WEBVTT 00:00:00.000 --> 00:00:10.640 Music. 00:00:10.823 --> 00:00:15.723 Welcome to the edition podcast. I, as ever, am your host, Charlotte Henry, 00:00:15.803 --> 00:00:18.243 and I have another fantastic guest for you. 00:00:18.463 --> 00:00:22.603 I'm joined by the author, political editor of the New European, 00:00:23.723 --> 00:00:27.563 previous accomplice of Julian Assange. 00:00:29.283 --> 00:00:33.343 You can't deny it, James Bull. I absolutely can't deny it. 00:00:33.463 --> 00:00:38.943 It will stay on my permanent record. I mean, your lives are very much diverge, 00:00:39.043 --> 00:00:41.343 but there are books that put you in there. 00:00:43.123 --> 00:00:47.423 How are you, James? I'm so thrilled to have you back. It's lovely to be here 00:00:47.423 --> 00:00:50.423 in this lovely, balmy, calm February. 00:00:51.503 --> 00:01:00.623 Yes, February, where we are in London, is as grey and as damp as one might imagine February in London. 00:01:01.123 --> 00:01:06.183 Things are a little bit more spicy over in Washington, D.C. 00:01:06.783 --> 00:01:10.243 I feel like I'm now going to have to have a new section every week called Elon 00:01:10.243 --> 00:01:14.523 Musk Watch, because every time I turn on any kind of news app, 00:01:14.823 --> 00:01:17.823 he's done something else completely mental, James. 00:01:18.123 --> 00:01:24.223 Now, as a transatlantic politics watcher and a very well-informed tech journalist, 00:01:24.903 --> 00:01:28.963 following these developments are just fascinating, aren't they, for you? 00:01:29.443 --> 00:01:34.063 I mean, they are for me, so I guess it's up for you. It's absolutely extraordinary. 00:01:34.883 --> 00:01:40.403 I'm actually having a trip to DC in a week or so and I'm sort of curious as 00:01:40.403 --> 00:01:45.783 to what it'll look like let alone what people will be feeling like but it's 00:01:45.783 --> 00:01:47.723 hard to explain Will anyone have jobs at this point? 00:01:48.043 --> 00:01:52.843 Well, it's hard to explain how far off the map we are you know, 00:01:53.443 --> 00:02:00.123 Musk doesn't have a formal role in the US government he has got access to... 00:02:01.029 --> 00:02:05.929 A lot of its most sensitive and its most important records and systems nonetheless, 00:02:06.609 --> 00:02:09.449 and he's essentially had to call the marshals to 00:02:09.449 --> 00:02:12.609 do it because some of these are so controlled that 00:02:12.609 --> 00:02:15.689 you need vetting you need sensitive access you need approval 00:02:15.689 --> 00:02:19.249 to people who actually are congressionally accountable 00:02:19.249 --> 00:02:22.089 and he's bypassed all of this and 00:02:22.089 --> 00:02:26.169 just got the access now that's kind 00:02:26.169 --> 00:02:28.929 of scary for what's he going to 00:02:28.929 --> 00:02:32.449 do with them etc he's got things like the payment database while 00:02:32.449 --> 00:02:36.309 being an active federal contractor um but 00:02:36.309 --> 00:02:39.209 because he's opened these up um wired have 00:02:39.209 --> 00:02:42.529 identified five people aged 19 to 24 00:02:42.529 --> 00:02:47.729 who suddenly got access to several of government's most sensitive databases 00:02:47.729 --> 00:02:54.389 now the level of cyber security threat this represents is off the charts you 00:02:54.389 --> 00:02:58.949 know if If these people are introducing vulnerabilities or errors or joining 00:02:58.949 --> 00:03:00.989 up systems, that's there. 00:03:01.129 --> 00:03:04.329 There's the risk of unvetted people themselves being in it. 00:03:04.509 --> 00:03:10.589 There is joining up systems that weren't joined up. And then there's putting them to improper uses. 00:03:11.029 --> 00:03:15.749 You've just explained so many of the key issues. But I obviously want to get 00:03:15.749 --> 00:03:17.329 to the media side of this. 00:03:17.769 --> 00:03:21.909 And I'm so pleased you mentioned the Wired story. And you hinted at a couple 00:03:21.909 --> 00:03:23.989 of things, which leads me to my question. 00:03:24.649 --> 00:03:28.089 Are we as journalists contextualizing and 00:03:28.089 --> 00:03:30.869 reporting this properly you described it as so far off 00:03:30.869 --> 00:03:34.449 the map and you're right but has the level 00:03:34.449 --> 00:03:39.249 of normalization already gone too far are we in the media explaining just how 00:03:39.249 --> 00:03:44.909 to use your phrase far off the map this is are we doing a good job at that i 00:03:44.909 --> 00:03:51.529 don't think there's a very good job being done at all um i think the what is 00:03:51.529 --> 00:03:54.169 getting covered quite well in most of the coverage. 00:03:54.409 --> 00:04:00.369 I think, and there's so much happening that it's almost impossible because so 00:04:00.369 --> 00:04:04.949 much, so many individual things should and would be leading the news at any point. 00:04:05.309 --> 00:04:07.629 I think the fact that most of it's happening. 00:04:08.474 --> 00:04:11.894 Very blatantly illegally is almost 00:04:11.894 --> 00:04:15.594 just being relegated down and not communicated you 00:04:15.594 --> 00:04:18.574 know the the u.s doesn't work like we do 00:04:18.574 --> 00:04:25.314 if congress um appropriates money the president has to spend it it is an order 00:04:25.314 --> 00:04:31.134 to do the thing it's not saying here is how much is available to spend and you 00:04:31.134 --> 00:04:37.814 the minister or you the secretary decides how to spend it as it is in the uk Yeah, you must spend it. 00:04:37.994 --> 00:04:41.454 We have a negative obligation to spend. You can spend up to this amount. 00:04:41.734 --> 00:04:43.674 The US has what's called a positive obligation. 00:04:44.294 --> 00:04:49.174 And so suddenly freezing things, not paying bills, etc, is illegal. 00:04:50.234 --> 00:04:55.394 And so some of these freezers of foreign aid, or there was a brief one of a 00:04:55.394 --> 00:04:59.434 lot of domestic spending that was very hastily reversed, aren't legal. 00:04:59.434 --> 00:05:03.874 Some of the idea of, you know, just not paying a bill isn't legal. 00:05:04.154 --> 00:05:09.334 And so you run into problems there. You run into problems around access to information. 00:05:09.334 --> 00:05:10.554 That's very controlled. 00:05:11.394 --> 00:05:17.394 There's this email, which, you know, let's stress, Elon isn't really trying to hide what he's doing. 00:05:17.574 --> 00:05:21.994 His email to Twitter employees was, the subject line was fork in the road. 00:05:22.294 --> 00:05:25.634 The subject line to federal employees was fork in the road. 00:05:26.454 --> 00:05:30.214 You can't just offer to pay. federal 00:05:30.214 --> 00:05:33.214 employees seven months wages to do nothing 00:05:33.214 --> 00:05:36.774 there are because because 00:05:36.774 --> 00:05:41.974 republicans don't like wasting government money there have been a lot of restrictions 00:05:41.974 --> 00:05:47.694 on things like how much you can pay federal workers to leave etc so you know 00:05:47.694 --> 00:05:52.074 this i think that's what that's what the media is missing though they're going 00:05:52.074 --> 00:05:56.654 trump's doing this trump's doing that but that makes it sound like a policy debate. 00:05:57.054 --> 00:06:01.694 To a lot of people, Trump suspends the operations of US aid. Well, good. 00:06:01.994 --> 00:06:07.134 I didn't want money going on foreign aid. That's what I elected him for. The actual... 00:06:08.150 --> 00:06:12.230 It is very significant that he's doing that, but actually not that surprising. 00:06:12.630 --> 00:06:16.770 The fact that he's ripping up the legal system and the constitution of the US 00:06:16.770 --> 00:06:18.830 to do it is, in fact, the top line. 00:06:19.130 --> 00:06:23.570 And I think by focusing on the what, they're actually just making it sound like, 00:06:23.650 --> 00:06:26.910 well, why hasn't previous Republicans just done this? 00:06:27.370 --> 00:06:32.470 And I think they're actually, through that omission and through that emphasis, 00:06:32.750 --> 00:06:37.650 quite badly misinforming the voters. because it does start to suggest that someone 00:06:37.650 --> 00:06:39.130 could have done this the whole time. 00:06:39.390 --> 00:06:44.770 But it also means that these quite fundamental checks and balances, 00:06:45.010 --> 00:06:47.730 and in fact the whole constitutional settlement of the US, 00:06:48.370 --> 00:06:56.210 is being rewritten on the fly extrajudicially and most Americans won't really 00:06:56.210 --> 00:06:57.830 think of that as the headline. 00:06:58.210 --> 00:07:02.070 Yeah, just because Elon Musk managed to become friends with Donald Trump? 00:07:02.710 --> 00:07:08.410 Yeah, pretty much. And there is a good chance that all of this overreaches and 00:07:08.410 --> 00:07:12.350 unravels before too long, probably very, very messily. 00:07:12.650 --> 00:07:17.490 But a lot of the Trump project, even beyond Musk, is doing things by ripping 00:07:17.490 --> 00:07:19.390 up the law and illegal protections. 00:07:19.830 --> 00:07:26.570 And I think, you know, maybe they're overcorrecting for getting too outraged 00:07:26.570 --> 00:07:28.470 at too many things in the first term. 00:07:28.650 --> 00:07:31.530 Right. That was the question I was going to ask you. you know 00:07:31.530 --> 00:07:36.810 it it is and they're trying to do just the facts you know not democracy dies 00:07:36.810 --> 00:07:42.170 in darkness but breaking down the constitutional order is is a factual thing 00:07:42.170 --> 00:07:47.310 is is something that should be in the news section and i think is just it's 00:07:47.310 --> 00:07:50.070 is the missed step it's the alarm that's not going off, 00:07:51.150 --> 00:07:56.330 well that's a positive way to start the show thank you enjoy your time in washington, 00:07:57.810 --> 00:08:03.610 well hopefully it's only a week and not uh 20 years yeah well we'll see we'll 00:08:03.610 --> 00:08:07.530 see if you carry up i'll join me to pause putting out the show until you're back safely. 00:08:09.890 --> 00:08:13.070 No i could never do that to you now we'll 00:08:13.070 --> 00:08:16.290 we'll get on to the main subject i had for you which is kind of a reverse of 00:08:16.290 --> 00:08:19.390 the show i did with with chris stoker walker where we mentioned deep seek and 00:08:19.390 --> 00:08:24.110 ai at the top of the show and then deep dived into sort of elon musk and those 00:08:24.110 --> 00:08:31.310 kind of things But I want to dig more into AI and its effects on journalism. 00:08:31.570 --> 00:08:34.130 And again, I think we need to have a conversation about whether journalism is 00:08:34.130 --> 00:08:36.490 covering this all properly or not. 00:08:36.630 --> 00:08:42.050 But just to kick things off, we had a flurry of excitement about DeepSeek and 00:08:42.050 --> 00:08:49.250 NVIDIA share prices and NVIDIA chips and DeepSeek versus ChatGPT. 00:08:50.317 --> 00:08:54.577 Will you forgive me and indulge me, James, by having a little chuckle at the 00:08:54.577 --> 00:09:00.837 expense of Sam Altman, who has got very upset that someone may have scraped and nicked his work, 00:09:02.137 --> 00:09:10.237 having spent years scraping and, dare I suggest, nicking a load of content and 00:09:10.237 --> 00:09:12.557 work that he hadn't made? 00:09:13.897 --> 00:09:18.577 Yeah, it's hard not to kind of go, no, nicked, isn't it? 00:09:19.477 --> 00:09:23.917 It's sad when people get to profit off your work and then come up with legal 00:09:23.917 --> 00:09:27.057 loopholes as to why they think they can, actually, isn't it? 00:09:27.257 --> 00:09:32.817 So sorry, Sam. And especially when one of Sam's key lieutenants sort of once 00:09:32.817 --> 00:09:38.317 said, well, yes, of course, generative AI may then lead to lots of media and 00:09:38.317 --> 00:09:42.197 creative jobs going, but maybe those jobs shouldn't have existed in the first place. 00:09:42.997 --> 00:09:48.177 This was an actual quote on a panel event by one of his sort of co-founders. 00:09:48.177 --> 00:09:52.597 And so there is a thing of, well, maybe open AI shouldn't have existed in the 00:09:52.597 --> 00:09:55.337 first place. You know, maybe you've just been outcompeted. 00:09:56.477 --> 00:10:01.297 Whoopsie. Yeah. Yeah. Schadenfreude is not a nice emotion, but it is fun. 00:10:02.337 --> 00:10:07.437 It is a thing that exists, James Ball. It is a thing. 00:10:09.337 --> 00:10:13.277 Well, having got that little chuckle out the way, just explain to us where we 00:10:13.277 --> 00:10:16.977 are in the development of this story. Because as I say, we had the beginning 00:10:16.977 --> 00:10:20.357 flurry of, oh, my goodness, isn't DeepSeek exciting? 00:10:20.717 --> 00:10:25.177 Oh, my goodness, look what's happened to NVIDIA shares, et cetera, et cetera, et cetera. 00:10:26.537 --> 00:10:30.057 Also, I made this point, I think, to Chris. Oh, my goodness, 00:10:30.177 --> 00:10:34.217 how did no journalist slash city analyst work out what was going on? 00:10:34.977 --> 00:10:38.757 And how did DeepSeek just arrive almost by, you know, accident? 00:10:39.057 --> 00:10:41.637 Not accident, but we took us by surprise. 00:10:42.397 --> 00:10:45.157 Um so where are we now or sort 00:10:45.157 --> 00:10:48.037 of a couple of weeks on so i 00:10:48.037 --> 00:10:50.857 i think i've got a couple of fairly contained takes and then 00:10:50.857 --> 00:10:53.777 i think that you know i think it has changed the game quite 00:10:53.777 --> 00:10:56.677 profoundly but i think that saves that for last i 00:10:56.677 --> 00:11:00.057 think the dumbest sort 00:11:00.057 --> 00:11:02.997 of first media stories were looking at the stock market 00:11:02.997 --> 00:11:08.497 reaction and so of course way too much of the media did that uh the stock market 00:11:08.497 --> 00:11:13.417 reacted a week late it took them a week to notice we keep getting told the market 00:11:13.417 --> 00:11:20.277 is very reactive and well informed and that that was sort of a long way down the line but, 00:11:21.411 --> 00:11:24.591 It's headline news when stocks fall and not when they rise. 00:11:24.911 --> 00:11:29.871 And this really irritated me. Almost every tech stock is up versus last month. 00:11:31.091 --> 00:11:36.631 Everything except NVIDIA is actually trading higher than it was a month ago. 00:11:37.491 --> 00:11:41.011 Yeah, although NVIDIA is the crucial one in this story. Well, 00:11:41.251 --> 00:11:44.491 yes and no. NVIDIA is still up versus six months ago. 00:11:44.811 --> 00:11:50.771 It lost 16% of its value. But it doesn't mean anything to go hot and heavy telling 00:11:50.771 --> 00:11:55.231 an audience that stocks have lost a trillion dollars in value when you don't 00:11:55.231 --> 00:11:59.311 tell them that they've gained about two and a half trillion in the previous six months. 00:12:00.311 --> 00:12:07.231 Like it just it's completely misinforming the audience as to what matters and what doesn't. 00:12:07.391 --> 00:12:11.111 And the stock prices of these companies are the least important thing there. 00:12:11.251 --> 00:12:15.191 That's essentially just going, what does the city think is important? it 00:12:15.191 --> 00:12:18.431 it matters on one level but actually that it wasn't 00:12:18.431 --> 00:12:21.931 a very important crash it wasn't a very systemic crash and 00:12:21.931 --> 00:12:25.691 it starts making people think that looking at 00:12:25.691 --> 00:12:28.491 that line tells you what you need to know and actually it 00:12:28.491 --> 00:12:31.551 tells you what a panicked herd of people 00:12:31.551 --> 00:12:36.691 trying to make profit on the day think and so i i think it was a very silly 00:12:36.691 --> 00:12:41.851 lens to look at this story through through and i wish it's very relevant to 00:12:41.851 --> 00:12:46.351 the business pages i think it's not relevant to the front pages and when the 00:12:46.351 --> 00:12:50.911 front pages and when tv news starts talking about stocks it does it incredibly badly. 00:12:52.451 --> 00:12:57.831 And and this is also part of why british people have such stupid attitudes towards 00:12:57.831 --> 00:13:03.751 property you know we consistently invest in property no no but we consistently 00:13:03.751 --> 00:13:08.031 invest in property which leaves us poorer because the only time any normal british 00:13:08.031 --> 00:13:11.071 person hears about stocks is to hear they've crashed catastrophically. 00:13:11.431 --> 00:13:15.151 America has the opposite problem in that all of their daily news talks about 00:13:15.151 --> 00:13:19.031 how often stocks go up. And so they all speculate and retail invest. 00:13:19.351 --> 00:13:21.931 I feel like this is a bit of a therapy session for you. 00:13:22.311 --> 00:13:25.011 It's just incredibly irritating coverage. 00:13:25.451 --> 00:13:27.051 Like America's too... 00:13:27.641 --> 00:13:30.721 America's too boosterish the uk kind of when 00:13:30.721 --> 00:13:34.301 it covers it just covers crashes and covers them very badly um 00:13:34.301 --> 00:13:37.521 big big tech is fine nvidia is fine like 00:13:37.521 --> 00:13:40.481 their material circumstance have actually changed very little 00:13:40.481 --> 00:13:43.381 i think the the other story that 00:13:43.381 --> 00:13:46.161 everyone did was look what happens when i typed 00:13:46.161 --> 00:13:49.001 tianuban square in it which sure yeah i saw 00:13:49.001 --> 00:13:51.941 a lot of that i think it's a 00:13:51.941 --> 00:13:54.941 bit more legit but i think was done in a very banal way 00:13:54.941 --> 00:13:58.381 um and there is a more interesting story 00:13:58.381 --> 00:14:01.221 going on here what i want to 00:14:01.221 --> 00:14:05.921 ask you is what should have happened so i think it's worth reporting what happens 00:14:05.921 --> 00:14:12.021 when you type it in i think what gets quite interesting is that these models 00:14:12.021 --> 00:14:17.441 are open source and the censorship is built into the app not into the model 00:14:17.441 --> 00:14:21.201 right and so if you put deep DeepSeek, 00:14:21.401 --> 00:14:25.181 one of the impressive things about it is the lightweight versions of it will 00:14:25.181 --> 00:14:27.481 run on your own laptop. You can run it locally. 00:14:27.841 --> 00:14:31.601 If you want to run a powerful version of it, you can do it relatively cheaply 00:14:31.601 --> 00:14:34.141 in the cloud. It is not that compute intensive. 00:14:35.541 --> 00:14:39.421 Which is the whole thing, forgive me to come back to it, which shocked people 00:14:39.421 --> 00:14:43.461 around NVIDIA, because everyone had said for ages that all these AI models are 00:14:43.461 --> 00:14:45.221 going to need huge computing power. 00:14:45.521 --> 00:14:47.301 And DeepSeek does not. 00:14:48.001 --> 00:14:52.501 Yes. Now, there is, just to do a tangent briefly, there are some questions being 00:14:52.501 --> 00:14:56.021 asked about whether it was quite so cheap to train as they say it was. 00:14:56.121 --> 00:14:57.961 They say it cost 5.6 million. 00:14:58.921 --> 00:15:04.041 There are some researchers speculating it could have been anything up to 500 million. 00:15:05.961 --> 00:15:10.801 There may be some overblown claims here, so it might not be as revolutionary as it appears. 00:15:11.081 --> 00:15:16.521 And I don't have a strong view either way, but I think it's an interesting line of inquiry. 00:15:17.041 --> 00:15:21.821 That seems unlikely, James. Well, the 5.6 million is... That you don't have 00:15:21.821 --> 00:15:23.741 a strong view on it. I don't know enough. 00:15:24.561 --> 00:15:31.121 5.6 million is a very strong claim to make, to just take on faith without evidence. 00:15:31.461 --> 00:15:35.621 Sure. You know, we don't sort of say that someone's discovered dark matter... 00:15:35.869 --> 00:15:41.369 And not ask for further proof, I would like some demonstration that it was done that cheaply. 00:15:42.109 --> 00:15:46.349 I certainly believe they've got some real innovations that it was done more cheaply. 00:15:46.969 --> 00:15:51.389 I would like to see some more on the extent of it. So I think that's some interesting 00:15:51.389 --> 00:15:54.949 coverage for some people to do. But why hasn't that been done? 00:15:55.229 --> 00:15:58.789 This is what I have to keep coming back with when you and I have these conversations. 00:15:59.589 --> 00:16:02.869 This is a media show, I want to analyze, you 00:16:02.869 --> 00:16:05.989 and I could have fastened conversations about AI models and 00:16:05.989 --> 00:16:08.789 so on but the key question for this show and 00:16:08.789 --> 00:16:12.169 my audience is why is this stuff consistently being 00:16:12.169 --> 00:16:15.769 missed is it because there's not enough expert reporters 00:16:15.769 --> 00:16:18.869 in these mainstream publications is it because expert 00:16:18.869 --> 00:16:22.929 and niche publications are doing this but it's getting lost in the noise what 00:16:22.929 --> 00:16:26.789 so how do you perceive all of that happening I mean I could be facetious and 00:16:26.789 --> 00:16:32.529 answer yes all of the above right okay but I think that there There are some 00:16:32.529 --> 00:16:36.169 niche publications doing some of this work, which is why I'm aware of some of it. 00:16:36.289 --> 00:16:40.149 But they are limited in what they can look at, how much they can run out at once. 00:16:40.369 --> 00:16:48.649 I think there is still a feeling in mainstream media that tech is not a serious beat or is a niche beat. 00:16:48.889 --> 00:16:52.649 Really? But also that superficial coverage of it is fine. 00:16:52.909 --> 00:16:56.829 And you can run at that and people don't want the more detailed version of the 00:16:56.829 --> 00:17:00.029 story, don't want the extra levels in it. 00:17:00.029 --> 00:17:04.629 And to finish this sort of censorship thing, because this is a coverage one, 00:17:04.829 --> 00:17:09.089 if you install DeepSeek on your own computer or on a thing and ask it about 00:17:09.089 --> 00:17:10.689 Tiananmen Square, it answers. 00:17:11.089 --> 00:17:16.969 You don't get, I can't answer this question, but you do get an answer that focuses 00:17:16.969 --> 00:17:22.749 a lot on China upholding the rule of law, acting within the law and etc. 00:17:22.749 --> 00:17:24.789 So it does say there was a crackdown. 00:17:25.249 --> 00:17:30.469 And this is more interesting for me, because this isn't about the model being 00:17:30.469 --> 00:17:33.389 told what to answer. This is about what the model was trained on. 00:17:34.009 --> 00:17:38.969 Right. And of course, a Chinese-based model, a non-Western model, 00:17:39.289 --> 00:17:47.049 has been fed a lot of information and a lot of background that isn't from the Western perspective. 00:17:47.429 --> 00:17:51.609 And if it's from a Chinese perspective, will have come from a very restricted 00:17:51.609 --> 00:17:55.049 lexicon of media, a very restricted lexicon of everything. 00:17:56.929 --> 00:18:01.909 And just going, well, is there a censorship thing? Can you type something to avoid it? 00:18:02.369 --> 00:18:06.509 Just makes it look a bit easy. what do you do if suddenly you have dominant 00:18:06.509 --> 00:18:10.689 technologies across the internet that aren't from a US perspective? 00:18:11.069 --> 00:18:14.889 And also, if we want to get sort of really deep... 00:18:16.488 --> 00:18:23.068 What has it been doing to us for the past 30 years to only have big tech from the U.S. perspective? 00:18:23.568 --> 00:18:28.528 What blind spots is it giving us? I don't think the U.S. has been engaged in 00:18:28.528 --> 00:18:29.648 a grand censorship project. 00:18:29.788 --> 00:18:33.788 I absolutely don't want to draw an equivalence between U.S. 00:18:34.268 --> 00:18:40.008 Biases and ideologies and China's outright censorship and draconian crackdown on free speech. 00:18:40.008 --> 00:18:43.228 But that more subtle question 00:18:43.228 --> 00:18:46.428 of deep seek when it gives you answers giving 00:18:46.428 --> 00:18:49.908 you answers that are different to what we're used to should we 00:18:49.908 --> 00:18:52.828 actually have more pluralism in tech what 00:18:52.828 --> 00:18:55.748 you know should we accept you know 00:18:55.748 --> 00:18:58.928 the versions of deep seek that give us different answers that 00:18:58.928 --> 00:19:01.788 emphasize i think falsely china's respect for 00:19:01.788 --> 00:19:04.588 rule of law etc that that to me 00:19:04.588 --> 00:19:08.088 is a much more interesting and textured debate than sort 00:19:08.088 --> 00:19:10.988 of going look what happens when I type Tiananmen Square and let 00:19:10.988 --> 00:19:13.888 me stress I was doing some look what happens when I type Tiananmen Square. 00:19:13.888 --> 00:19:16.688 Stuff as well it's it's it's there I mean 00:19:16.688 --> 00:19:19.868 what was interesting to me was it would give me several interesting answers 00:19:19.868 --> 00:19:23.368 on Taiwan and chip manufacture and 00:19:23.368 --> 00:19:26.468 it would type out the whole answer and you I screen recorded 00:19:26.468 --> 00:19:29.468 this and then delete it oh and so 00:19:29.468 --> 00:19:32.388 it would give you actually I thought on 00:19:32.388 --> 00:19:35.628 Taiwan and chip manufacturer and geopolitical sensitivity 00:19:35.628 --> 00:19:38.728 a genuinely balanced answer 00:19:38.728 --> 00:19:42.028 i you know what i thought was quite a high quality response. 00:19:42.028 --> 00:19:44.908 And then delete it uh but 00:19:44.908 --> 00:19:47.988 because it chunked the answer out quite slowly you 00:19:47.988 --> 00:19:50.948 could read the whole answer um and again 00:19:50.948 --> 00:19:54.468 some of that gets you to a much more interesting 00:19:54.468 --> 00:19:57.508 debate about what tech is going to do to 00:19:57.508 --> 00:20:00.428 the information ecosystem to the media ecosystem to 00:20:00.428 --> 00:20:03.488 perspectives than just that that chinese 00:20:03.488 --> 00:20:06.548 that you know chinese censorship one and so 00:20:06.548 --> 00:20:09.928 i i wish more places hadn't just 00:20:09.928 --> 00:20:12.868 stopped at the gotcha which is 00:20:12.868 --> 00:20:15.588 a legitimate gotcha i i'm yeah i've done a lot of 00:20:15.588 --> 00:20:18.408 slagging off mainstream journalism in this one there's nothing wrong 00:20:18.408 --> 00:20:21.328 with that but i wish more had done the 00:20:21.328 --> 00:20:25.368 follow-ups and maybe some places will yeah um 00:20:25.368 --> 00:20:28.088 i this brings me back to the key point that i bore on 00:20:28.088 --> 00:20:30.928 about every time i have a guest on to talk about 00:20:30.928 --> 00:20:33.628 ai i think i did it with comrade cruelty harper last week 00:20:33.628 --> 00:20:37.248 i did it with chris stoker walker before um i think you're the first person 00:20:37.248 --> 00:20:40.828 on this show this year that hasn't got a double barrel name james actually now 00:20:40.828 --> 00:20:45.568 i think about it so if you could sort that out that would be good um so i'll 00:20:45.568 --> 00:20:51.128 work on it i promise thank you thank you but i mentioned this a lot because um. 00:20:52.389 --> 00:20:56.629 People have forgotten. Again, I think this might be a nature of the coverage, 00:20:56.809 --> 00:21:03.809 but also just the nature of, you know, the way quite legitimately normal people engage with this stuff. 00:21:03.809 --> 00:21:07.589 Like it's not most people's job to worry about AI models and so on. 00:21:09.129 --> 00:21:15.969 People have forgotten, I think, that AI tools, however clever they are, 00:21:16.129 --> 00:21:18.109 at the current moment where you and I sit here, 00:21:18.629 --> 00:21:25.789 can only put out stuff based on what has been put in by other people. 00:21:26.849 --> 00:21:33.049 And I think we have totally lost sight of that. Yes, I think most journalists 00:21:33.049 --> 00:21:36.649 don't understand much about how AI models work. 00:21:36.829 --> 00:21:41.269 And I don't necessarily blame journalists for that. They're very complex. 00:21:41.489 --> 00:21:46.109 It's very hard. And also, the way a lot of AI companies talk about the tech. 00:21:46.109 --> 00:21:49.269 Is getting detached from what it is. 00:21:49.269 --> 00:21:52.809 But essentially they are they average 00:21:52.809 --> 00:21:56.129 out responses that they've had and they're. 00:21:56.129 --> 00:21:59.469 Getting better and better at finding out what's relevant at what's. 00:21:59.469 --> 00:22:02.269 There but they can't create and they 00:22:02.269 --> 00:22:05.429 can't innovate so they can write very good simple practical 00:22:05.429 --> 00:22:08.189 code um they're increasingly good 00:22:08.189 --> 00:22:11.449 at sub-editing uh like they 00:22:11.449 --> 00:22:14.609 can do things like that but they won't turn 00:22:14.609 --> 00:22:17.949 out wonderful innovative in terms of phrase they will flatten 00:22:17.949 --> 00:22:21.049 everything out they will average everything out but they draw 00:22:21.049 --> 00:22:23.869 on what trains their model by and large 00:22:23.869 --> 00:22:27.029 some of them are plugged to the internet so they can get some information 00:22:27.029 --> 00:22:30.089 from there too but what they look for 00:22:30.089 --> 00:22:32.709 and what they prioritize will again be based on the 00:22:32.709 --> 00:22:35.709 data set so if you feed 00:22:35.709 --> 00:22:39.589 a model entirely chinese media what even 00:22:39.589 --> 00:22:42.349 what it would think to access on the 00:22:42.349 --> 00:22:45.869 internet and what it would search and how it would present it would 00:22:45.869 --> 00:22:48.989 be quite radically different from a u.s media trained 00:22:48.989 --> 00:22:51.849 model and so you know 00:22:51.849 --> 00:22:54.849 it weirdly it's it is the the issue 00:22:54.849 --> 00:22:57.989 with any automated system is always that it can 00:22:57.989 --> 00:23:01.009 average real world biases so if you 00:23:01.009 --> 00:23:04.009 if you did a sentencing algorithm that looked at 00:23:04.009 --> 00:23:07.249 historical sentencing data in america 00:23:07.249 --> 00:23:10.089 it would probably be much more likely to jail black 00:23:10.089 --> 00:23:12.789 offenders than white offenders and you could say 00:23:12.789 --> 00:23:16.909 it's an algorithm it can't possibly be racist but if the data that's trained 00:23:16.909 --> 00:23:21.809 it is racist it will and so you have all of these complications and all of these 00:23:21.809 --> 00:23:27.209 factors and ai doesn't get you anywhere away from them even with these models. 00:23:27.209 --> 00:23:31.429 And the fact that they operate as a black box is really difficult because. 00:23:32.069 --> 00:23:36.409 No one can go here's how it reached this decision and here's how it worked including 00:23:36.409 --> 00:23:43.569 the person that made it yeah and ai's and this gets a little bit this is another 00:23:43.569 --> 00:23:48.349 one that sounds like you've been on weed all night or something you can ask 00:23:48.349 --> 00:23:51.289 an ai model to explain to you how it reached a decision, 00:23:52.407 --> 00:23:55.967 That may have nothing to do with how it actually reached a decision. 00:23:56.227 --> 00:24:02.567 That may well just be it coming up with the best explanation to tell you for 00:24:02.567 --> 00:24:03.987 how it reached a decision. 00:24:04.447 --> 00:24:09.527 Yeah. And there's no way to align that it's definitely the case. 00:24:10.067 --> 00:24:13.407 Yeah. And there's all the issues with hallucinations and so on. 00:24:13.407 --> 00:24:19.227 But how do we as journalists, tech on the tech beat and otherwise, 00:24:19.867 --> 00:24:25.307 explain all this to people who might just see these terrifying stories and go, 00:24:25.407 --> 00:24:26.967 oh, my God, AI is coming for my job? 00:24:28.147 --> 00:24:34.307 I think it's about trying to be as clear eyed as possible. And I think in this we actually need to. 00:24:35.127 --> 00:24:39.627 I am usually quite in favor of journalists having a voice and having an opinion. 00:24:39.627 --> 00:24:43.207 Um and i think it's usually useful to 00:24:43.207 --> 00:24:46.107 try and go he is what i think he is where i'm coming from 00:24:46.107 --> 00:24:48.967 i think on ai we're getting a bit too 00:24:48.967 --> 00:24:52.007 much of it i think there are journalists who 00:24:52.007 --> 00:24:55.587 i quite like and admire who are justifiably 00:24:55.587 --> 00:24:58.527 burned after years of tech saying the metaverse is 00:24:58.527 --> 00:25:01.487 the next thing cryptocurrency is the next thing nfts are 00:25:01.487 --> 00:25:05.687 there and them all being crap um who 00:25:05.687 --> 00:25:08.767 i think are just kind of looking at AI and 00:25:08.767 --> 00:25:11.547 going this is the next crapshoot this isn't 00:25:11.547 --> 00:25:16.287 this doesn't work this isn't interesting this isn't uh and we're journalists 00:25:16.287 --> 00:25:20.467 so we need to be cynical about it yes and I think you've got others who are 00:25:20.467 --> 00:25:24.807 going we need to turn this off right now the sky is falling and you've got others 00:25:24.807 --> 00:25:28.867 who are going it's going to completely transform society within years I think, 00:25:29.641 --> 00:25:34.261 All of those are perfectly reasonable opinions to have. They're legitimate opinions. 00:25:35.261 --> 00:25:42.741 But actually, there is a working operating technology now that needs to be explained in factual terms. 00:25:42.981 --> 00:25:47.681 And I think we're all trying to get so much into what it means in five years' 00:25:47.821 --> 00:25:51.881 time that we haven't explained the technology that's in front of us, 00:25:51.901 --> 00:25:57.521 which is capable of some genuinely impressive things it couldn't do a year or two years ago. 00:25:57.521 --> 00:26:00.321 But also we were hearing 00:26:00.321 --> 00:26:03.781 in 2023 that these things would be replacing call 00:26:03.781 --> 00:26:06.941 center workers by now and that we'd all have ai personal assistants 00:26:06.941 --> 00:26:09.701 and that we'd all have and a lot of that's been a lot 00:26:09.701 --> 00:26:12.801 slower than people thought because it's really 00:26:12.801 --> 00:26:17.001 hard to get from 95 accurate to 99.9 and 00:26:17.001 --> 00:26:19.861 if you're gonna let something have access to your credit card and 00:26:19.861 --> 00:26:23.381 tell it to book hotels for you or this kind of thing they've 00:26:23.381 --> 00:26:26.281 got to be a lot better than they are and they're finding that side 00:26:26.281 --> 00:26:29.561 of things really difficult and so why you 00:26:29.561 --> 00:26:32.541 know what can they do what are they good at what are they bad at what 00:26:32.541 --> 00:26:35.641 did we think they'd be good at that they're bad at how do they work 00:26:35.641 --> 00:26:38.561 you know you tend to have to explain these things 00:26:38.561 --> 00:26:41.361 10 15 20 times before they really 00:26:41.361 --> 00:26:46.061 break through i think we're all trying to skip a page and do the big is this 00:26:46.061 --> 00:26:51.141 the end of humanity is this the end of work does this mean you know fully automated 00:26:51.141 --> 00:26:56.761 luxury communism for us all um i think we actually need to take a step back 00:26:56.761 --> 00:26:59.341 and do more of the fundamentals that we have been doing. 00:27:00.821 --> 00:27:03.621 Yeah yeah i think that that is a fair 00:27:03.621 --> 00:27:11.301 point and i think sometimes in journalism because we by definition journalists 00:27:11.301 --> 00:27:16.581 pick up the story earlier than the public i mean that's but you know yes if 00:27:16.581 --> 00:27:22.801 we're any good at our job yes correct and so that leads to a tendency, 00:27:23.581 --> 00:27:28.581 to go oh we've known this for ages it's not interesting yeah except actually 00:27:28.581 --> 00:27:31.521 the people who need to know about it have not known about it for ages, 00:27:32.693 --> 00:27:38.953 And sometimes we skip off reassessing, you know, if some people I think played with chat GPT-3, 00:27:39.813 --> 00:27:43.333 realised it wasn't nearly as clever as it was being made out to be and haven't 00:27:43.333 --> 00:27:49.173 really kept playing with the later models, which are absolutely night and day to it. 00:27:49.253 --> 00:27:52.653 You know, I use AI most days and it is extraordinary. 00:27:53.793 --> 00:27:56.433 I'm pleased you brought that up because I did want to ask you, 00:27:56.513 --> 00:27:58.653 how do you use AI in your journalism? 00:27:59.393 --> 00:28:02.513 I try and use it for discovery. so i 00:28:02.513 --> 00:28:05.593 i before i send things to editors i i do 00:28:05.593 --> 00:28:08.593 can you sub this uh for british english 00:28:08.593 --> 00:28:11.793 for grammatical errors subject verb disagreement 00:28:11.793 --> 00:28:14.633 and it's good at things that 00:28:14.633 --> 00:28:17.833 spell check miss like if you have if instead of of or 00:28:17.833 --> 00:28:21.473 on instead of in yeah you know guilty and 00:28:21.473 --> 00:28:24.513 guilty like they're hard mistakes to catch 00:28:24.513 --> 00:28:27.453 they're hard for subs to catch you know spell check 00:28:27.453 --> 00:28:30.393 doesn't get them so basic practical stuff 00:28:30.393 --> 00:28:33.613 like that is great um i like 00:28:33.613 --> 00:28:36.473 to have it suggest sources or 00:28:36.473 --> 00:28:40.113 suggest angles or suggest um you 00:28:40.113 --> 00:28:43.173 know are there some journal articles i i should read it 00:28:43.173 --> 00:28:46.013 hallucinates journal articles quite often but it 00:28:46.013 --> 00:28:48.853 might suggest an author that i don't know it tends to 00:28:48.853 --> 00:28:52.113 be that the author has written a journal article fairly. 00:28:52.113 --> 00:28:55.533 Similar to the one it says um but that 00:28:55.533 --> 00:28:58.313 that rate goes down i don't ask it to 00:28:58.313 --> 00:29:01.813 do research for me um sometimes i 00:29:01.813 --> 00:29:05.413 will ask it to come up with a four different 00:29:05.413 --> 00:29:09.833 paragraph long summaries of an idea um and 00:29:09.833 --> 00:29:12.773 so if i'm trying to think of a way to define a 00:29:12.773 --> 00:29:15.653 concept etc and i want to check my 00:29:15.653 --> 00:29:18.713 understanding of it uh i you 00:29:18.713 --> 00:29:21.813 know what's you know uh who 00:29:21.813 --> 00:29:27.113 coined the late stage capitalism for example and you know it's marx it's 150 00:29:27.113 --> 00:29:31.813 years old it's and you know why did he say it was late stage if you're just 00:29:31.813 --> 00:29:36.533 trying to get into a concept i think the issue is people use it as a search 00:29:36.533 --> 00:29:39.473 engine and that's the thing it's worst at yes uh. 00:29:40.354 --> 00:29:43.314 Uh using it as a tool for discovery using it 00:29:43.314 --> 00:29:46.314 as a tool for refinement actually works 00:29:46.314 --> 00:29:48.994 a lot better i i mean quite often i'm playing with 00:29:48.994 --> 00:29:52.234 it also i have it write code um so 00:29:52.234 --> 00:29:55.634 i've i've had it do things like um i 00:29:55.634 --> 00:29:59.114 had i wanted to make sure i had screen grabs of 00:29:59.114 --> 00:30:01.814 a thousand web pages that i'd looked at 00:30:01.814 --> 00:30:04.834 in the research for a big project because this 00:30:04.834 --> 00:30:07.814 is the most james ball thing i've ever heard yeah of course you did that if 00:30:07.814 --> 00:30:10.714 they were taken offline would i still be able to prove 00:30:10.714 --> 00:30:13.894 that it said it and so it's a 00:30:13.894 --> 00:30:16.854 while since i've had to write code to do that and so 00:30:16.854 --> 00:30:19.774 essentially i went can you do something that will go through open the 00:30:19.774 --> 00:30:22.794 link in an excel sheet uh open it 00:30:22.794 --> 00:30:26.794 screen grab it close it i repeat um and 00:30:26.794 --> 00:30:29.614 it took it about three tries to get the code so it 00:30:29.614 --> 00:30:33.474 worked but in each instance i had it run the code said here's the error message 00:30:33.474 --> 00:30:38.554 or here's what it did can you fix it and by the third time it fixed it and so 00:30:38.554 --> 00:30:44.334 it would have taken me about 90 minutes to write that properly because i'm i'm 00:30:44.334 --> 00:30:49.514 not good at this uh and i don't do it often enough to get better, 00:30:50.174 --> 00:30:54.614 uh but because i knew it was a simple task and i knew what i was asking it to do. 00:30:55.474 --> 00:31:02.094 I could get it to do it in 10 minutes and then so it's useful yeah this is a 00:31:02.094 --> 00:31:05.434 really i'm pleased we've had this kind of more balanced conversation because 00:31:05.434 --> 00:31:07.594 it feels to me like you think there are, 00:31:08.314 --> 00:31:11.994 genuinely good use cases for the 00:31:11.994 --> 00:31:16.274 tools both as they currently are and how they will probably develop for within 00:31:16.274 --> 00:31:21.554 journalism yes instead of us all freaking out going oh my god it's going to 00:31:21.554 --> 00:31:27.494 take my job and write the articles for me i i mean i i worry i worry about how 00:31:27.494 --> 00:31:28.734 it will actually be used, 00:31:28.854 --> 00:31:36.114 and I suspect that leads to the bad side because I look at the model of some of the companies, 00:31:36.914 --> 00:31:38.954 you know, look, let's name them. 00:31:39.054 --> 00:31:42.594 I look at Reach and go, I know exactly how they're going to use AI, 00:31:42.934 --> 00:31:45.074 and it's going to be terrible. 00:31:46.278 --> 00:31:50.398 Newsrooms that want to help good journalism, if you can do things like automate 00:31:50.398 --> 00:31:55.518 tasks that are repetitive and boring to sort of make it quicker, 00:31:55.718 --> 00:32:01.258 if you can smooth the production process, if you can use it to enhance the abilities 00:32:01.258 --> 00:32:07.558 of human reporters and human editors, it actually has a lot of potential to improve it. 00:32:07.558 --> 00:32:13.818 But that needs enlightened owners who sort of go, we can make the product better 00:32:13.818 --> 00:32:20.058 rather than we can make the product only slightly worse, but at half the price. 00:32:20.258 --> 00:32:25.058 Essentially, I think AI can make the work and the output of human reporters 00:32:25.058 --> 00:32:28.158 and editors quite a lot better for very little cost. 00:32:28.158 --> 00:32:33.898 It can let newsrooms get a lot more bang for the buck, or it could let newsrooms 00:32:33.898 --> 00:32:40.178 make slightly worse output for a much lower cost by getting rid of reporters. 00:32:40.358 --> 00:32:45.038 And it can make average copy and averagely synthesize events. 00:32:45.278 --> 00:32:50.518 And what I worry is that the industry has spent the last two decades or parts 00:32:50.518 --> 00:32:56.278 of the industry has spent the last two decades making the output worse to reduce costs. 00:32:56.278 --> 00:33:01.738 And so I worry they'll look at a technology that could actually make journalism 00:33:01.738 --> 00:33:05.538 better and use it to carry on making it much worse. 00:33:05.698 --> 00:33:08.398 But that's the fault of media bosses, not AI. 00:33:08.878 --> 00:33:12.238 Listening to you describe all this and knowing some of the big investigations 00:33:12.238 --> 00:33:16.318 you've worked on, how much would you have loved to have some of these tools 00:33:16.318 --> 00:33:19.938 to dig through all the thousands of records, millions of records, 00:33:20.078 --> 00:33:23.198 I think in some cases, you had to for some of the big stories you've worked on. 00:33:23.258 --> 00:33:26.498 How much easier would that have made your life? it would be amazing 00:33:26.498 --> 00:33:29.618 wouldn't it but i i don't think we ever could have um 00:33:29.618 --> 00:33:33.098 okay i wonder if we could have put a on 00:33:33.098 --> 00:33:36.858 an air-gapped computer that had never touched the internet could 00:33:36.858 --> 00:33:39.838 we have run a model like llama or like deep 00:33:39.838 --> 00:33:46.798 seek and pointed it at the data so that we knew it couldn't phone home to anywhere 00:33:46.798 --> 00:33:52.578 maybe um i wonder if it would have tripped any protections because it was obviously 00:33:52.578 --> 00:33:57.478 looking at classified documents I wonder if they have anything at that but obviously. 00:33:58.314 --> 00:34:04.174 Nothing to do with Snowden could ever touch the website, could ever touch the Internet. 00:34:04.234 --> 00:34:06.914 You know, we did that completely on air-gapped computers. 00:34:07.254 --> 00:34:13.294 And I mean, we also, all the laptops we did on Snowden, I bought in cash from Best Buy. 00:34:13.534 --> 00:34:17.734 And we bought different models of laptop from different Best Buys at different 00:34:17.734 --> 00:34:24.974 times so that they couldn't be intercepted by the NSA on their way being delivered from Amazon, etc. 00:34:24.974 --> 00:34:29.354 Or, of course, by Russian intelligence or Chinese intelligence to put something 00:34:29.354 --> 00:34:30.974 in them to try and get documents. 00:34:31.454 --> 00:34:35.514 You know, this was about trying to keep this secure. And so I think loading 00:34:35.514 --> 00:34:39.834 it into an AI that we couldn't control and couldn't tell you the building of 00:34:39.834 --> 00:34:43.354 and the waiting of would probably have gone quite against it. 00:34:43.494 --> 00:34:47.934 But I think the potential for what they could have done in terms of semantic 00:34:47.934 --> 00:34:53.474 search and understanding could have saved us weeks of work or months of work. 00:34:53.474 --> 00:34:56.414 But i think we would have probably ruled against 00:34:56.414 --> 00:34:59.434 using it even if we could have done it solving the 00:34:59.434 --> 00:35:02.054 internet problem which is a shame i mean 00:35:02.054 --> 00:35:05.994 security tends to run against convenience and i 00:35:05.994 --> 00:35:12.134 think that's going to stay particularly true in uh in the ai era yeah and i 00:35:12.134 --> 00:35:15.574 think it's particularly true and i mentioned this previously uh when you're 00:35:15.574 --> 00:35:18.274 talking about deep seek because everyone's freaking out about having tiktok 00:35:18.274 --> 00:35:24.234 on their phone but seems to be happily installing a chinese built AI model app onto their phone. 00:35:24.794 --> 00:35:28.094 James, I could talk to you about this for ages, and I probably will in the pub, 00:35:28.334 --> 00:35:30.674 but sadly, the show has to come to an end. 00:35:31.054 --> 00:35:33.994 Just tell people where they can keep up with you and all of your fantastic work. 00:35:34.434 --> 00:35:38.454 So they can subscribe to the New European, where I'm political editor, 00:35:38.654 --> 00:35:42.294 and I have 1600 words essay on AI this week. 00:35:43.574 --> 00:35:46.814 They can follow me on twitter on x still just 00:35:46.814 --> 00:35:49.554 about at james rb uk i'm on 00:35:49.554 --> 00:35:52.274 blue sky at james rball.com or you 00:35:52.274 --> 00:35:56.174 know i'm just floating around in the ether um we'll 00:35:56.174 --> 00:35:59.314 find that and i'll link to your substat newsletter which occasionally goes out 00:35:59.314 --> 00:36:02.914 to people yes it is very intermittent but 00:36:02.914 --> 00:36:05.834 it's always interesting when it's there i'm at charlotte a 00:36:05.834 --> 00:36:09.414 henry across various social media platforms of 00:36:09.414 --> 00:36:12.594 course if you listen to this i hope you know about the edition.net where 00:36:12.594 --> 00:36:17.594 i post newsletters blogs and indeed these podcast episodes and I would be thrilled 00:36:17.594 --> 00:36:21.594 if you signed up for a paid subscription that would be really lovely because 00:36:21.594 --> 00:36:27.194 it helps keep the whole show on the road so that is at theedition.net James 00:36:27.194 --> 00:36:30.814 thank you once again for joining me for this fascinating conversation and I'll see you all next week. 00:36:31.440 --> 00:36:40.325 Music.