WEBVTT - Week in Tech: You, Me & Ghibli 0:00:00.520 --> 0:00:04.320 Welcome to Tech Stuff, a production of iHeart Podcasts and Kaleidoscope. 0:00:04.720 --> 0:00:07.160 I'm mos Vloscian and today Cara Price and I will 0:00:07.160 --> 0:00:10.680 bring you the headlines this week, including the fashion models 0:00:10.720 --> 0:00:14.320 getting Ai twins. Then on Tech Support, we'll talk to 0:00:14.320 --> 0:00:17.720 Garrett da vinc a reporter at the Washington Post about 0:00:17.760 --> 0:00:20.520 all the news in AI from Open Ai, it's huge 0:00:20.520 --> 0:00:25.920 new fundraise, to Xai's acquisition of x formerly Twitter, all 0:00:25.960 --> 0:00:27.320 of that on the Weekend. 0:00:27.000 --> 0:00:37.400 Tech It's Friday, April fourth, Cara Price. 0:00:37.440 --> 0:00:40.560 As Volashan, you're looking like me today. 0:00:40.680 --> 0:00:43.400 Yeah, we are matching. 0:00:43.720 --> 0:00:46.600 We are very matchie matchie. I was at. 0:00:46.560 --> 0:00:49.600 McLaren f one's headquarters as you know, earlier this week 0:00:49.720 --> 0:00:51.519 and I sent you a picture that I was very 0:00:51.520 --> 0:00:55.760 proud of me and Zach Brown, the CEO of McLaren Racing, 0:00:56.200 --> 0:00:59.280 And your only comment was, oh, black trainers. 0:00:58.920 --> 0:01:02.280 Always black six sneakers. You can sponsor us, but yes, 0:01:02.360 --> 0:01:04.680 always black A six sneakers, even if he's wearing a tie. 0:01:05.800 --> 0:01:08.120 But it's, you know, for a tech podcast, it's very 0:01:08.160 --> 0:01:09.880 funny how much we talk about fashion. And it's not 0:01:09.920 --> 0:01:10.680 because I'm a woman. 0:01:10.920 --> 0:01:13.120 I mean, it is also ironic that we both have 0:01:13.240 --> 0:01:16.960 our own uniforms. We're wearing blue blue shirts, dark blue pants, shirt, 0:01:17.000 --> 0:01:20.520 blubby pants, and a cool hat. 0:01:20.560 --> 0:01:23.280 Though I am wearing a cool hat, I'm always wearing well, 0:01:23.319 --> 0:01:24.760 I wouldn't call it a cool hat, but if you 0:01:24.760 --> 0:01:27.600 want to call it a cool hat, it's an appropriated 0:01:27.840 --> 0:01:31.080 Yankees city by hat that is completely illegal and was 0:01:31.080 --> 0:01:33.280 taken off the internet the minute it went up, So 0:01:33.480 --> 0:01:34.960 I'm very proud to be wearing it now. 0:01:35.000 --> 0:01:36.959 The reason we're talking about this is not just because 0:01:37.000 --> 0:01:40.520 we're totally self indulgent, but because the topic of tech 0:01:40.640 --> 0:01:44.680 folks and their uniforms is pretty fascinating. Earlier this year, 0:01:44.680 --> 0:01:48.160 we talked about Mark Zuckerberg, who had this famous, weirdly 0:01:48.240 --> 0:01:51.880 shaped T shirt with the Latin phrase out zuk out 0:01:52.000 --> 0:01:54.880 nihil printed on it, which means all zook ll all 0:01:54.920 --> 0:01:58.400 nothing and is a is a reference to Julius Caesar. 0:01:58.520 --> 0:01:59.680 Of course I can't. 0:02:00.080 --> 0:02:04.960 It's just insane to me. The level of It's not narcissism. 0:02:05.080 --> 0:02:09.520 It's this sense of like an obsession with being cool that, 0:02:09.680 --> 0:02:14.359 while chased, is the antithesis of anything cool and fashionable. 0:02:14.639 --> 0:02:17.679 I thought that one of the genuinely coolest tech swipes 0:02:17.720 --> 0:02:20.760 of the year was when Jay Graber, who's the CEO 0:02:20.919 --> 0:02:24.200 of blue Sky, the kind of left leaning social media platform, 0:02:24.520 --> 0:02:28.320 wore a T shirt at south By Southwest that said 0:02:28.680 --> 0:02:32.480 in Latin, no more Caesars, and it was a real 0:02:32.639 --> 0:02:35.880 sort of shots fired at Zuck, I think, and. 0:02:35.800 --> 0:02:38.080 Also just again in Latin. I mean, these are the 0:02:38.120 --> 0:02:39.720 nerdiest people in human history. 0:02:39.919 --> 0:02:42.280 This was a huge best seller for Blue Sky. I 0:02:42.320 --> 0:02:43.359 saw that almost immediately. 0:02:43.440 --> 0:02:46.720 I bet it was, you know, it's tech Titans, as 0:02:46.760 --> 0:02:48.840 I was saying, just seemed to want to have a 0:02:48.880 --> 0:02:51.160 signature look. And I don't know if part of it 0:02:51.240 --> 0:02:58.240 is pathology, psychological pathology, I mean, certainly minus psychological pathology. 0:02:58.280 --> 0:03:00.040 I like to wear the same thing every day. I 0:03:00.040 --> 0:03:02.600 don't have to think about it. Like our friend Elizabeth Holmes, 0:03:02.600 --> 0:03:04.720 who also rocked the Black Turble. 0:03:04.560 --> 0:03:05.960 She flex hard on the Steve Jobs. 0:03:06.200 --> 0:03:09.800 She flexed and copped. She coughed something from Steve Jobs. 0:03:09.840 --> 0:03:11.360 And I guess it's sort of worked for her because 0:03:11.360 --> 0:03:12.440 we're still talking about it. 0:03:12.480 --> 0:03:14.600 But I think the most sort of talked about tech 0:03:14.680 --> 0:03:17.160 CEO of the moment is none other than Jensen Hwang 0:03:17.320 --> 0:03:18.840 of Invidia. 0:03:18.520 --> 0:03:20.040 The man who made a lot of people a lot 0:03:20.040 --> 0:03:21.839 of money and now is making a lot of people 0:03:21.880 --> 0:03:22.800 a little less money. 0:03:22.720 --> 0:03:25.600 And still still a lot, but little less exactly. He 0:03:25.760 --> 0:03:29.079 has a signature. Look, can you describe it. 0:03:29.080 --> 0:03:32.120 It's sort of a Marlon Brando, sort of on the 0:03:32.160 --> 0:03:36.480 waterfront black leather jacket that is I think Tom Ford. 0:03:36.680 --> 0:03:38.440 I think it is Tom forty, which is like a 0:03:38.480 --> 0:03:39.320 ten thousand dollars. 0:03:39.320 --> 0:03:41.440 That's an extremely expensive jacket. 0:03:41.680 --> 0:03:43.960 But there's a lot of Invidia stands. It turns out, 0:03:44.080 --> 0:03:46.880 and if you build it, they will come to buy 0:03:46.920 --> 0:03:48.120 a fake black leather jacket. 0:03:48.360 --> 0:03:49.640 If you build it, they will come and buy it 0:03:49.640 --> 0:03:50.960 a lot cheaper exactly. 0:03:51.080 --> 0:03:53.800 So our friends at four or for Media reported that 0:03:53.840 --> 0:03:56.920 there are all these websites popping up with knockoff Jensen 0:03:56.960 --> 0:04:01.240 Hwang in video jackets for which have creative titles like 0:04:01.320 --> 0:04:05.960 Jensen wang in Video Jacket totally optimized their CEO. 0:04:06.080 --> 0:04:07.960 People, They're going to find my phone and be like 0:04:08.160 --> 0:04:09.960 that was the last thing that she was looking at. 0:04:10.960 --> 0:04:12.920 But I think I don't sure if you saw it. 0:04:12.920 --> 0:04:15.600 But this week Jensen Hwang went to visit a company 0:04:15.720 --> 0:04:19.120 called x one Robots, and they had one of their 0:04:19.200 --> 0:04:22.279 robots come up to Jensen whanging humanoid robot and present 0:04:22.360 --> 0:04:25.120 him with a new black leather jacket, not a tom 0:04:25.240 --> 0:04:29.000 Ford one. This one was bedazzled with the Nvidia stock 0:04:29.080 --> 0:04:32.159 ticker on the front left pocket and the logo in 0:04:32.240 --> 0:04:33.920 kind of Swarovsky style crystals on. 0:04:33.839 --> 0:04:36.839 The back to imagine if the humanoid robot is also sentient, 0:04:37.080 --> 0:04:39.120 if the robot was like, I don't want to do this, 0:04:40.640 --> 0:04:43.040 like the people programming mean, please give me something better 0:04:43.080 --> 0:04:43.360 to do. 0:04:44.320 --> 0:04:46.599 But there's a more serious story about fashion AI here, 0:04:46.680 --> 0:04:49.240 and it comes from The Guardian. So according to their story, 0:04:49.320 --> 0:04:51.520 the clothing retailer H and M announced that they're going 0:04:51.560 --> 0:04:55.200 to create so called AI twins of thirty models, which 0:04:55.200 --> 0:04:57.400 they're going to use for social media marketing posts. 0:04:57.440 --> 0:04:59.320 The way, I wish I had this for a podcast. 0:04:59.480 --> 0:05:02.800 As much as I enjoy doing it with you, I wish. 0:05:02.680 --> 0:05:06.760 You could just have a I'd have a relic. Would 0:05:06.760 --> 0:05:08.560 you rather use your replica or my replica? 0:05:09.080 --> 0:05:10.480 Well, if we wanted me to be on time, I 0:05:10.480 --> 0:05:13.919 would use your replica. But sorry, keep going, because I 0:05:13.920 --> 0:05:14.600 do love the story. 0:05:14.720 --> 0:05:17.960 No, So these models, the real models, have given their 0:05:17.960 --> 0:05:20.719 permission to H and M to use their likeness with AI. 0:05:21.600 --> 0:05:23.359 You know again, as you said, sounds good, you know, 0:05:23.400 --> 0:05:25.880 why why sharp yourself you can send your digital twin. 0:05:26.160 --> 0:05:28.880 Well that and also I think there's something that's very 0:05:28.920 --> 0:05:31.520 interesting here that reminds me of the Atlantic sort of 0:05:31.520 --> 0:05:34.400 selling its data, which is these are people that, like 0:05:34.480 --> 0:05:37.839 in the film industry, are cannibalizing their own job. 0:05:39.560 --> 0:05:42.680 Well, of course, just like with the Atlantic and Open Ai, 0:05:43.200 --> 0:05:46.320 where open Ai actually compensates the Atlantic for using its archive, 0:05:46.880 --> 0:05:50.280 these models are also being compensated for their image. They 0:05:50.279 --> 0:05:52.920 own the rights of their twins, and their twins can 0:05:52.920 --> 0:05:55.880 work for other brands, not just H and M, and 0:05:56.160 --> 0:05:59.440 they'll get paid. One of the thirty models said, quote, 0:05:59.600 --> 0:06:01.320 she's like me without the jet lag. 0:06:04.320 --> 0:06:06.400 So the other thing is, and it's worth saying, the 0:06:06.440 --> 0:06:10.320 images in the business of fashion right up look amazingly similar. 0:06:10.360 --> 0:06:11.479 Like it's not distinguishable. 0:06:11.640 --> 0:06:13.520 Yeah, it's not like who was that person that we 0:06:13.600 --> 0:06:16.880 used to cover on Sleepwalkers. She was a model Lil Michayla. 0:06:17.000 --> 0:06:20.240 Oh yeah, like littl Michayla looked like AI. I think 0:06:20.279 --> 0:06:22.760 you know, after four years, we have now gotten to 0:06:22.800 --> 0:06:27.000 a place where it's just an AI replica, which is incredible. 0:06:27.200 --> 0:06:30.480 Yeah, and there's no sick finger. As you pointed out, 0:06:30.640 --> 0:06:32.640 it does raise concerns about the future of modeling in 0:06:32.640 --> 0:06:36.479 the fashion industry because although these models who are participation 0:06:36.520 --> 0:06:38.840 in the partnership will get paid every time that digital 0:06:38.839 --> 0:06:41.279 twin is used, you know, just like in the film industry, 0:06:41.320 --> 0:06:43.920 also raises questions about what happens to all the people 0:06:43.960 --> 0:06:45.840 who work on sets. I mean the hair siders, the 0:06:45.880 --> 0:06:47.400 makeup artist, the lighting designers. 0:06:47.560 --> 0:06:49.840 Yeah, you know, I was actually talking to our producer Eliza. 0:06:49.920 --> 0:06:52.240 Shout out to Eliza, I like to keep her in 0:06:52.279 --> 0:06:55.360 the mix. And she actually knows a hairstylist who works 0:06:55.360 --> 0:06:58.360 with a company that's testing out replacing many of their 0:06:58.520 --> 0:07:02.640 e commerce models with AI, and you know, work has 0:07:02.680 --> 0:07:04.080 slowed way down for her. 0:07:04.480 --> 0:07:05.239 So this is happening. 0:07:05.560 --> 0:07:07.560 I mean, think about it. You don't need to be 0:07:07.640 --> 0:07:09.239 doing a blood on a digital avatar. 0:07:09.720 --> 0:07:11.800 Yeah, and I mean even for models themselves, if I 0:07:11.800 --> 0:07:13.640 guess if you're you know, for these thirty who have 0:07:13.960 --> 0:07:16.360 gotten first through the door. It's one thing, but this 0:07:16.440 --> 0:07:20.000 may affect the kind of future job landscape for model 0:07:20.040 --> 0:07:21.680 as well as people work around modeling. 0:07:22.520 --> 0:07:25.360 But this is like a This is Zulander three where 0:07:25.400 --> 0:07:30.800 Derek realizes we're in jeopardy. But in all seriousness, you know, 0:07:30.880 --> 0:07:34.120 what what does it look like for models who are 0:07:34.120 --> 0:07:36.120 no longer in demand? And it might not seem important 0:07:36.160 --> 0:07:39.720 to like a lay person, but I do think it's 0:07:39.760 --> 0:07:42.440 a harbinger of things to come, which is like, if 0:07:42.480 --> 0:07:45.800 you can replace something that is so sort of ubiquitous 0:07:45.840 --> 0:07:49.160 for you know, a century, what does that look like? 0:07:49.600 --> 0:07:51.840 Yeah, I mean wile I find most interesting about these 0:07:51.880 --> 0:07:54.520 like AI digital twins or AI actors or whatever, or 0:07:55.040 --> 0:07:58.360 you know, chatbots of famous people from history is their 0:07:58.400 --> 0:08:01.440 effect is kind of to lock in the very few 0:08:01.520 --> 0:08:04.480 most famous people in the world as the only characters 0:08:04.520 --> 0:08:07.480 who are worth interacting with. I mean, you know, if 0:08:07.480 --> 0:08:10.280 you think about an action movie, why would you not 0:08:10.280 --> 0:08:12.120 make it with Tom Cruise? Or talking about doing a 0:08:12.120 --> 0:08:14.240 fashion shoot, why would you not make it with el McPherson. 0:08:14.280 --> 0:08:16.600 So kind of, I think there's a longer term kind 0:08:16.600 --> 0:08:19.080 of chilling effect on the pipeline and new talent in 0:08:19.160 --> 0:08:22.480 creative industries, which you know, which will pretty interesting, you know, 0:08:22.560 --> 0:08:24.960 and somewhat disturbing. To be fair to H and M, 0:08:24.960 --> 0:08:27.200 they're being very upfront about their use of AI. They're 0:08:27.200 --> 0:08:29.600 going to watermark the images of the AI twins in 0:08:29.640 --> 0:08:32.600 their ads so people will know they're the AI versions, 0:08:33.120 --> 0:08:35.160 And by doing this, the company will also be abiding 0:08:35.240 --> 0:08:37.959 with the EU's AI Act, which is coming into effect 0:08:38.000 --> 0:08:41.000 in twenty twenty six, and it will require all AI 0:08:41.160 --> 0:08:44.000 images to be labeled as AI images, to. 0:08:44.000 --> 0:08:47.000 Which I say, who's looking for that? Like, I you 0:08:47.040 --> 0:08:48.480 know what I mean? Like the role of AI and 0:08:48.559 --> 0:08:50.800 creative industries, which you know, is something I'm obsessed with 0:08:50.800 --> 0:08:53.040 and sort of how to regulate it is something that's 0:08:53.080 --> 0:08:56.160 obviously going to keep coming up, and I'm curious to 0:08:56.160 --> 0:08:58.560 see where it goes next. I wonder to what extent 0:08:58.720 --> 0:09:02.680 the lay person cares that, Like where they're buying a shirt, 0:09:03.120 --> 0:09:06.240 is either modeling that shirt with a fake twin or 0:09:06.280 --> 0:09:07.440 the real person. 0:09:07.320 --> 0:09:10.880 Probably not, Probably not. The next headline comes from the 0:09:10.880 --> 0:09:14.439 stuff of nightmares for anyone who's a frequent traveler, and 0:09:14.520 --> 0:09:16.520 it has to do me, and it has to do 0:09:16.600 --> 0:09:18.720 with an airport being shut down for twenty four hours. 0:09:18.720 --> 0:09:19.520 I heard it was Heathrow. 0:09:19.640 --> 0:09:21.880 It was London heath Throw Airport, where I flew out 0:09:21.880 --> 0:09:24.559 of just this week. It shut down, leading to over 0:09:24.600 --> 0:09:28.559 a thousand council flights after a fire caused a power outage, 0:09:28.880 --> 0:09:31.400 and Bloomberg reported that the outage could be traced back 0:09:31.440 --> 0:09:35.559 to a single point of failure, a burned transformer and 0:09:35.640 --> 0:09:39.800 twenty five thousand liters of transformer cooling oil that was 0:09:39.840 --> 0:09:42.760 a blaze for several hours. It's a fascinating story, and 0:09:42.800 --> 0:09:45.640 Bloomberg had such a great headline, which was the device 0:09:45.920 --> 0:09:50.000 throttling the world's electrified future. But Kara, do you know 0:09:50.040 --> 0:09:51.120 what a transformer is? 0:09:52.120 --> 0:09:54.439 I know what the movie Transformers is. I know it's 0:09:54.440 --> 0:09:55.880 a car that turns into a man. 0:09:56.480 --> 0:10:00.000 So most simply put, a transformer is something that changed 0:10:00.320 --> 0:10:04.000 is voltage. So when you create electricity in a power plant, 0:10:04.320 --> 0:10:06.600 you actually want to increase the voltage as much as 0:10:06.640 --> 0:10:09.040 you can because that allows it to travel far and 0:10:09.200 --> 0:10:12.480 fast with less loss of electricity along the way. So 0:10:12.920 --> 0:10:15.360 use a transformer on the way out of the power plant. 0:10:15.640 --> 0:10:17.240 But then when it gets to your home or the 0:10:17.280 --> 0:10:19.400 local electric grid or whatever it may be, you actually 0:10:19.400 --> 0:10:21.880 want to use a transformer to turn the voltage back 0:10:21.920 --> 0:10:24.200 down because otherwise it blows up your stuff. 0:10:24.559 --> 0:10:27.720 So ostensibly, it's sort of like when I'm staying at 0:10:28.600 --> 0:10:33.280 not the greatest hotel and I turn a blow dryer 0:10:33.360 --> 0:10:36.079 on and all of a sudden, the entire room short 0:10:36.120 --> 0:10:38.640 circuits and all of the electricity goes off because the 0:10:38.720 --> 0:10:39.760 voltage is too high. 0:10:39.960 --> 0:10:42.080 Yeah, exactly mean that is a short circuit where there's 0:10:42.200 --> 0:10:45.360 likely been a problem with the transformer being bypassed or 0:10:45.400 --> 0:10:46.319 not functioning correctly. 0:10:46.480 --> 0:10:48.000 Right. And I think that in a storm or a 0:10:48.080 --> 0:10:51.840 natural disaster, they can sometimes explode and it's very loud. 0:10:51.920 --> 0:10:54.160 Yeah, it sounds like fireworks or a bomb going off. 0:10:54.200 --> 0:10:56.880 And I mean during hurricanes and other kind of natural disasters, 0:10:56.920 --> 0:10:59.160 these things come under a lot of pressure when they 0:10:59.160 --> 0:11:01.800 do go out. Follow And so what happened to Heathrow 0:11:01.840 --> 0:11:04.760 Airport was this fire broke out in a substation which 0:11:04.760 --> 0:11:08.240 houses transformers, and it took the firefighters seven hours to 0:11:08.240 --> 0:11:10.439 get it under control. The airport, in order to come 0:11:10.440 --> 0:11:13.280 back online, was able to accept power from other substations, 0:11:13.760 --> 0:11:15.199 but even the time it took them to do that, 0:11:15.360 --> 0:11:18.800 many many flights were canceled. It was chaos. 0:11:19.840 --> 0:11:22.080 She made it, but it was not a pretty team. 0:11:22.160 --> 0:11:25.200 Yeah him. Back in twenty thirteen, there was actually a 0:11:25.240 --> 0:11:29.400 sniper attack on a substation in California which caused a 0:11:29.440 --> 0:11:32.800 fire that burned seventeen transformers and almost knocked out all 0:11:32.800 --> 0:11:35.480 of the power of Silicon Valley. Now This actually led 0:11:35.520 --> 0:11:38.760 to a lot more security around substations and even stockpiling 0:11:38.800 --> 0:11:43.360 of transformers, which of course sparked another problem, a shortage 0:11:43.360 --> 0:11:46.679 of transformers, transformer shortage and obviously with the supply chain 0:11:46.720 --> 0:11:51.160 issues recently, the lead time for delivering a new large 0:11:51.200 --> 0:11:55.240 transformer is now about three to five years for a 0:11:55.280 --> 0:11:58.520 single transformer, and bear in mind these can be huge. Nonetheless, 0:11:58.559 --> 0:12:00.800 the scale of transformer that was Heathrow, it takes a 0:12:00.840 --> 0:12:03.679 long hundre replace and they've got a much more expensive 0:12:03.760 --> 0:12:05.520 So you know, now we're living in the error of 0:12:05.559 --> 0:12:09.120 EVS and the AI boom powered by data centers and 0:12:09.280 --> 0:12:12.720 transformers are also required to bring renewable power onto the grid, 0:12:13.080 --> 0:12:15.040 and so you know, one of the interesting implications of 0:12:15.080 --> 0:12:18.480 this story, as Imaza, who's on the show, recently told 0:12:18.520 --> 0:12:21.960 me that actually the US's struggles to onboard new power 0:12:21.960 --> 0:12:24.880 to the grid maybe the reason why the US ultimately 0:12:24.880 --> 0:12:26.280 falls behind China in AI. 0:12:26.840 --> 0:12:29.840 So it's ironic actually that the increase in transformer prices 0:12:29.880 --> 0:12:33.160 could be further increased by President Trump's on again, off 0:12:33.200 --> 0:12:36.400 again relationship with imposing tariffs in Canada and Mexico, which 0:12:36.440 --> 0:12:39.520 is where we import a lot of our large transformers from. 0:12:39.520 --> 0:12:41.320 And that's why I love this Bloomberg story because it's 0:12:41.320 --> 0:12:43.680 fun think about all the sexy stuff like new chips 0:12:43.679 --> 0:12:46.640 and data center construction, you know, but there's still this 0:12:46.679 --> 0:12:49.920 one hundred year old technology that hasn't changed and that 0:12:50.040 --> 0:12:53.960 has to be imported and which is absolutely critical for infrastructure, 0:12:54.040 --> 0:12:55.760 digital and otherwise, you know. 0:12:55.720 --> 0:12:58.840 Speaking of the AI boom and things that use a lot, 0:12:58.880 --> 0:13:01.000 a lot, a lot of energy. This next story that 0:13:01.000 --> 0:13:02.920 I want to tell you is for those people who've 0:13:02.920 --> 0:13:06.679 been sitting here thinking about lms and not really understanding 0:13:06.679 --> 0:13:11.040 how they work. I have news for you. Much like 0:13:11.080 --> 0:13:17.040 Trump's decisions on tariffs, nobody knows how they work. And Anthropic, 0:13:17.120 --> 0:13:19.680 which is the same company that makes the AI model Claud, 0:13:20.000 --> 0:13:22.840 has been trying to figure out how the hell these things. 0:13:22.840 --> 0:13:25.040 This is to solve the so called black books. 0:13:24.720 --> 0:13:27.040 The black box problem. So Anthropic, which is the same 0:13:27.080 --> 0:13:29.679 company that makes the A model Claud, has been trying 0:13:29.720 --> 0:13:32.880 to figure out sort of what's under the hood, and 0:13:32.960 --> 0:13:36.679 they recently released two reports on how llms do things 0:13:36.720 --> 0:13:42.800 like complete sentences, solve math problems, and suppress hallucinations. And 0:13:42.840 --> 0:13:45.840 they use a technique called circuit tracing which let them 0:13:45.920 --> 0:13:49.520 track an LM's decision making process for ten different tasks 0:13:49.800 --> 0:13:52.680 by working back from the solution to the query. 0:13:53.040 --> 0:13:56.280 Huh okay, that makes sense. But I just think more broadly, 0:13:56.480 --> 0:13:59.360 as more and more decisions are taken for us by 0:13:59.679 --> 0:14:03.640 AI or outcomes kind of determined by AI, it's kind 0:14:03.640 --> 0:14:06.600 of remarkable that this huge elephant is still in the 0:14:06.640 --> 0:14:09.560 corner of the room, which is we can't understand how 0:14:09.559 --> 0:14:10.520 they make their decisions. 0:14:10.640 --> 0:14:12.560 But it also makes you think of like, I don't 0:14:12.600 --> 0:14:14.840 really know. I mean, I know how a car works, 0:14:15.000 --> 0:14:19.640 but like the I'm not an engineer, and yet increasingly, 0:14:19.880 --> 0:14:22.600 I mean, you know, since whenever the nineteen twenties, people 0:14:22.640 --> 0:14:24.480 have used cars more and more to get around, and 0:14:24.480 --> 0:14:26.760 we're just kind of like, well, this thing's gonna work 0:14:26.800 --> 0:14:29.160 until it blows up, you know. That's why I feel whatever. 0:14:29.400 --> 0:14:33.680 But it's important to note that lllm's, like Claude, which 0:14:33.760 --> 0:14:37.040 was the focus of this study, are trained, which I 0:14:37.080 --> 0:14:40.400 thought this was really interesting. They're trained, not programmed on 0:14:40.440 --> 0:14:43.520 a bunch of data. They create their own rules based 0:14:43.560 --> 0:14:46.600 on the data they ingest. But up until now we 0:14:46.760 --> 0:14:49.560 haven't been able to see into the models to know 0:14:49.600 --> 0:14:52.320 what those rules actually are. Let alone how the models 0:14:52.400 --> 0:14:53.080 generate them. 0:14:53.240 --> 0:14:55.800 Yeah, and I think what the work the anthropic is 0:14:55.840 --> 0:14:59.880 doing is all about basically understanding decision making. So then 0:15:00.120 --> 0:15:02.040 not yet at the stage of being able to understand 0:15:02.040 --> 0:15:04.480 how the models generate their own rules, But I think 0:15:04.520 --> 0:15:07.720 this story is all about how it's starting to become 0:15:08.520 --> 0:15:12.480 easier or possible to basically backfigure out how a model 0:15:12.480 --> 0:15:13.240 has made a decision. 0:15:13.400 --> 0:15:18.520 Yes, and it's not that simple. I think the researchers 0:15:18.560 --> 0:15:21.880 in this particular case were inspired by brain scan techniques 0:15:22.240 --> 0:15:25.360 that are used in neuroscience. They found that llms, which 0:15:25.400 --> 0:15:28.520 again I'm so interested in how we anthropomorphize lms, But 0:15:28.840 --> 0:15:33.080 they found that llms store different constellations of knowledge in 0:15:33.120 --> 0:15:35.400 different parts of their model. So, for example, the concept 0:15:35.440 --> 0:15:39.800 of smallness or the idea of a rabbit. Anthropic was 0:15:39.840 --> 0:15:43.560 actually able to identify and then turn off certain parts 0:15:43.600 --> 0:15:47.120 of the model. So like the idea rabbit and tune 0:15:47.120 --> 0:15:49.160 it down so that it couldn't be part of a 0:15:49.240 --> 0:15:50.160 queer result. 0:15:50.360 --> 0:15:52.760 And for exac like what eats carrots? It would be like. 0:15:54.080 --> 0:15:58.080 People right, a dog right exactly, And so the same 0:15:58.200 --> 0:16:01.480 query would have a different answer if the rabbit part 0:16:01.520 --> 0:16:03.280 of the model was dealt up or down, which you. 0:16:03.280 --> 0:16:06.720 Say said by contrast, you might you might say, you know, 0:16:07.000 --> 0:16:09.000 what's a mammal? It always asks rabbit if you dial 0:16:09.000 --> 0:16:10.240 it up, and if you if you dial it down, 0:16:10.240 --> 0:16:11.120 it would never own a rabbit. 0:16:11.280 --> 0:16:11.640 Correct. 0:16:11.720 --> 0:16:13.680 So I mean, and it's similar. I mean it's similar 0:16:13.680 --> 0:16:15.960 in human beings, which is like I don't drink anymore, 0:16:15.960 --> 0:16:17.560 so I go to a bar and I'm like water. 0:16:19.520 --> 0:16:20.680 It's a similar kind of thing. 0:16:21.120 --> 0:16:23.080 I mean that that actually clarifies for me, and I 0:16:23.240 --> 0:16:25.520 kind of We've done a bunch of coverage of AI 0:16:25.600 --> 0:16:28.000 and spoken to Jeffrey Hinton and others, but like that 0:16:28.120 --> 0:16:31.640 idea of a neural network is to me really clarified 0:16:31.640 --> 0:16:32.280 by this study. 0:16:32.400 --> 0:16:34.080 Well, and how could it not be. It's something that 0:16:34.080 --> 0:16:36.000 we've created, you know what I mean, it's going to 0:16:36.040 --> 0:16:40.400 reflect a sort of human centric way of working. I 0:16:40.440 --> 0:16:44.480 think this story actually came from the MIT Technology Review 0:16:44.480 --> 0:16:47.800 with the headline anthropic can now track the bizarre inner 0:16:47.840 --> 0:16:51.280 workings of a large language model, to which I said, Now, 0:16:52.480 --> 0:16:54.280 what really blew my mind, though, is the way that 0:16:54.360 --> 0:16:57.280 this thing solves math problems, because it's not the way 0:16:57.280 --> 0:16:58.200 that humans do math. 0:16:58.320 --> 0:16:59.400 Okay, tell me more. 0:16:59.360 --> 0:17:04.400 So, research asked it to solve the equation thirty six 0:17:04.480 --> 0:17:07.640 plus fifty nine, and while Claude came up with there, 0:17:07.760 --> 0:17:13.880 what's the correct answer? Well, well, Claude took a very 0:17:13.920 --> 0:17:17.440 circuitous route to get that answer. Anthropic found that Claude 0:17:17.520 --> 0:17:22.280 used multiple computation paths in parallel to get its final answer, 0:17:22.359 --> 0:17:25.280 unlike you, who just used your brain. So one path 0:17:25.359 --> 0:17:28.720 added a bunch of numbers. I love this. This is 0:17:28.760 --> 0:17:31.600 like nerd alert close to thirty six and fifty nine 0:17:31.720 --> 0:17:36.000 to approximate the total like thirty five and sixty, while 0:17:36.160 --> 0:17:40.920 another path focused on determining the last digit of the sum, 0:17:41.040 --> 0:17:43.879 so it actually added the last two digits of thirty 0:17:43.880 --> 0:17:47.040 six and fifty nine, six and nine to know that 0:17:47.119 --> 0:17:50.200 the answer had to be a multiple of five. So 0:17:50.240 --> 0:17:52.480 Claude used these two paths to come up with the 0:17:52.520 --> 0:17:54.800 correct answer, which is as you said. 0:17:54.800 --> 0:17:58.320 Ninety five. You know, I find particularly fascinating about this, 0:17:58.760 --> 0:18:01.320 so I always struggled to be with math. My dad, 0:18:01.359 --> 0:18:04.679 on the other hand, is a crazy math murder. He 0:18:04.800 --> 0:18:07.880 was like the under thirteen chess champion of Britain, etc. 0:18:08.840 --> 0:18:10.560 You know, he said to me about the most important 0:18:10.640 --> 0:18:14.200 thing about how to do math well, is how approximate 0:18:14.240 --> 0:18:16.480 the answer before you work it out. He basically told 0:18:16.560 --> 0:18:20.000 me to do exactly what this model does. Basically break 0:18:20.000 --> 0:18:22.320 it down to much simpler calculation, and then when you 0:18:22.359 --> 0:18:24.560 do the actual work, you will know whether or not 0:18:24.600 --> 0:18:25.359 you're in the range. 0:18:25.440 --> 0:18:25.640 Right. 0:18:25.760 --> 0:18:28.919 Well, Claude does math like your dad. When asked by 0:18:28.960 --> 0:18:31.639 a user how it got the answer ninety five, it 0:18:31.680 --> 0:18:34.840 claims to do it by the book, for example, simple addition, 0:18:35.000 --> 0:18:38.760 carrying the one and one of the explanations anthropic positive. 0:18:38.440 --> 0:18:41.200 Positive for the fact that, when asked how it got 0:18:41.200 --> 0:18:44.280 to the answer, did it shared a response that was not, 0:18:44.320 --> 0:18:45.840 in fact, how it got to the answer. 0:18:45.760 --> 0:18:49.560 Right because Claude's written answer and I quote may reflect 0:18:49.560 --> 0:18:52.920 the fact that the model learns to explain math by 0:18:53.000 --> 0:18:56.800 simulating explanations written by people. So when asked to do 0:18:56.880 --> 0:19:00.120 math without being taught how to do it, it may 0:19:00.119 --> 0:19:04.600 develop its own internal strategies to do so. Remember it's 0:19:04.640 --> 0:19:05.720 trade not programmed. 0:19:12.280 --> 0:19:14.359 We're going to take a quick break, but stick around 0:19:14.520 --> 0:19:17.439 and we'll be back with this week's tech support, all 0:19:17.480 --> 0:19:34.240 about AI competition among tech giants. For our next segment, 0:19:34.240 --> 0:19:37.719 we're going to be talking about all things AI. Surprise, Surprise, 0:19:37.760 --> 0:19:41.280 on tech stuff Rise tell surprise, But in all seriousness, 0:19:41.400 --> 0:19:43.960 there's just so much happening in AI all the time. 0:19:44.040 --> 0:19:45.760