799 [applause] 857 00:27:26,799 --> 00:27:30,200 We did it. 858 00:27:30,618 --> 00:27:32,638 [applause] 859 00:27:33,120 --> 00:27:34,880 Thank you for keeping up the claps until 860 00:27:34,880 --> 00:27:37,440 I made it up on stage. Daisy Benu, thank 861 00:27:37,440 --> 00:27:39,120 you so much for this wonderful 862 00:27:39,120 --> 00:27:41,440 presentation. And I have some wonderful 863 00:27:41,440 --> 00:27:43,279 Python AU Max for you. 864 00:27:43,279 --> 00:27:45,120 Thank you for having us. 865 00:27:45,120 --> 00:27:46,994 All right. 866 00:27:46,994 --> 00:27:49,014 [applause] 867 00:27:50,640 --> 00:27:53,279 Just it is it is not filled with piss 868 00:27:53,279 --> 00:27:56,240 thankfully. So that's good 869 00:27:56,240 --> 00:27:58,159 yet. 870 00:27:58,159 --> 00:27:59,760 So 871 00:27:59,760 --> 00:28:01,279 in your [snorts] own time. 872 00:28:01,279 --> 00:28:03,760 All right. Thank you all very much for 873 00:28:03,760 --> 00:28:05,600 being here and we're going to have a 874 00:28:05,600 --> 00:28:07,279 quick changeover before we're going to 875 00:28:07,279 --> 00:28:11,360 talk, I think, about launching rockets.