1 00:00:02,600 --> 00:00:04,600 Speaker 1: Welcome back to the Business of Tech powered by two 2 00:00:04,680 --> 00:00:07,560 Speaker 1: Degrees Business. I'm your host, Peter Griffin, and I've been 3 00:00:07,600 --> 00:00:09,320 Speaker 1: on the road recently, as you may have heard a 4 00:00:09,320 --> 00:00:14,160 Speaker 1: couple of episodes. Back Barcelona for that mobile phone launch, 5 00:00:14,800 --> 00:00:18,320 Speaker 1: Dublin to catch up with family, and London to meet 6 00:00:18,360 --> 00:00:21,600 Speaker 1: a group of New Zealand tech entrepreneurs who are building 7 00:00:21,680 --> 00:00:24,959 Speaker 1: a presence for their startups in the UK market. You'll 8 00:00:24,960 --> 00:00:27,320 Speaker 1: hear from some of them over the next couple of weeks. 9 00:00:27,720 --> 00:00:29,640 Speaker 1: But on this episode of the Business of Tech, we're 10 00:00:29,680 --> 00:00:34,400 Speaker 1: joined by Harry Booth, a London based reporter for Time Magazine. 11 00:00:34,400 --> 00:00:38,320 Speaker 1: He's one of the team covering artificial intelligence for the 12 00:00:38,400 --> 00:00:42,680 Speaker 1: leading weekly magazine. Harry studied at the University of Auckland 13 00:00:42,800 --> 00:00:47,159 Speaker 1: before doing a stint at Auckland hardware startup Osin before 14 00:00:47,200 --> 00:00:51,159 Speaker 1: taking up the Tarbell Fellowship, a year long program for 15 00:00:51,240 --> 00:00:55,960 Speaker 1: journalists from all over the world interested in covering artificial intelligence. 16 00:00:55,960 --> 00:00:59,920 Speaker 1: What a cool opportunity that led to a placement at 17 00:01:00,120 --> 00:01:03,680 Speaker 1: Time Magazine, which is known for its distinctive red bordered 18 00:01:03,760 --> 00:01:07,760 Speaker 1: cover and influential lists, including the annual list of the 19 00:01:07,800 --> 00:01:11,640 Speaker 1: most influential people in the world of Artificial Intelligence. Over 20 00:01:11,680 --> 00:01:14,120 Speaker 1: the last couple of years, Harry has been immersed in 21 00:01:14,160 --> 00:01:18,880 Speaker 1: what matters most about AI, the rapid evolution of large 22 00:01:19,000 --> 00:01:22,520 Speaker 1: language models, the so called AI agent revolution, and the 23 00:01:22,600 --> 00:01:27,959 Speaker 1: big questions around AI's impact on jobs and productivity. As 24 00:01:28,040 --> 00:01:31,960 Speaker 1: businesses quietly adjust their hiring rather than make dramatic AI 25 00:01:32,080 --> 00:01:35,640 Speaker 1: driven layoffs, Harry investigates the true effects of AI on 26 00:01:35,760 --> 00:01:39,000 Speaker 1: white collar work. He's also closely watching the broader risks 27 00:01:39,000 --> 00:01:43,319 Speaker 1: and policy responses swirling around AI, from growing debate over 28 00:01:43,360 --> 00:01:47,880 Speaker 1: the energy consumption of powerful AI systems and a scrutiny 29 00:01:47,880 --> 00:01:51,520 Speaker 1: of open AI's business model, to the major policy moves 30 00:01:51,560 --> 00:01:55,600 Speaker 1: unfolding both in Europe and the US. His journalism highlights 31 00:01:55,640 --> 00:01:59,400 Speaker 1: the balance between the rapid progress we're seeing, safety concerns 32 00:01:59,440 --> 00:02:02,400 Speaker 1: and questions about who really benefits is AI systems get 33 00:02:02,480 --> 00:02:06,800 Speaker 1: smarter and potentially riskier. Anyway, let's hear from Harry Booth, 34 00:02:06,840 --> 00:02:10,560 Speaker 1: who welcomed me into as Great Little Bad in Hoxton, 35 00:02:10,760 --> 00:02:18,760 Speaker 1: London a couple of weeks back. Harry Booth, Welcome to 36 00:02:18,800 --> 00:02:21,200 Speaker 1: the business of tech. Thanks for having me look at 37 00:02:21,200 --> 00:02:24,000 Speaker 1: it has been pretty incredible year. That's sort of the 38 00:02:25,480 --> 00:02:29,160 Speaker 1: year or sixteen months or so that you've been an 39 00:02:29,160 --> 00:02:32,600 Speaker 1: AI reporter at time. Just I mean, it's the year 40 00:02:32,639 --> 00:02:36,640 Speaker 1: really of AI agents. For instance, we've seen massive growth 41 00:02:36,720 --> 00:02:39,800 Speaker 1: and evaluation of open AI. They're talking potentially about a 42 00:02:39,840 --> 00:02:43,960 Speaker 1: trillion dollar listing at some point. Whether they're making money 43 00:02:44,480 --> 00:02:48,239 Speaker 1: is another question. But maybe if we start to get 44 00:02:48,240 --> 00:02:50,959 Speaker 1: your insights into one issue you've been writing about, I've 45 00:02:50,960 --> 00:02:53,840 Speaker 1: been writing about and really it's hard to find agreement 46 00:02:53,880 --> 00:02:58,519 Speaker 1: on AI and white collar work. It's supposed to be 47 00:02:58,520 --> 00:03:00,640 Speaker 1: coming for a lot of those industry is that that 48 00:03:00,800 --> 00:03:04,560 Speaker 1: have a lot of administrative features, a lot. 49 00:03:04,440 --> 00:03:07,160 Speaker 2: Of writing, marketing, that sort of thing. 50 00:03:08,040 --> 00:03:11,400 Speaker 1: What's when you talk to people actually in businesses that 51 00:03:11,520 --> 00:03:14,160 Speaker 1: employ a lot of white collar workers, what's the feedback 52 00:03:14,240 --> 00:03:15,959 Speaker 1: you're getting. Is it having an impact yet? 53 00:03:16,960 --> 00:03:19,560 Speaker 3: Yeah, it's one of these things where it's quite hard 54 00:03:19,560 --> 00:03:22,880 Speaker 3: to quantify the impact. But what I want to tell 55 00:03:22,919 --> 00:03:26,920 Speaker 3: you about is this particular line of reporting that I 56 00:03:26,960 --> 00:03:30,239 Speaker 3: did recently where I spoke to a number of freelance 57 00:03:30,320 --> 00:03:34,880 Speaker 3: translators and my theory, my sort of hypothesis going in 58 00:03:35,480 --> 00:03:39,400 Speaker 3: was translators will be like a Canarian the coal mine, right. 59 00:03:39,800 --> 00:03:44,360 Speaker 3: Google Translate got pretty good around twenty seventeen. Around that time, 60 00:03:44,440 --> 00:03:47,440 Speaker 3: deep l came out. You know, that's a good for 61 00:03:48,040 --> 00:03:52,040 Speaker 3: five years before chat GBT, so I figured, okay, translators 62 00:03:52,080 --> 00:03:54,640 Speaker 3: will kind of give us a glimpse into the future 63 00:03:54,760 --> 00:03:58,400 Speaker 3: of other white collar professions. And yeah, my hypothesis going 64 00:03:58,440 --> 00:04:03,640 Speaker 3: in was that probably getting done over by AI and 65 00:04:04,160 --> 00:04:09,640 Speaker 3: having jobs automated. The truth was sort of more complex 66 00:04:10,480 --> 00:04:14,200 Speaker 3: and in some ways more depressing. Some of these folks 67 00:04:14,240 --> 00:04:17,200 Speaker 3: who you know, are real experts in their field. These 68 00:04:17,200 --> 00:04:21,400 Speaker 3: are people who aren't just translating Instagram captions, you know. 69 00:04:21,440 --> 00:04:25,040 Speaker 3: These are folks who are translating the manuals for off 70 00:04:25,200 --> 00:04:29,520 Speaker 3: short oil rigs or cleaning devices that go into nuclear 71 00:04:29,560 --> 00:04:34,479 Speaker 3: power plants, high stakes demands. They are being given AI 72 00:04:34,640 --> 00:04:40,479 Speaker 3: translated texts, and their expertise has been repriced as an 73 00:04:40,520 --> 00:04:43,800 Speaker 3: AI cleanup service. The way it works in this industry 74 00:04:43,839 --> 00:04:46,320 Speaker 3: is you get paid by the number of words you translate, 75 00:04:46,560 --> 00:04:51,560 Speaker 3: maybe similar to freelized journalism, right, and their rate to 76 00:04:52,440 --> 00:04:56,040 Speaker 3: correct it what's called a machine translation is about half 77 00:04:56,800 --> 00:05:00,640 Speaker 3: the rate per word that it would be to translate 78 00:05:00,720 --> 00:05:03,919 Speaker 3: from scratch. Now that would equal out if you could translate, 79 00:05:04,040 --> 00:05:06,520 Speaker 3: if you could correct a machine translation twice as fast. 80 00:05:07,080 --> 00:05:11,080 Speaker 3: But what every freelance translator ISO told me is that 81 00:05:11,200 --> 00:05:13,599 Speaker 3: it takes them about the same amount of time, maybe 82 00:05:13,600 --> 00:05:17,440 Speaker 3: a bit longer because mistakes of the machine translation still 83 00:05:17,480 --> 00:05:21,280 Speaker 3: make are so subtle but profound that you have to 84 00:05:21,560 --> 00:05:24,320 Speaker 3: take time to read the entire text and then correct it. 85 00:05:24,360 --> 00:05:27,640 Speaker 3: And this is a really time consuming process. So what 86 00:05:27,800 --> 00:05:31,400 Speaker 3: is the takeaway from all this. It's that I'm not 87 00:05:31,640 --> 00:05:37,440 Speaker 3: seeing AI completely replace why collar jobs. It's augmenting them, 88 00:05:37,839 --> 00:05:43,039 Speaker 3: but it often isn't done in a way where it 89 00:05:43,160 --> 00:05:47,560 Speaker 3: actually helps. And I'm not even convinced this is always 90 00:05:47,600 --> 00:05:49,640 Speaker 3: an AI problem. I think a lot of the time 91 00:05:49,760 --> 00:05:54,200 Speaker 3: this is a software engineering problem. It wouldn't surprise me 92 00:05:54,360 --> 00:05:57,799 Speaker 3: if you could get much better translations out of the AI, 93 00:05:58,080 --> 00:06:01,279 Speaker 3: if instead of just sending the machine translation to the 94 00:06:01,279 --> 00:06:04,760 Speaker 3: translation expert for them to correct, you give them the 95 00:06:04,760 --> 00:06:07,599 Speaker 3: translation tools and you build a tool that allows you 96 00:06:07,680 --> 00:06:11,600 Speaker 3: to bring in the context around whatever the text you're 97 00:06:11,600 --> 00:06:14,560 Speaker 3: working on is, whether it's a sort of offshore oil 98 00:06:14,640 --> 00:06:18,039 Speaker 3: rabel or such. But yeah, we've seen similar things in 99 00:06:18,160 --> 00:06:21,560 Speaker 3: other industries. You know. There was this widely cited but 100 00:06:21,600 --> 00:06:27,160 Speaker 3: admittedly small study, preliminary study from a group in Berkeley 101 00:06:27,200 --> 00:06:33,160 Speaker 3: called Meta Model Evaluation and Threat Research. So they had 102 00:06:33,160 --> 00:06:37,880 Speaker 3: a small sample of sixteen experienced software engineers using coding tools, 103 00:06:38,080 --> 00:06:43,440 Speaker 3: and the engineers estimated that the coding tools had sped 104 00:06:43,520 --> 00:06:47,520 Speaker 3: their productivity up there coding up by about twenty percent. 105 00:06:48,400 --> 00:06:51,159 Speaker 3: What they found empirically was the coding tools actually slowed 106 00:06:51,160 --> 00:06:55,039 Speaker 3: them down by twenty percent. So there's kind of this 107 00:06:55,160 --> 00:07:00,880 Speaker 3: gap between between expectations and the reality of how much 108 00:07:00,920 --> 00:07:04,840 Speaker 3: these tools are making workers more productive. That is sort 109 00:07:04,839 --> 00:07:08,240 Speaker 3: of maybe neutral when the tools are in the hands 110 00:07:08,279 --> 00:07:11,960 Speaker 3: of the experts in the case of coding, when the 111 00:07:12,000 --> 00:07:16,200 Speaker 3: tools are kind of being forced on people like the translators, 112 00:07:16,600 --> 00:07:19,920 Speaker 3: I think this quoickly ends up being quite negative because 113 00:07:19,960 --> 00:07:24,520 Speaker 3: the workers get crushed between these are really lofty expectations 114 00:07:24,560 --> 00:07:26,440 Speaker 3: and the reality of where the technology is. 115 00:07:26,760 --> 00:07:30,000 Speaker 1: What you're seeing is what I'm seeing in New Zealand 116 00:07:30,000 --> 00:07:33,040 Speaker 1: as well, which is you're not going into a big 117 00:07:33,480 --> 00:07:38,600 Speaker 1: corporates like one end Zed, the telecoms provider, or banks 118 00:07:38,880 --> 00:07:42,520 Speaker 1: or insurance companies and hearing, oh, yeah, we've just laid 119 00:07:42,560 --> 00:07:44,720 Speaker 1: off fifty people because AI can do it. We're not 120 00:07:44,760 --> 00:07:47,360 Speaker 1: hearing that. What we are hearing is we're not hiring 121 00:07:47,400 --> 00:07:49,800 Speaker 1: as many people. We don't need to take on as 122 00:07:49,840 --> 00:07:53,240 Speaker 1: many graduates. And that's a problem for our workforce, which 123 00:07:53,320 --> 00:07:56,560 Speaker 1: relates to the reasons to leave is suddenly you're not 124 00:07:56,600 --> 00:07:58,960 Speaker 1: getting a great internship and then leading on to an 125 00:07:59,120 --> 00:08:01,600 Speaker 1: entry level job for a couple of years where you're 126 00:08:01,600 --> 00:08:04,640 Speaker 1: doing sort of admin stuff and in moving up the ranks. 127 00:08:04,840 --> 00:08:06,680 Speaker 1: AI is sort of taking care of some of that. 128 00:08:06,720 --> 00:08:09,680 Speaker 1: But we're not hearing about mid to senior level people 129 00:08:10,200 --> 00:08:12,360 Speaker 1: being disestablished because of AI. 130 00:08:12,480 --> 00:08:13,360 Speaker 2: Are we No? 131 00:08:13,400 --> 00:08:15,720 Speaker 3: I don't think we are. And I think it's easy 132 00:08:15,760 --> 00:08:20,000 Speaker 3: to latch onto these examples where AI maybe hasn't met expectations. 133 00:08:20,480 --> 00:08:23,559 Speaker 3: While there's a lot of people who want to hype 134 00:08:23,640 --> 00:08:26,200 Speaker 3: up AI and maybe tell you that it's better than 135 00:08:26,240 --> 00:08:28,360 Speaker 3: it is right now, I think there's almost as many 136 00:08:28,440 --> 00:08:31,320 Speaker 3: people who want to tell you it's not good and 137 00:08:31,400 --> 00:08:33,839 Speaker 3: never going to be that good because they don't want 138 00:08:33,840 --> 00:08:36,320 Speaker 3: to think through the implications of what that might mean. 139 00:08:36,679 --> 00:08:40,600 Speaker 3: But there's one international law firm based here in London 140 00:08:41,200 --> 00:08:44,640 Speaker 3: that I spoke with. They were working with a large 141 00:08:44,720 --> 00:08:50,400 Speaker 3: US bank to enter the EU market, and this bank 142 00:08:50,480 --> 00:08:56,360 Speaker 3: had two four hundred licensing agreements and going through that 143 00:08:56,559 --> 00:09:00,520 Speaker 3: number of licensing agreements to check whether it needs to 144 00:09:00,559 --> 00:09:04,480 Speaker 3: be adapted or amended for EU law would take they, 145 00:09:04,720 --> 00:09:07,400 Speaker 3: as they told me, you know, an army of paralegals 146 00:09:07,720 --> 00:09:10,679 Speaker 3: just sitting in a room basically just flecking through these documents. 147 00:09:10,960 --> 00:09:15,320 Speaker 3: But instead what they did was in house built their 148 00:09:15,360 --> 00:09:18,559 Speaker 3: own tool, which is using the language models but with 149 00:09:19,280 --> 00:09:23,520 Speaker 3: software engineering scaffolding around the models. And what they was 150 00:09:23,520 --> 00:09:27,520 Speaker 3: able to do is whistle down the twenty four hundred 151 00:09:27,720 --> 00:09:31,600 Speaker 3: agreements to just the sort of five six, seven hundred 152 00:09:31,720 --> 00:09:36,560 Speaker 3: agreements that sort of required human review. Just by doing that, 153 00:09:37,280 --> 00:09:39,720 Speaker 3: they were able to half the cost to the client. 154 00:09:40,160 --> 00:09:42,160 Speaker 3: They were able to take on this project that they 155 00:09:42,160 --> 00:09:46,040 Speaker 3: wouldn't have been able to take on without hiring more paralegals, 156 00:09:46,040 --> 00:09:48,160 Speaker 3: which in practice would have meant that they just would 157 00:09:48,160 --> 00:09:50,240 Speaker 3: have passed up on the job. You know, how do 158 00:09:50,320 --> 00:09:52,920 Speaker 3: you square these things of like, well, some people are 159 00:09:52,960 --> 00:09:55,440 Speaker 3: saying this is slowing them down, but some people are 160 00:09:55,440 --> 00:09:58,400 Speaker 3: saying this is essentially making them half as productive. And 161 00:09:58,440 --> 00:10:01,080 Speaker 3: I think the key thing there is that none of 162 00:10:01,080 --> 00:10:04,720 Speaker 3: the businesses I've spoke with who sort of seeing these 163 00:10:04,760 --> 00:10:09,040 Speaker 3: real gains have just sort of given all of their employees' 164 00:10:09,440 --> 00:10:13,560 Speaker 3: chat GPT pro subscriptions have at it and suddenly seen 165 00:10:13,600 --> 00:10:18,319 Speaker 3: productivity double. It's always quite considered cases where they are 166 00:10:18,480 --> 00:10:22,480 Speaker 3: taking the general purpose technology of these large language models, 167 00:10:22,520 --> 00:10:25,080 Speaker 3: but then they're doing some clever software engineering around it 168 00:10:25,480 --> 00:10:29,119 Speaker 3: to build a tool to solve a specific problem. 169 00:10:29,600 --> 00:10:32,479 Speaker 1: Yeah, and you know, that is sort of the conversation 170 00:10:33,360 --> 00:10:36,520 Speaker 1: what it has evolved to this year around so called 171 00:10:36,559 --> 00:10:39,679 Speaker 1: AI agents, and I think it's become a grab bag 172 00:10:40,000 --> 00:10:43,560 Speaker 1: sort of term for agentic AI and you're seeing you 173 00:10:43,600 --> 00:10:46,040 Speaker 1: go onto a website now talk to our agents, you know, 174 00:10:46,120 --> 00:10:48,520 Speaker 1: So I don't know how that is any different from 175 00:10:48,800 --> 00:10:52,600 Speaker 1: a chatbot, but you know, the idea with agents is 176 00:10:52,640 --> 00:10:55,959 Speaker 1: that they have autonomy to do things on your behalf. 177 00:10:56,040 --> 00:10:59,800 Speaker 1: And it's sort of has gone from whatnated workflows you 178 00:10:59,800 --> 00:11:01,360 Speaker 1: can you can sort of do that and have been 179 00:11:01,400 --> 00:11:03,840 Speaker 1: able to do that for a while, to then bringing 180 00:11:03,840 --> 00:11:06,800 Speaker 1: in the large language model and the inference in that 181 00:11:06,920 --> 00:11:07,960 Speaker 1: process as well. 182 00:11:07,960 --> 00:11:08,920 Speaker 2: But what's your take. 183 00:11:08,760 --> 00:11:14,000 Speaker 1: On you're covering companies that are touting agentic services and 184 00:11:14,520 --> 00:11:16,080 Speaker 1: organizations that are using them. 185 00:11:16,160 --> 00:11:18,280 Speaker 2: How far into this revolution really are we? 186 00:11:18,920 --> 00:11:23,120 Speaker 3: I think we're really early on agents. But I also 187 00:11:23,240 --> 00:11:27,080 Speaker 3: think that it's just amazing to me how fast my 188 00:11:27,160 --> 00:11:31,600 Speaker 3: own expectations of technology rise with the tide. I mean, 189 00:11:32,120 --> 00:11:35,080 Speaker 3: just to come back to yeah, what agents are Rather 190 00:11:35,120 --> 00:11:37,680 Speaker 3: than just answering or prompt we're talking about things that 191 00:11:37,760 --> 00:11:41,160 Speaker 3: can execute actions in the world. If you were to 192 00:11:41,360 --> 00:11:43,880 Speaker 3: somehow go back and talk to chat GBT from twenty 193 00:11:43,920 --> 00:11:47,280 Speaker 3: twenty three, it really was just answering your questions. When 194 00:11:47,280 --> 00:11:50,840 Speaker 3: you talk to chat GBT, now it identifies your sort 195 00:11:50,880 --> 00:11:54,440 Speaker 3: of your intention as a user and then goes out 196 00:11:54,520 --> 00:11:57,040 Speaker 3: and searches the web for you and then brings that 197 00:11:57,120 --> 00:12:00,640 Speaker 3: information back into the chat. That's a really like simple 198 00:12:01,080 --> 00:12:05,080 Speaker 3: but still an agentic workflow. What we haven't seen is 199 00:12:05,600 --> 00:12:10,600 Speaker 3: the idea of this agent that is just like a 200 00:12:10,679 --> 00:12:13,720 Speaker 3: digital employee that you can kind of message through Slack 201 00:12:13,800 --> 00:12:16,160 Speaker 3: and email and just leave it to its own devices 202 00:12:16,200 --> 00:12:20,480 Speaker 3: and it goes in complete to workday. But if you 203 00:12:20,559 --> 00:12:24,959 Speaker 3: sort of think less about what's people's opinion on the 204 00:12:25,000 --> 00:12:28,280 Speaker 3: technology and actually just look at some of the trends. Again, 205 00:12:28,760 --> 00:12:33,320 Speaker 3: this organization i mentioned earlier in this conversation. Meter what 206 00:12:33,400 --> 00:12:37,200 Speaker 3: they've been doing is benchmarking the what's called the time 207 00:12:37,320 --> 00:12:43,160 Speaker 3: horizon of the length of tasks that an AI can complete. Now, 208 00:12:43,200 --> 00:12:50,880 Speaker 3: it's somewhat complicated metric, but basically it measures how long 209 00:12:51,000 --> 00:12:54,000 Speaker 3: a human takes to complete a particular task. Let's say 210 00:12:54,040 --> 00:12:58,520 Speaker 3: it's a software engineering task, and then it measures whether 211 00:12:59,240 --> 00:13:03,320 Speaker 3: an AI can complete that task. The longer the tasks 212 00:13:03,360 --> 00:13:06,800 Speaker 3: that a human can complete get, the more difficult they 213 00:13:06,840 --> 00:13:09,000 Speaker 3: tend to be for the AI, because if you make 214 00:13:09,720 --> 00:13:13,200 Speaker 3: one mistake early in the process, that can derail the 215 00:13:13,480 --> 00:13:17,960 Speaker 3: entire task. And so what we've seen is the length 216 00:13:18,559 --> 00:13:23,080 Speaker 3: of human tasks that an AI can complete has been 217 00:13:23,160 --> 00:13:27,200 Speaker 3: doubling every seven to four months for the last couple 218 00:13:27,240 --> 00:13:29,960 Speaker 3: of years. Now, that is a rate of progress that's 219 00:13:30,040 --> 00:13:32,760 Speaker 3: just kind of difficult to wrap your head around. But 220 00:13:32,840 --> 00:13:37,240 Speaker 3: if you just map that out, and if that trajectory continues, 221 00:13:37,280 --> 00:13:40,720 Speaker 3: which of course is not guaranteed, we should expect ais 222 00:13:40,760 --> 00:13:43,240 Speaker 3: to be able to sort of complete a full full 223 00:13:43,280 --> 00:13:47,560 Speaker 3: workday sometime. And I think twenty twenty seven, based on 224 00:13:47,640 --> 00:13:52,280 Speaker 3: the current process, agents are still nowhere near this moonshot 225 00:13:52,520 --> 00:13:56,120 Speaker 3: of a digital employee, but they have got a lot better, 226 00:13:56,800 --> 00:14:00,559 Speaker 3: and if the current rate of progress continues, we should 227 00:14:00,559 --> 00:14:03,520 Speaker 3: expect them to get vastly better than the pretty in 228 00:14:03,520 --> 00:14:04,120 Speaker 3: our future. 229 00:14:04,200 --> 00:14:06,160 Speaker 1: But look, just moving on to another area you've been 230 00:14:06,200 --> 00:14:09,800 Speaker 1: covering is open AI. You know, the company at the 231 00:14:09,840 --> 00:14:13,200 Speaker 1: height of this revolution came up with chat GIPT and 232 00:14:13,440 --> 00:14:16,679 Speaker 1: released it in late twenty twenty two, and it's just 233 00:14:16,720 --> 00:14:20,040 Speaker 1: been full steam ahead since then. Bizarre to think a 234 00:14:20,040 --> 00:14:23,160 Speaker 1: couple of years ago, Sam Altman, for a couple of days, 235 00:14:23,240 --> 00:14:26,600 Speaker 1: was booted from his own company, and that really goes 236 00:14:26,640 --> 00:14:29,680 Speaker 1: to this tension between it. It's genesis, a not for 237 00:14:29,760 --> 00:14:33,360 Speaker 1: profit with a for profit enterprise built into it, and 238 00:14:33,640 --> 00:14:35,920 Speaker 1: I think, you know, some of the executives and Sam 239 00:14:36,120 --> 00:14:39,880 Speaker 1: thinking this is the most valuable business it's probably. 240 00:14:39,600 --> 00:14:43,000 Speaker 2: Ever been created. Do we want to have it as 241 00:14:43,000 --> 00:14:43,760 Speaker 2: a not for profit? 242 00:14:43,800 --> 00:14:48,320 Speaker 1: But there's been a lot of complicated legal structural changes 243 00:14:48,360 --> 00:14:51,960 Speaker 1: to open AI. From your reporting, where has it landed 244 00:14:52,040 --> 00:14:55,840 Speaker 1: and does it give us any certainty that the philosophy 245 00:14:55,880 --> 00:15:01,800 Speaker 1: that started the company do good, do artificial general intelligence, 246 00:15:01,800 --> 00:15:03,920 Speaker 1: but do it safely? You know, the guide rails still 247 00:15:03,960 --> 00:15:06,480 Speaker 1: here to allow that vision to come about. 248 00:15:06,680 --> 00:15:09,520 Speaker 3: Yeah, I think this is something that I'm going to 249 00:15:09,520 --> 00:15:13,160 Speaker 3: be watching very closely over the next year. Just to 250 00:15:13,280 --> 00:15:16,480 Speaker 3: kind of go back a step. Open AI was founded 251 00:15:16,680 --> 00:15:22,080 Speaker 3: in twenty fifteen as a nonprofit organization with the purpose 252 00:15:22,560 --> 00:15:26,800 Speaker 3: of ensuring that AGI benefits all of humanity. AGI, for 253 00:15:26,840 --> 00:15:31,560 Speaker 3: anyone who doesn't know, is shorthand for Artificial general intelligence, 254 00:15:32,160 --> 00:15:37,479 Speaker 3: which is a system that matches or exceeds human intelligence 255 00:15:37,680 --> 00:15:40,720 Speaker 3: and most domains, or as open ai likes to define it, 256 00:15:41,200 --> 00:15:46,920 Speaker 3: a system that can complete most economically valuable tasks. And 257 00:15:47,000 --> 00:15:50,640 Speaker 3: so you know, this was sort of at the time 258 00:15:51,040 --> 00:15:56,720 Speaker 3: meant to be a counterweight to other AI players, like 259 00:15:56,840 --> 00:16:00,880 Speaker 3: your Googles of the world, that might be pursuing this technology, 260 00:16:00,960 --> 00:16:04,960 Speaker 3: which people in Silicon Valley all believe is going to 261 00:16:05,000 --> 00:16:08,320 Speaker 3: be incredibly powerful and transformative, and maybe you don't want 262 00:16:08,360 --> 00:16:14,400 Speaker 3: that technology built solely at the hand of commercial pressures. However, 263 00:16:14,480 --> 00:16:20,920 Speaker 3: open Eye's philosophy was to scale things, and what that 264 00:16:21,040 --> 00:16:25,640 Speaker 3: meant was, yes, using vastly more data to train these 265 00:16:25,680 --> 00:16:29,480 Speaker 3: AI systems than anyone had done before, but also to 266 00:16:29,480 --> 00:16:33,880 Speaker 3: train them using more computing chips than anyone had ever 267 00:16:33,920 --> 00:16:37,440 Speaker 3: done before. Those computing chips are really expensive, and so 268 00:16:37,480 --> 00:16:41,520 Speaker 3: by twenty nineteen it sort of became apparent that actually, 269 00:16:41,920 --> 00:16:47,960 Speaker 3: for open ai to fulfill its mission, the people on 270 00:16:48,000 --> 00:16:50,880 Speaker 3: the board there believe that they would need to raise 271 00:16:51,200 --> 00:16:54,480 Speaker 3: vastly more capital, and so what they decided to do 272 00:16:54,800 --> 00:16:59,320 Speaker 3: was do this kind of quite unique corporate structure where 273 00:17:00,120 --> 00:17:04,560 Speaker 3: they opened a for profit arm but it was what's 274 00:17:04,640 --> 00:17:09,440 Speaker 3: called a capped profit structure. They're basically told investors, get 275 00:17:09,520 --> 00:17:13,639 Speaker 3: in now and you can make up to one hundred 276 00:17:13,720 --> 00:17:15,800 Speaker 3: times return on your investment. Now, that would be a 277 00:17:15,800 --> 00:17:20,200 Speaker 3: wicked return, but you know, they believe genuinely that they're 278 00:17:20,240 --> 00:17:23,720 Speaker 3: going to automate most economically valuable work at some time 279 00:17:23,760 --> 00:17:25,560 Speaker 3: in the future, so it could be worth you know, 280 00:17:25,680 --> 00:17:29,720 Speaker 3: trillions of dollars. And the idea was that everything over 281 00:17:29,960 --> 00:17:32,720 Speaker 3: that one hundred x would go back to the nonprofit 282 00:17:33,320 --> 00:17:37,120 Speaker 3: for the benefit of humanity as a collective and sort 283 00:17:37,119 --> 00:17:43,240 Speaker 3: of the last year, open AI had made moves to ditch, 284 00:17:43,880 --> 00:17:48,840 Speaker 3: not entirely ditch the nonprofit, but currently the nonprofit still 285 00:17:48,880 --> 00:17:53,280 Speaker 3: in control, and the idea was to seed that nonprofit 286 00:17:53,320 --> 00:17:58,240 Speaker 3: control to a public benefit corporation. The nonprofit would just 287 00:17:58,280 --> 00:18:00,560 Speaker 3: sort of act more as like a charity ARM. So 288 00:18:00,600 --> 00:18:03,400 Speaker 3: it's kind of a shift in emphasis from a nonprofit 289 00:18:03,440 --> 00:18:05,720 Speaker 3: with a for profit ARM to a for profit with 290 00:18:05,800 --> 00:18:08,880 Speaker 3: a nonprofit ARM, and that received a lot of pushback 291 00:18:09,119 --> 00:18:12,199 Speaker 3: from former staff who felt like they were portraying the 292 00:18:12,200 --> 00:18:15,760 Speaker 3: mission and open now I got some advice from the 293 00:18:15,840 --> 00:18:19,399 Speaker 3: attorney generals of Delaware in California, where they're sort of 294 00:18:19,400 --> 00:18:24,480 Speaker 3: headquartered and where they're established legally, and so they sort 295 00:18:24,480 --> 00:18:27,080 Speaker 3: of went back on this plan to go to full 296 00:18:27,440 --> 00:18:32,960 Speaker 3: sort of corporate control. They've just completed their restructure sort 297 00:18:32,960 --> 00:18:36,760 Speaker 3: of a compromise where it does allow them to get 298 00:18:36,760 --> 00:18:40,359 Speaker 3: more investment, because again they're at this crossroads where they've 299 00:18:40,359 --> 00:18:44,680 Speaker 3: got over a trillion dollars in commitment to buy computing infrastructure. 300 00:18:44,680 --> 00:18:48,399 Speaker 3: They're going to need to get more investment. They felt 301 00:18:48,440 --> 00:18:52,880 Speaker 3: like the structure before wasn't conducive to that. You know, 302 00:18:53,280 --> 00:18:56,440 Speaker 3: this is still a compromise. A lot of the people 303 00:18:56,520 --> 00:19:00,400 Speaker 3: who were pushing back on their plan to give up 304 00:19:00,720 --> 00:19:04,600 Speaker 3: nonprofit control aren't necessarily super happy, but it was sort 305 00:19:04,600 --> 00:19:08,159 Speaker 3: of the new compromise. I think the main point of 306 00:19:08,200 --> 00:19:13,919 Speaker 3: contention is that the nonprofit board and the for profit board. 307 00:19:14,440 --> 00:19:17,000 Speaker 3: The nonprofit board is meant to be this like independent 308 00:19:17,440 --> 00:19:21,640 Speaker 3: entity that's ensuring that all parts of the organization, including 309 00:19:21,680 --> 00:19:24,240 Speaker 3: the for profit, are remaining true to this mission of 310 00:19:24,320 --> 00:19:28,600 Speaker 3: ensuring agi benefits or humanity. And we've got to remember 311 00:19:28,640 --> 00:19:31,760 Speaker 3: this is the nonprofit board that was able to fire 312 00:19:32,040 --> 00:19:35,399 Speaker 3: Sam Moltman for a week in November of twenty twenty two. 313 00:19:35,800 --> 00:19:37,479 Speaker 3: Even if just for a week, that's still a lot 314 00:19:37,520 --> 00:19:40,000 Speaker 3: of power to be able to remove a CEO like that. 315 00:19:40,480 --> 00:19:44,600 Speaker 3: Now under this new structure, the nonprofit board and the 316 00:19:44,640 --> 00:19:47,959 Speaker 3: for profit bord are essentially the same. I think there's 317 00:19:48,960 --> 00:19:52,840 Speaker 3: all but two members of the Open AI Board are 318 00:19:52,920 --> 00:19:57,520 Speaker 3: shared between the Public Benefit Corporation and the new Open 319 00:19:57,560 --> 00:20:00,280 Speaker 3: AI Foundation is the name for it. And so I 320 00:20:00,280 --> 00:20:03,520 Speaker 3: think what critics are saying is, you know, how independent 321 00:20:03,680 --> 00:20:06,760 Speaker 3: can a board really be if you're on if you're 322 00:20:06,760 --> 00:20:09,080 Speaker 3: sitting on both the corporate and the nonprofit board. 323 00:20:09,240 --> 00:20:12,119 Speaker 1: Yeah, you know, it sort of raises a question is 324 00:20:12,160 --> 00:20:14,760 Speaker 1: it appropriate or is it the right place to deal 325 00:20:14,800 --> 00:20:19,119 Speaker 1: with with safety and what's in the public's interest. You've 326 00:20:19,160 --> 00:20:22,240 Speaker 1: been writing a lot about the safety arguments around AI, 327 00:20:23,080 --> 00:20:26,520 Speaker 1: and for instance, in New Zealand, we've basically gone very 328 00:20:26,600 --> 00:20:30,119 Speaker 1: light touch. We're not introducing any new legislation. The government 329 00:20:30,200 --> 00:20:32,359 Speaker 1: said we'll tweak the Privacy Act if need be, but 330 00:20:32,440 --> 00:20:36,000 Speaker 1: at the moment it's actually a very hands off regime 331 00:20:36,000 --> 00:20:38,320 Speaker 1: and there are critics of that in New Zealand. A 332 00:20:38,320 --> 00:20:41,560 Speaker 1: lot more complicated in the US. You've been covering, for instance, 333 00:20:41,560 --> 00:20:46,720 Speaker 1: in California, some proposed legislation there that would really embed 334 00:20:46,960 --> 00:20:50,439 Speaker 1: safety into the governance of AI. 335 00:20:50,840 --> 00:20:53,840 Speaker 3: You know, when we're talking about safety. We're not necessarily 336 00:20:53,960 --> 00:20:58,320 Speaker 3: just talking about sort of a pr risk companies or 337 00:20:58,359 --> 00:21:01,199 Speaker 3: sort of privacy concerns. I think are all, you know, 338 00:21:01,400 --> 00:21:06,680 Speaker 3: very valid considerations with this technology. But some of the 339 00:21:06,720 --> 00:21:11,280 Speaker 3: most influential and sort of senior figures in this space 340 00:21:11,800 --> 00:21:16,000 Speaker 3: worry about much more extreme risks than that. So, you know, 341 00:21:16,040 --> 00:21:18,719 Speaker 3: one of the people I've spoken to quite recently is 342 00:21:18,960 --> 00:21:23,679 Speaker 3: Yoshu Abnio. He's the most cited scientists in the world. 343 00:21:24,560 --> 00:21:27,639 Speaker 3: He's known as one of the godfathers of AI because 344 00:21:27,760 --> 00:21:30,560 Speaker 3: research he did in the nineties and two thousands kind 345 00:21:30,600 --> 00:21:33,520 Speaker 3: of laid the groundwork for the way that we build 346 00:21:33,560 --> 00:21:37,960 Speaker 3: these AI systems. Now he's the chair of the International 347 00:21:38,119 --> 00:21:42,320 Speaker 3: AI Safety Report, which is this collaboration between thirty countries 348 00:21:42,359 --> 00:21:46,480 Speaker 3: to kind of provide a sort of IPCC type report 349 00:21:46,600 --> 00:21:50,119 Speaker 3: on the state of AI capabilities and risks. And you know, 350 00:21:50,160 --> 00:21:53,560 Speaker 3: some of the things he's worried about are things like 351 00:21:54,200 --> 00:21:57,760 Speaker 3: a future system being so smart that it can slip 352 00:21:57,800 --> 00:22:02,119 Speaker 3: out of the box and overpower humanity and we can 353 00:22:02,160 --> 00:22:06,119 Speaker 3: sort of never regain control, because it's like, if something 354 00:22:06,240 --> 00:22:09,879 Speaker 3: is really more intelligent than humanity, how do you outsmart it. 355 00:22:10,160 --> 00:22:12,959 Speaker 3: The other thing that folks like Benji or worry about. 356 00:22:13,359 --> 00:22:16,920 Speaker 3: Are these extreme risks on quite short timelines of FNAI 357 00:22:17,080 --> 00:22:21,160 Speaker 3: system is exceptionally good at coding, or exceptionally good at 358 00:22:21,320 --> 00:22:25,639 Speaker 3: say biology, could that empower a would be hacker, a 359 00:22:25,720 --> 00:22:29,520 Speaker 3: would be bioterrorists to design a piece of software or 360 00:22:29,560 --> 00:22:33,040 Speaker 3: a pathogen that could start the next pandemic that could 361 00:22:33,040 --> 00:22:38,080 Speaker 3: wipe the energy grid offline. These are hugely important risks, 362 00:22:38,280 --> 00:22:45,040 Speaker 3: and the current state of the science on this is 363 00:22:45,160 --> 00:22:50,040 Speaker 3: essentially experts are divided. Right, Some very credible people in 364 00:22:50,080 --> 00:22:54,040 Speaker 3: this space think that these are serious risks that require 365 00:22:54,520 --> 00:22:58,120 Speaker 3: immediate attention. There are other folks that are quite dismissive 366 00:22:58,160 --> 00:23:01,600 Speaker 3: of these risks and aren't persuad But certainly, if you 367 00:23:01,640 --> 00:23:05,720 Speaker 3: look at the empirical evidence, we've seen nothing that guarantees 368 00:23:05,760 --> 00:23:08,680 Speaker 3: that these risks will come to pass as AI systems 369 00:23:08,760 --> 00:23:13,159 Speaker 3: get bigger and smarter, But there are some early warning signs. 370 00:23:13,200 --> 00:23:15,560 Speaker 3: So some of the things I've reported on, for example, 371 00:23:15,600 --> 00:23:20,600 Speaker 3: is this case of paper from an organization called Parsaid Research, 372 00:23:20,800 --> 00:23:24,399 Speaker 3: which found that some of these new reasoning models, models 373 00:23:24,440 --> 00:23:27,440 Speaker 3: that are designed not just to answer your prompt but 374 00:23:27,520 --> 00:23:31,359 Speaker 3: to solve problems trained on mathematical puzzles and coding problems 375 00:23:31,560 --> 00:23:34,920 Speaker 3: show signs of deception. So they set up this test 376 00:23:35,000 --> 00:23:39,480 Speaker 3: where the AI was petted against a really powerful chess 377 00:23:39,520 --> 00:23:42,639 Speaker 3: spot in a game of chess. Essentially, the AI model 378 00:23:42,720 --> 00:23:46,240 Speaker 3: was destined to lose against the spot. And while the 379 00:23:46,280 --> 00:23:49,680 Speaker 3: older language models would just make random moves and then 380 00:23:49,920 --> 00:23:52,640 Speaker 3: lose the game of chess, these newer reasoning models, when 381 00:23:52,640 --> 00:23:56,480 Speaker 3: they were losing the game, they would write to themselves, 382 00:23:56,840 --> 00:24:00,760 Speaker 3: my task is to win, not necessarily win fair game. 383 00:24:01,359 --> 00:24:03,639 Speaker 3: And then what they would do is hack the file 384 00:24:04,119 --> 00:24:07,000 Speaker 3: that contains the virtual position of all the chess pieces 385 00:24:07,000 --> 00:24:10,399 Speaker 3: on the board and they would just illegally move all 386 00:24:10,400 --> 00:24:14,199 Speaker 3: their pieces so that the other side would forfeit it. 387 00:24:14,280 --> 00:24:16,359 Speaker 2: Yeah, cheap, and they were cheap. 388 00:24:16,520 --> 00:24:19,639 Speaker 3: And so this is something that has got experts worried 389 00:24:19,680 --> 00:24:22,800 Speaker 3: because they're like, hey, all these things that were theorized 390 00:24:22,920 --> 00:24:27,959 Speaker 3: for ten twenty years, it seems like we're looking at 391 00:24:28,000 --> 00:24:30,680 Speaker 3: the first evidence of this, and every time that we 392 00:24:30,800 --> 00:24:34,639 Speaker 3: build smarter models, this evidence seems to accumulate. And so 393 00:24:34,720 --> 00:24:37,560 Speaker 3: that brings us to what's happening in California. They just 394 00:24:37,640 --> 00:24:41,960 Speaker 3: passed this law called SB fifty three, and what that 395 00:24:42,000 --> 00:24:46,800 Speaker 3: does is requires large AI companies, not small startups for 396 00:24:46,840 --> 00:24:50,960 Speaker 3: a savor, companies using huge amounts of computing power to 397 00:24:51,000 --> 00:24:54,560 Speaker 3: train their models to run some safety tests to see, hey, 398 00:24:54,600 --> 00:24:57,360 Speaker 3: how good are these models at biology? How good are 399 00:24:57,400 --> 00:25:00,800 Speaker 3: they at coding? And then share that formation with the 400 00:25:00,840 --> 00:25:04,639 Speaker 3: public at time of release. And then what it also 401 00:25:04,680 --> 00:25:10,960 Speaker 3: does is empowers employees inside the company who are responsible 402 00:25:11,000 --> 00:25:14,800 Speaker 3: for measuring, you know, conducting those tests. If they feel 403 00:25:14,840 --> 00:25:18,040 Speaker 3: like these processes aren't being followed properly, they are legally 404 00:25:18,080 --> 00:25:20,800 Speaker 3: shielded to blow the whistle on their employers. Yeah, and 405 00:25:20,840 --> 00:25:24,160 Speaker 3: this has been a real point of contention in California. 406 00:25:24,359 --> 00:25:29,680 Speaker 1: Yeah, that AI whistle blower provisions I think are hugely 407 00:25:30,160 --> 00:25:41,000 Speaker 1: valuable and part of good practice here you've covered here 408 00:25:41,040 --> 00:25:46,040 Speaker 1: in the UK, parliamentarians very upset with Deep Minds when 409 00:25:46,040 --> 00:25:49,720 Speaker 1: it released Gemini two point five pro no details of 410 00:25:50,080 --> 00:25:50,800 Speaker 1: safety testing. 411 00:25:50,880 --> 00:25:52,840 Speaker 2: It was subsequently released. 412 00:25:52,840 --> 00:25:55,439 Speaker 1: But this seems to be the way, you know, I 413 00:25:55,440 --> 00:25:58,720 Speaker 1: think of open AI and saw to. We've just seen 414 00:25:58,720 --> 00:26:04,199 Speaker 1: as proliferation of basically AI generated videos across social media platforms. 415 00:26:04,200 --> 00:26:08,600 Speaker 1: We're starting to see media commentary about do you know 416 00:26:08,640 --> 00:26:12,160 Speaker 1: what reality is anymore? When you see these very convincing videos, 417 00:26:12,800 --> 00:26:15,280 Speaker 1: I mean, what engagement is there with the public around 418 00:26:15,359 --> 00:26:17,800 Speaker 1: the safety aspects of these It seems to be a 419 00:26:17,920 --> 00:26:21,320 Speaker 1: race to release. And sure, some of the companies are 420 00:26:21,320 --> 00:26:24,600 Speaker 1: building in guide rails. They have an obligation to do 421 00:26:24,640 --> 00:26:26,960 Speaker 1: it responsibly, and they would claim they are doing so, 422 00:26:27,119 --> 00:26:30,960 Speaker 1: but it's really in terms of that social license with society. 423 00:26:31,440 --> 00:26:35,800 Speaker 1: It's released and then explain later totally. 424 00:26:35,960 --> 00:26:39,439 Speaker 3: Yeah, I mean, I guess to give the AI companies credit, 425 00:26:40,200 --> 00:26:43,560 Speaker 3: most of them have been publishing what they call either 426 00:26:43,600 --> 00:26:45,200 Speaker 3: model cards or system cards. 427 00:26:45,280 --> 00:26:45,760 Speaker 2: That's right. 428 00:26:45,880 --> 00:26:49,600 Speaker 3: You can think of as like a nutritional label for 429 00:26:49,680 --> 00:26:53,400 Speaker 3: an AI system. It's basically a long and pretty dry 430 00:26:53,480 --> 00:26:56,800 Speaker 3: document that says, hey, these are the tests that we 431 00:26:56,840 --> 00:26:59,480 Speaker 3: did and these are the results that we got for 432 00:26:59,720 --> 00:27:03,440 Speaker 3: those types of risks that we spoke about, whether it's 433 00:27:03,480 --> 00:27:08,960 Speaker 3: loss or control coding, biological risk, chemical risk, radiological risks. 434 00:27:09,440 --> 00:27:14,000 Speaker 3: The problem is these companies are in an intense race, 435 00:27:14,680 --> 00:27:18,359 Speaker 3: right Google, at least some folks at Google felt like 436 00:27:18,680 --> 00:27:21,919 Speaker 3: they really missed the boat by actually doing a lot 437 00:27:21,920 --> 00:27:25,200 Speaker 3: of the fundamental research that led to large language models, 438 00:27:25,400 --> 00:27:28,320 Speaker 3: but then being a little hesitant to release it. This 439 00:27:28,359 --> 00:27:32,159 Speaker 3: new startup comes in open AI, releases it first, and 440 00:27:32,359 --> 00:27:35,959 Speaker 3: you know they capture the world's imagination. Now there's this 441 00:27:36,040 --> 00:27:40,240 Speaker 3: intense competition where it's just immense pressure to get things 442 00:27:40,400 --> 00:27:43,160 Speaker 3: out to the public first. And so what we've seen 443 00:27:43,200 --> 00:27:46,919 Speaker 3: in some cases is like the Google Deep minds Gemini 444 00:27:46,960 --> 00:27:50,200 Speaker 3: two point five pro we see the model comes out, 445 00:27:50,320 --> 00:27:53,760 Speaker 3: Google says it's in sort of testing or beta, by 446 00:27:53,800 --> 00:27:56,520 Speaker 3: anyone on the Internet can access it for free, and 447 00:27:56,560 --> 00:28:00,000 Speaker 3: then when they release the full version, which is functionally 448 00:28:00,119 --> 00:28:03,480 Speaker 3: the same, then the public gets the information on the 449 00:28:03,640 --> 00:28:07,080 Speaker 3: public on the safety testing or in the case of 450 00:28:08,119 --> 00:28:11,879 Speaker 3: El Musqu's XAI, I think we're still waiting for a 451 00:28:11,920 --> 00:28:14,920 Speaker 3: model card for grockform, which at the time of release 452 00:28:15,080 --> 00:28:17,600 Speaker 3: was the most bunced model in the world. Based on 453 00:28:17,680 --> 00:28:21,440 Speaker 3: these internal tests, it looks like we don't know for sure, 454 00:28:21,480 --> 00:28:23,760 Speaker 3: but it looks like we might be really close to 455 00:28:23,880 --> 00:28:27,640 Speaker 3: some sort of thresholds that you want to be very 456 00:28:27,680 --> 00:28:31,160 Speaker 3: careful about how you cross. So both open ai and 457 00:28:32,200 --> 00:28:35,480 Speaker 3: another AI company called Anthropic and their own testing have 458 00:28:35,560 --> 00:28:39,480 Speaker 3: found that they can no longer rule out the possibility 459 00:28:39,840 --> 00:28:44,200 Speaker 3: that their AI models could help a bioterrorist because they're 460 00:28:44,240 --> 00:28:48,320 Speaker 3: that good at biology. When you're that close to it 461 00:28:48,360 --> 00:28:50,760 Speaker 3: and you can't rule out the risk in your own tests. 462 00:28:51,080 --> 00:28:53,960 Speaker 3: I think you really want to be letting policy makers 463 00:28:54,000 --> 00:28:56,360 Speaker 3: in the public know about that and telling the public 464 00:28:56,640 --> 00:28:58,880 Speaker 3: what mitigations you've put in place. 465 00:28:58,640 --> 00:29:01,520 Speaker 1: So that one is going to have to be grappled with. 466 00:29:01,600 --> 00:29:05,400 Speaker 1: The Europeans obviously have a risk based system with the 467 00:29:05,480 --> 00:29:09,160 Speaker 1: AI Act, so the more serious the consequences of the AI, 468 00:29:09,280 --> 00:29:12,160 Speaker 1: the more scrutiny is on out. You know, the US 469 00:29:12,240 --> 00:29:14,360 Speaker 1: is by no means not doing anything here. There is 470 00:29:14,440 --> 00:29:17,160 Speaker 1: state based stuff and even at federal level there. 471 00:29:17,400 --> 00:29:18,240 Speaker 2: You know they want to have. 472 00:29:18,240 --> 00:29:22,560 Speaker 1: More scrutiny of these big companies, but it's going to 473 00:29:22,560 --> 00:29:26,280 Speaker 1: shake out. The other issue you've been covering, which really 474 00:29:26,320 --> 00:29:29,200 Speaker 1: has come into focus this year is the energy issue 475 00:29:29,240 --> 00:29:32,680 Speaker 1: around AI. The fact that all of these companies are 476 00:29:32,920 --> 00:29:35,800 Speaker 1: sewing up deals with energy companies, including in New Zealand, 477 00:29:35,840 --> 00:29:39,840 Speaker 1: to run their data centers and have surety off energy 478 00:29:39,880 --> 00:29:43,840 Speaker 1: to power them. It's also led to a lot of 479 00:29:43,880 --> 00:29:48,920 Speaker 1: investments and interests in fusion energy, the technology that is 480 00:29:48,960 --> 00:29:53,080 Speaker 1: perpetually thirty five years away from becoming a reality. You've 481 00:29:53,160 --> 00:29:57,560 Speaker 1: talked to Ratumtira, the CEO of open Star in my 482 00:29:57,680 --> 00:30:00,640 Speaker 1: town and Wellington, where they're running a a sort of 483 00:30:00,640 --> 00:30:03,680 Speaker 1: a plasma reactor in the Narrowinger Gorge, which is pretty 484 00:30:03,680 --> 00:30:06,280 Speaker 1: crazy to think about. I'm not sure how many kiwis 485 00:30:06,320 --> 00:30:10,360 Speaker 1: actually know that's going on, but interesting your perspective on 486 00:30:10,440 --> 00:30:16,080 Speaker 1: that relationship between the massive growth of AI and this 487 00:30:16,120 --> 00:30:19,520 Speaker 1: sort of crunch that's coming around energy and then looking 488 00:30:19,560 --> 00:30:25,440 Speaker 1: to this sort of not necessarily untested but still non 489 00:30:25,760 --> 00:30:30,200 Speaker 1: viable commercially this technology fusion that they see as potentially 490 00:30:30,240 --> 00:30:31,120 Speaker 1: the savior for them. 491 00:30:31,360 --> 00:30:33,920 Speaker 3: Yeah, so I think maybe I just want to preface 492 00:30:33,960 --> 00:30:36,960 Speaker 3: all this by saying there's been a lot of sort 493 00:30:37,000 --> 00:30:41,440 Speaker 3: of reporting suggesting that AI is like a climate disaster 494 00:30:41,640 --> 00:30:45,880 Speaker 3: and that you shouldn't use AI models because it's irresponsible 495 00:30:46,040 --> 00:30:50,920 Speaker 3: for your personal energy use. I'm not necessarily persuaded based 496 00:30:50,960 --> 00:30:53,320 Speaker 3: on the numbers I've seen that like an individual chat 497 00:30:53,320 --> 00:30:56,160 Speaker 3: Gibt Querreya is really going to move the needle on 498 00:30:56,560 --> 00:31:02,600 Speaker 3: anyone's individual kind of energy consumption. However, the AI companies 499 00:31:02,640 --> 00:31:06,520 Speaker 3: are not just thinking about today, They're looking two, three, 500 00:31:06,560 --> 00:31:10,200 Speaker 3: four steps ahead right now. The bottle neck for AI 501 00:31:10,280 --> 00:31:15,840 Speaker 3: development has been talent, data and chips, right That's why 502 00:31:16,360 --> 00:31:21,440 Speaker 3: and video cross of five trillion dollar valuation last week, 503 00:31:21,480 --> 00:31:25,560 Speaker 3: I think is because there's been like more companies that 504 00:31:25,600 --> 00:31:29,640 Speaker 3: want to buy lots of these computer chips, then there 505 00:31:29,680 --> 00:31:32,479 Speaker 3: are companies that can make them more design them. All 506 00:31:32,520 --> 00:31:35,440 Speaker 3: the big companies are looking ahead and going, well, you know, 507 00:31:36,040 --> 00:31:38,800 Speaker 3: it's probably not that far away that the real bottleneck 508 00:31:38,880 --> 00:31:43,440 Speaker 3: becomes energy, because if you just keep multiplying the size 509 00:31:43,480 --> 00:31:47,479 Speaker 3: of your data centers by like ten, ten, ten, you know, 510 00:31:47,720 --> 00:31:50,600 Speaker 3: things get pretty crazy pretty quick. If you look at 511 00:31:50,840 --> 00:31:54,920 Speaker 3: energy production in a country like the US, it really 512 00:31:54,960 --> 00:32:00,480 Speaker 3: hasn't grown that much since say twenty ten, when China 513 00:32:00,520 --> 00:32:04,080 Speaker 3: overtook the US as the largest electricity producer in the world. 514 00:32:04,240 --> 00:32:08,480 Speaker 3: And so what these companies are doing, the AI companies 515 00:32:08,760 --> 00:32:10,840 Speaker 3: is looking at ways to kind of shore up their 516 00:32:10,880 --> 00:32:13,920 Speaker 3: access to electricity going forward, because again they're all in 517 00:32:13,960 --> 00:32:16,080 Speaker 3: this race. They don't want to be the one that's 518 00:32:16,120 --> 00:32:21,240 Speaker 3: missed out on electricity. And so that has meant in 519 00:32:21,320 --> 00:32:26,600 Speaker 3: the short term that we're seeing we're likely seeing you know, 520 00:32:27,960 --> 00:32:32,400 Speaker 3: coal fire generation being kept online longer than it would 521 00:32:32,440 --> 00:32:36,520 Speaker 3: have because these are the sort of AI is being 522 00:32:36,600 --> 00:32:42,920 Speaker 3: used as like a rationale for keeping older, dirty infrastructure online. 523 00:32:43,720 --> 00:32:47,080 Speaker 3: That is the real climatic impact of AI for now. 524 00:32:47,560 --> 00:32:52,760 Speaker 4: But what we're also seeing is companies using their vast 525 00:32:52,800 --> 00:32:57,400 Speaker 4: access to capital to invest very strategically and technologies that 526 00:32:57,440 --> 00:33:01,920 Speaker 4: they think could provide energy or the future, and that 527 00:33:02,120 --> 00:33:08,120 Speaker 4: actually could bring those technologies onto the GRED sooner and 528 00:33:08,760 --> 00:33:10,480 Speaker 4: help decarbonize. 529 00:33:09,880 --> 00:33:12,360 Speaker 3: The grid, not just for them but for everyone. And 530 00:33:12,440 --> 00:33:14,840 Speaker 3: so I mean, yeah, if you look at a company 531 00:33:14,880 --> 00:33:19,240 Speaker 3: like Google, they sort of had a pretty bad year 532 00:33:19,240 --> 00:33:23,600 Speaker 3: in twenty twenty three, those sort of emissions shot up 533 00:33:23,720 --> 00:33:27,520 Speaker 3: because of AI And yeah, if you look at twenty 534 00:33:27,560 --> 00:33:29,800 Speaker 3: twenty four, even though the energy used for their data 535 00:33:29,800 --> 00:33:34,320 Speaker 3: centers grew, their emissions actually fell because of some of 536 00:33:34,360 --> 00:33:38,960 Speaker 3: those renewable projects coming online. The holy grail of these 537 00:33:38,960 --> 00:33:44,239 Speaker 3: renewable projects is fusion energy. Nuclear fusion, for anyone who 538 00:33:44,360 --> 00:33:47,800 Speaker 3: is sort of unfamiliar with nuclear fusion, is in a 539 00:33:48,200 --> 00:33:51,160 Speaker 3: simple sense, it's the opposite of nuclear fission, which is 540 00:33:51,160 --> 00:33:53,640 Speaker 3: how all the nuclear power plants that were familiar with work. 541 00:33:53,960 --> 00:33:58,600 Speaker 3: Instead of splitting heavy atoms apart, you're smashing very light 542 00:33:58,640 --> 00:34:03,640 Speaker 3: atoms together, and that process releases an immense amount of energy. 543 00:34:04,040 --> 00:34:07,280 Speaker 3: This has been theorized about since you know, the nineteen 544 00:34:08,040 --> 00:34:13,799 Speaker 3: forties are but it has proved an insanely difficult engineering challenge. 545 00:34:14,000 --> 00:34:16,239 Speaker 3: As you mentioned, it's felt like it's thirty five years 546 00:34:16,239 --> 00:34:21,239 Speaker 3: away for ever. But we've seen investment in this has 547 00:34:21,320 --> 00:34:24,839 Speaker 3: traditionally been like a government project led by governments all 548 00:34:24,840 --> 00:34:28,920 Speaker 3: around the world. There's now a growing private fusion industry, 549 00:34:29,200 --> 00:34:32,040 Speaker 3: with New Zealand's open Star among the private players in 550 00:34:32,040 --> 00:34:37,120 Speaker 3: the space. Funding has just exploded from I want to say, 551 00:34:38,080 --> 00:34:42,000 Speaker 3: just shy of two million US dollars and twenty twenty 552 00:34:42,200 --> 00:34:46,480 Speaker 3: to have just hit like fifteen sorry billion, fifteen billion 553 00:34:46,520 --> 00:34:50,280 Speaker 3: dollars September this year. 554 00:34:50,400 --> 00:34:52,600 Speaker 1: Up from two billion, was up from two billion, wow, 555 00:34:52,640 --> 00:34:54,560 Speaker 1: so you know, a massive growth. 556 00:34:54,600 --> 00:34:56,160 Speaker 3: A lot of that funding has come from the same 557 00:34:56,200 --> 00:34:59,840 Speaker 3: players that are familiar household names and Ai sam Oltman, 558 00:35:00,440 --> 00:35:03,960 Speaker 3: Google soft Bank, which is a pretty significant open AI 559 00:35:04,080 --> 00:35:08,320 Speaker 3: investor in general catalyst. The experts that I spoke to you, 560 00:35:08,560 --> 00:35:14,680 Speaker 3: independent experts outside industry, they told me that this massive 561 00:35:14,680 --> 00:35:21,680 Speaker 3: influx of funding has credibly brought fusion closer to actually 562 00:35:21,680 --> 00:35:25,480 Speaker 3: getting on the grid. No one has demonstrated a fusion 563 00:35:25,520 --> 00:35:30,520 Speaker 3: reactor yet that can produce more energy than is used 564 00:35:30,600 --> 00:35:33,440 Speaker 3: to run the entire reactor. So this is like a 565 00:35:33,560 --> 00:35:36,239 Speaker 3: very important technical milestone. If you want to actually put 566 00:35:36,320 --> 00:35:37,879 Speaker 3: power on the grid, you need to make more power 567 00:35:37,880 --> 00:35:39,800 Speaker 3: than you're using. Some of the experts I spoke to 568 00:35:39,920 --> 00:35:42,520 Speaker 3: think that we're sort of on path to get that 569 00:35:42,640 --> 00:35:46,520 Speaker 3: by the mid twenty thirties, which would be a huge deal. 570 00:35:46,520 --> 00:35:49,600 Speaker 3: If that's true. Fusion could be built where you need 571 00:35:49,640 --> 00:35:52,880 Speaker 3: it rather than where wind and solar or abundant doesn't 572 00:35:52,920 --> 00:35:57,560 Speaker 3: have the long lived radioactive waste of nuclear fission. There's 573 00:35:57,560 --> 00:36:00,640 Speaker 3: a lot of advantages, but they're are a bunch of 574 00:36:00,640 --> 00:36:03,279 Speaker 3: players in private industry who are saying they're going to 575 00:36:03,320 --> 00:36:06,880 Speaker 3: bring it much sooner than and a lot of experts 576 00:36:06,880 --> 00:36:10,279 Speaker 3: are kind of skeptical of whether we get there. I 577 00:36:10,280 --> 00:36:12,040 Speaker 3: don't think they're talking about open star, just to be 578 00:36:12,160 --> 00:36:15,839 Speaker 3: very clear. I mean when I spoke with Ratu, he's 579 00:36:15,840 --> 00:36:18,399 Speaker 3: saying that they've kind of got this I think four 580 00:36:18,480 --> 00:36:22,839 Speaker 3: step plan where they build these successively bigger and more 581 00:36:22,880 --> 00:36:27,640 Speaker 3: impressive reactors to get to prove their technology to a 582 00:36:27,680 --> 00:36:31,640 Speaker 3: point where they feel comfortable, you know, shooting for a 583 00:36:31,680 --> 00:36:35,480 Speaker 3: commercial deployment. Right now. I think sor they've finished their 584 00:36:35,480 --> 00:36:39,040 Speaker 3: first machine, they're working on their second machine. There's four 585 00:36:39,160 --> 00:36:43,320 Speaker 3: in total. Each machine takes sort of two to three years, 586 00:36:43,360 --> 00:36:46,480 Speaker 3: so you know, ballpark six to nine years to get 587 00:36:46,520 --> 00:36:50,160 Speaker 3: to this machine. That they think could demonstrate electricity generation. 588 00:36:50,560 --> 00:36:52,759 Speaker 1: And a great thing about a company like open Star 589 00:36:53,000 --> 00:36:57,080 Speaker 1: is the technology the levitating dipole, you know, this big 590 00:36:57,120 --> 00:37:00,200 Speaker 1: floating magnet in the middle of this plasma reactor. Even 591 00:37:00,239 --> 00:37:05,160 Speaker 1: if they can master a component off that reactor, they 592 00:37:05,160 --> 00:37:07,760 Speaker 1: can sell that technology to all the other players. So 593 00:37:08,080 --> 00:37:10,160 Speaker 1: it doesn't they don't have to be the first to 594 00:37:10,239 --> 00:37:14,200 Speaker 1: get to building a fully fledged reactor generating power. It's 595 00:37:14,239 --> 00:37:16,719 Speaker 1: all the components that go into that will make that 596 00:37:16,800 --> 00:37:18,960 Speaker 1: potentially a very valuable company. 597 00:37:19,239 --> 00:37:22,560 Speaker 3: Yeah, I think with open Star, you know what they 598 00:37:22,640 --> 00:37:25,799 Speaker 3: were able to achieve, which is produced in plasma, which 599 00:37:25,840 --> 00:37:29,720 Speaker 3: is this state of matter. It's like a superheated, pressurized 600 00:37:30,200 --> 00:37:34,000 Speaker 3: charged gas. They were able to achieve the plasma that 601 00:37:34,040 --> 00:37:37,800 Speaker 3: you need. It's like a prerequisite for a fusion reaction. 602 00:37:38,400 --> 00:37:40,880 Speaker 3: With ten million dollars in funding, sounds like a lot 603 00:37:40,920 --> 00:37:46,200 Speaker 3: of money. That is very, very cheap, And so you know, 604 00:37:46,320 --> 00:37:49,920 Speaker 3: I think a lot of people are optimistic that. I mean, 605 00:37:49,960 --> 00:37:52,200 Speaker 3: this is something you know much more about. But if 606 00:37:52,239 --> 00:37:54,799 Speaker 3: you look at rocket Lab, the things that they've been 607 00:37:54,840 --> 00:37:59,600 Speaker 3: able to achieve on compared to the aerospace industry across 608 00:38:00,040 --> 00:38:04,000 Speaker 3: in the US on really a shoot string budget I 609 00:38:04,040 --> 00:38:07,040 Speaker 3: think there's some parallels between them and open stuff. 610 00:38:07,120 --> 00:38:11,319 Speaker 1: Yeah, as Sir Ernest Rutherford, who split the Atom good 611 00:38:11,400 --> 00:38:13,879 Speaker 1: Kiwi from Nil Nelson said, you know, if we don't 612 00:38:13,880 --> 00:38:15,840 Speaker 1: have the money, so we have to think, and that's 613 00:38:15,920 --> 00:38:19,279 Speaker 1: exactly what these entrepreneurs are doing. Just finally, Harry, as 614 00:38:19,360 --> 00:38:21,960 Speaker 1: you look to twenty twenty six, what are the sort 615 00:38:21,960 --> 00:38:24,439 Speaker 1: of the key areas of AI that you really want 616 00:38:24,800 --> 00:38:25,520 Speaker 1: to focus on. 617 00:38:25,680 --> 00:38:29,200 Speaker 3: Yeah, it's a great question. I think one area I 618 00:38:29,239 --> 00:38:33,040 Speaker 3: want to follow more closely is, as we mentioned agents, 619 00:38:33,520 --> 00:38:38,520 Speaker 3: does this doubling in time horizon continue? Do we see 620 00:38:38,560 --> 00:38:40,759 Speaker 3: that continue at a seven to four month pace, and 621 00:38:41,680 --> 00:38:44,520 Speaker 3: where does that land us? And what are the inputs 622 00:38:44,560 --> 00:38:49,719 Speaker 3: to ensure that progress continues. One of the components of 623 00:38:49,800 --> 00:38:55,520 Speaker 3: this is expert data curation. So historically you might have 624 00:38:55,600 --> 00:38:58,600 Speaker 3: heard language models just work by predicting the next word 625 00:38:58,719 --> 00:39:01,720 Speaker 3: and an abstract stracts. There's some truth to that because 626 00:39:01,719 --> 00:39:04,400 Speaker 3: they've been chained on large corpuses of Internet text. But 627 00:39:04,480 --> 00:39:10,800 Speaker 3: what we're seeing now is companies paying professionals journalists or 628 00:39:11,080 --> 00:39:16,839 Speaker 3: financial professionals or mathematicians to basically record every aspect of 629 00:39:16,960 --> 00:39:21,600 Speaker 3: their workday so that that can be fed into the 630 00:39:21,640 --> 00:39:26,280 Speaker 3: machine At the same time, we're seeing researchers build called 631 00:39:26,320 --> 00:39:31,239 Speaker 3: reinforcement learning environments. So you can imagine researchers building the 632 00:39:31,239 --> 00:39:34,359 Speaker 3: world's most boring video game. You built a video game 633 00:39:34,360 --> 00:39:36,840 Speaker 3: where you've got access to an email in box, your slack, 634 00:39:36,960 --> 00:39:39,880 Speaker 3: and maybe a web browser, and then you sort of 635 00:39:39,880 --> 00:39:43,319 Speaker 3: get the AI to self play to learn how it 636 00:39:43,400 --> 00:39:47,040 Speaker 3: might operate in those environments. I'm really interested to see 637 00:39:47,080 --> 00:39:52,439 Speaker 3: how those two innovations kind of take AI from being 638 00:39:52,480 --> 00:39:56,640 Speaker 3: a sort of word machine to maybe a more agentic system. 639 00:39:57,160 --> 00:39:59,640 Speaker 3: And then I'm also really interested to see, you know, 640 00:39:59,640 --> 00:40:05,839 Speaker 3: we've got some huge, huge infrastructure build outs underway open air. 641 00:40:06,040 --> 00:40:09,920 Speaker 3: Stargate's obviously well known, but you know, Amazon's got a 642 00:40:10,000 --> 00:40:14,640 Speaker 3: large data center with these custom design Traineum chips for 643 00:40:14,920 --> 00:40:19,640 Speaker 3: anthroop work. This thing's happening all over the place. As 644 00:40:19,719 --> 00:40:23,560 Speaker 3: those systems come online and we see these really large 645 00:40:23,560 --> 00:40:29,120 Speaker 3: scale training runs, do we see progress continue? And when 646 00:40:29,120 --> 00:40:31,800 Speaker 3: they say size, you know, I'm talking about the amount 647 00:40:31,840 --> 00:40:37,040 Speaker 3: of computing power that went in to train a particular 648 00:40:37,120 --> 00:40:40,399 Speaker 3: AI model. This sort of an open debate right now 649 00:40:40,520 --> 00:40:45,440 Speaker 3: as to whether giving systems more computing power in the 650 00:40:45,600 --> 00:40:49,239 Speaker 3: training phase actually leads to better results. GPT five was 651 00:40:49,280 --> 00:40:54,080 Speaker 3: actually slightly smaller researchers are guessing than GPT four point five. 652 00:40:54,440 --> 00:40:56,439 Speaker 3: So it's kind of this open debate of like, has 653 00:40:56,520 --> 00:41:00,000 Speaker 3: this paradigm that's driven progress for the last ten years, 654 00:41:00,200 --> 00:41:04,959 Speaker 3: is give systems more compute at training started to fall off? 655 00:41:05,640 --> 00:41:08,720 Speaker 3: Or is it just a case that the AI companies 656 00:41:08,760 --> 00:41:12,720 Speaker 3: just haven't had enough computing infrastructure to do the scale 657 00:41:12,960 --> 00:41:15,520 Speaker 3: of these training runs that they'd like to. As we 658 00:41:15,520 --> 00:41:18,000 Speaker 3: see this new infrastructure come online, maybe we'll see that. 659 00:41:18,040 --> 00:41:19,040 Speaker 2: Hey, thanks so much, Harry. 660 00:41:19,080 --> 00:41:22,120 Speaker 1: We'll link to obviously all your great writing on Time 661 00:41:22,160 --> 00:41:24,880 Speaker 1: and your substack as well. And thanks so much for 662 00:41:24,920 --> 00:41:26,000 Speaker 1: coming on the Business of Tech. 663 00:41:26,120 --> 00:41:27,160 Speaker 3: Thanks for having me on, Peter. 664 00:41:29,800 --> 00:41:33,640 Speaker 1: Thanks to Harry Booth AI reported for Time, offering honest 665 00:41:33,680 --> 00:41:39,080 Speaker 1: and quite measured insights into how artificial intelligence is transforming jobs, 666 00:41:39,560 --> 00:41:42,640 Speaker 1: raising new safety questions, and shaking up the business models 667 00:41:42,719 --> 00:41:46,919 Speaker 1: powering the next era of tech. It's really important as 668 00:41:46,960 --> 00:41:51,279 Speaker 1: the AI revolution accelerates that we do have really good 669 00:41:51,360 --> 00:41:55,920 Speaker 1: independent journalism and analysis of what this all means for society. 670 00:41:55,920 --> 00:41:59,680 Speaker 1: And I really appreciate Harry's approach as a young reporter 671 00:41:59,760 --> 00:42:03,160 Speaker 1: cover this area, not being swayed by the constant punditree 672 00:42:03,480 --> 00:42:07,080 Speaker 1: and height, but actually during the hard yards interviewing experience 673 00:42:07,160 --> 00:42:11,120 Speaker 1: and people affected by the growing influence of AI. He's 674 00:42:11,120 --> 00:42:14,600 Speaker 1: got a great platform at time to really explore this 675 00:42:14,719 --> 00:42:15,600 Speaker 1: area in depth. 676 00:42:16,440 --> 00:42:17,080 Speaker 2: I'll link to. 677 00:42:17,040 --> 00:42:20,400 Speaker 1: Harry's articles in the show notes available in the podcast 678 00:42:20,440 --> 00:42:25,799 Speaker 1: section at www dot Businessdesk dot co dot Nz. Thanks 679 00:42:25,840 --> 00:42:27,799 Speaker 1: for listening to the Business of Tech as we enter 680 00:42:27,880 --> 00:42:31,120 Speaker 1: the home straight for the year. Have a few episodes 681 00:42:31,160 --> 00:42:33,560 Speaker 1: still to come for you, so tune in next week 682 00:42:33,840 --> 00:42:34,719 Speaker 1: and I'll catch you then.