1 00:00:02,400 --> 00:00:06,760 Speaker 1: Bloomberg Audio Studios, podcasts, radio news. 2 00:00:08,480 --> 00:00:11,119 Speaker 2: I'm Stephen Carol and this is Here's Why, where we 3 00:00:11,160 --> 00:00:13,320 Speaker 2: take one news story and explain it in just a 4 00:00:13,320 --> 00:00:15,840 Speaker 2: few minutes with our experts here at Bloomberg. 5 00:00:18,800 --> 00:00:21,400 Speaker 3: The development we've seen with Anthropic and Metals is a 6 00:00:21,440 --> 00:00:26,000 Speaker 3: good example of a responsible company that is suddenly thinking, ah, 7 00:00:26,120 --> 00:00:28,320 Speaker 3: that could be really good, but if it falls in 8 00:00:28,360 --> 00:00:30,080 Speaker 3: the wrong hand, could be really bad. 9 00:00:30,240 --> 00:00:33,879 Speaker 4: The issue is to what extent is this new version 10 00:00:34,280 --> 00:00:37,840 Speaker 4: of the product going to be able to, in a sense, 11 00:00:38,080 --> 00:00:42,960 Speaker 4: identify vulnerabilities in other systems which can be exploited for 12 00:00:43,000 --> 00:00:48,880 Speaker 4: cyber attack purposes. This Anthropic Mythos model was a step 13 00:00:48,920 --> 00:00:53,440 Speaker 4: function change in abilities, learning capabilities. 14 00:00:53,520 --> 00:00:57,520 Speaker 1: I would put less emphasis specifically on this release of Mythos, 15 00:00:57,520 --> 00:01:00,000 Speaker 1: but really acknowledge that the frontier is moving, and it's 16 00:01:00,080 --> 00:01:02,480 Speaker 1: moving much more quickly than security teams. 17 00:01:02,200 --> 00:01:04,520 Speaker 2: Are able to keep up. We've been told for years 18 00:01:04,560 --> 00:01:08,680 Speaker 2: about the transformative effect of artificial intelligence, but what happens 19 00:01:08,720 --> 00:01:12,319 Speaker 2: when AI developers find their creations have gone too far. 20 00:01:13,000 --> 00:01:16,399 Speaker 2: Anthropic held back a general release of its Mythos tool 21 00:01:16,680 --> 00:01:20,039 Speaker 2: over fears it would help hackers steal data or disrupt 22 00:01:20,120 --> 00:01:23,840 Speaker 2: critical infrastructure. The worries were enough for the US Treasury 23 00:01:23,840 --> 00:01:26,640 Speaker 2: and Federal Reserve to summon Wall Street Bank bosses to 24 00:01:26,680 --> 00:01:29,600 Speaker 2: discuss it. What does it mean for the development and 25 00:01:29,680 --> 00:01:33,480 Speaker 2: regulation of AI in the future. Here's why some AI 26 00:01:33,840 --> 00:01:39,240 Speaker 2: might be too dangerous to release. Bomberg opinion columnist parme 27 00:01:39,319 --> 00:01:42,160 Speaker 2: Olsen joins us now for more, Permi. You had a 28 00:01:42,160 --> 00:01:45,600 Speaker 2: great piece recently about why you're skeptical of these kinds 29 00:01:45,600 --> 00:01:47,760 Speaker 2: of warnings from AI companies. 30 00:01:47,800 --> 00:01:50,600 Speaker 5: Tell us why, Yes, I am a little bit skeptical 31 00:01:50,640 --> 00:01:52,320 Speaker 5: of some of these warnings. And part of the reason 32 00:01:52,400 --> 00:01:56,600 Speaker 5: is because we have long had a history of AI 33 00:01:56,720 --> 00:02:00,880 Speaker 5: leaders using what I would call kind of paradox marketing 34 00:02:00,920 --> 00:02:05,480 Speaker 5: strategy of talking up the dangers of the technology that 35 00:02:05,520 --> 00:02:08,160 Speaker 5: they are selling. So a couple of years ago, not 36 00:02:08,280 --> 00:02:11,520 Speaker 5: long after the release of chat GPT, the thing that 37 00:02:11,639 --> 00:02:13,880 Speaker 5: AI leaders were sometimes. 38 00:02:13,480 --> 00:02:16,359 Speaker 1: Talking about was the civilizational. 39 00:02:15,919 --> 00:02:19,359 Speaker 5: Risk of AI, that it could eventually become so powerful 40 00:02:19,400 --> 00:02:22,040 Speaker 5: that it could take over our critical infrastructure and even 41 00:02:22,160 --> 00:02:25,440 Speaker 5: potentially wipe out humanity. And you might remember some tech 42 00:02:25,520 --> 00:02:28,960 Speaker 5: leaders signed a kind of moratorium on building AI systems 43 00:02:28,960 --> 00:02:32,760 Speaker 5: because they could be so dangerous. Nowadays, they're not talking 44 00:02:32,800 --> 00:02:36,000 Speaker 5: about that so much anymore. They're talking more about the 45 00:02:36,800 --> 00:02:40,960 Speaker 5: doomsday risk to jobs and also in the case of 46 00:02:41,000 --> 00:02:45,160 Speaker 5: Mythos cybersecurity. One of the things that Anthropic CEO said 47 00:02:45,280 --> 00:02:47,600 Speaker 5: just last year and also earlier this year was that 48 00:02:47,800 --> 00:02:51,520 Speaker 5: half of all entry level white collar jobs would probably 49 00:02:51,560 --> 00:02:54,840 Speaker 5: disappear in the next five years because of AI. The 50 00:02:54,880 --> 00:02:58,640 Speaker 5: head of AI at Microsoft, Mustafasuleman, seemed to kind of 51 00:02:58,680 --> 00:03:02,079 Speaker 5: one up that by saying that all professional tasks would 52 00:03:02,080 --> 00:03:05,600 Speaker 5: be automated by AI in just eighteen months. These are 53 00:03:05,720 --> 00:03:09,200 Speaker 5: very scary numbers to throw out there, and you can 54 00:03:09,280 --> 00:03:12,200 Speaker 5: sort of arguably take them seriously because these people are 55 00:03:12,200 --> 00:03:14,920 Speaker 5: close to the technology. But actually, if you look at 56 00:03:14,960 --> 00:03:20,320 Speaker 5: the statistics from labor organizations from economic institutes, there's not 57 00:03:20,440 --> 00:03:24,440 Speaker 5: a lot of signs yet that labor markets are significantly 58 00:03:24,480 --> 00:03:27,800 Speaker 5: being disrupted by AI, and companies themselves say that they're 59 00:03:27,840 --> 00:03:31,760 Speaker 5: not necessarily seeing much of an impact on employment. Yes, 60 00:03:31,840 --> 00:03:34,840 Speaker 5: we are seeing some difficulties for young people to get 61 00:03:34,960 --> 00:03:38,520 Speaker 5: entry level jobs, but by and large, across the board, 62 00:03:38,560 --> 00:03:41,200 Speaker 5: there hasn't been that much impact. So I think in 63 00:03:41,240 --> 00:03:44,240 Speaker 5: a weird way, when companies like Anthropic and OpenAI have 64 00:03:44,360 --> 00:03:47,640 Speaker 5: IPOs coming up and it works in their interests to 65 00:03:47,800 --> 00:03:50,960 Speaker 5: talk about how powerful their systems are. This kind of 66 00:03:51,040 --> 00:03:53,600 Speaker 5: rhetoric can stand them in good stead. And just in 67 00:03:53,640 --> 00:03:57,000 Speaker 5: the case of Mythos, you were saying that Anthropic talked 68 00:03:57,000 --> 00:03:58,920 Speaker 5: about it but then chose not to release it. 69 00:03:59,000 --> 00:04:01,240 Speaker 1: We saw similar their tactic back. 70 00:04:01,040 --> 00:04:04,920 Speaker 5: In twenty nineteen when OpenAI announced GPT two and said 71 00:04:04,920 --> 00:04:06,960 Speaker 5: it was too dangerous to release, and so they held 72 00:04:07,000 --> 00:04:09,200 Speaker 5: it back. But then eventually they did release it and 73 00:04:09,280 --> 00:04:12,200 Speaker 5: nothing happened. But along the way they had generated quite 74 00:04:12,240 --> 00:04:14,000 Speaker 5: a lot of hype about this model. 75 00:04:14,320 --> 00:04:16,839 Speaker 2: Yeah, and it's very interesting to trace the evolution of 76 00:04:16,880 --> 00:04:19,920 Speaker 2: those warnings, as you say, from going from existential risk 77 00:04:20,040 --> 00:04:25,279 Speaker 2: to jobs risk to cybersecurity in this case, do we 78 00:04:25,360 --> 00:04:29,080 Speaker 2: actually know about Mythos and the risks associated with us 79 00:04:29,120 --> 00:04:31,000 Speaker 2: separately to what we've heard from the company. 80 00:04:31,440 --> 00:04:34,240 Speaker 5: I feel like as we were just discussing a little 81 00:04:34,240 --> 00:04:37,480 Speaker 5: bit of skepticism about the way that Anthropic has announced 82 00:04:37,480 --> 00:04:40,919 Speaker 5: this new model. I mean, why why even talk about 83 00:04:40,920 --> 00:04:44,720 Speaker 5: it publicly if it's so dangerous? Why produce a kind 84 00:04:44,720 --> 00:04:49,040 Speaker 5: of slick video of Anthropic executives talking about how powerful 85 00:04:49,080 --> 00:04:51,680 Speaker 5: this model is and then show that to the world. 86 00:04:52,200 --> 00:04:54,240 Speaker 5: But I think two things can be true at the 87 00:04:54,279 --> 00:04:58,160 Speaker 5: same time, based on what we heard from the UK's 88 00:04:58,200 --> 00:05:01,400 Speaker 5: AI Security Institute, which just over the years become the 89 00:05:01,440 --> 00:05:05,320 Speaker 5: sort of global neutral arbiter of what makes secure and 90 00:05:05,360 --> 00:05:09,440 Speaker 5: safe AI. Actually, Mythos is a step up in capabilities 91 00:05:09,480 --> 00:05:12,080 Speaker 5: compared to other AI models in terms of its ability 92 00:05:12,200 --> 00:05:17,080 Speaker 5: to engage in cyber attacks and help hackers do these attacks. 93 00:05:17,560 --> 00:05:19,560 Speaker 5: But at the same time, I think the framing is 94 00:05:19,839 --> 00:05:23,400 Speaker 5: very helpful to Anthropic in terms of the partnerships that 95 00:05:23,440 --> 00:05:25,799 Speaker 5: it's now creating with these companies to whom it's given 96 00:05:25,839 --> 00:05:29,520 Speaker 5: access to Mythos, and also the publicity that is generating. 97 00:05:30,320 --> 00:05:32,760 Speaker 1: But yes, I think one of. 98 00:05:32,760 --> 00:05:35,320 Speaker 5: The really interesting things that came out of this report 99 00:05:35,440 --> 00:05:38,800 Speaker 5: from the AI Security Institute was that it said, actually 100 00:05:39,160 --> 00:05:42,760 Speaker 5: the bigger problem of this model is for so called 101 00:05:42,839 --> 00:05:47,760 Speaker 5: weakly defended or simple systems, which actually I think means that, 102 00:05:47,800 --> 00:05:49,960 Speaker 5: in spite of what we've seen on Wall Street, the 103 00:05:49,960 --> 00:05:53,320 Speaker 5: Secretary of the Treasury bessnt bringing together all these banks 104 00:05:53,360 --> 00:05:57,559 Speaker 5: to discuss the dangers of this model. Actually, I don't 105 00:05:57,720 --> 00:06:01,200 Speaker 5: think banks are really the ones so much at risk 106 00:06:01,279 --> 00:06:04,160 Speaker 5: here from this kind of technology. It's smaller companies who 107 00:06:04,240 --> 00:06:07,800 Speaker 5: don't have the resources to put in secure measures in place. 108 00:06:08,160 --> 00:06:13,280 Speaker 5: We're talking hospitals, mom and pop shops, small e commerce businesses, 109 00:06:13,480 --> 00:06:17,400 Speaker 5: SMEs basically who I think really just don't have the 110 00:06:17,440 --> 00:06:20,040 Speaker 5: resources to invest in security that might be most at risk. 111 00:06:20,520 --> 00:06:23,359 Speaker 2: Is there any effective way, though, to place guardrails on 112 00:06:23,480 --> 00:06:27,400 Speaker 2: technology that's developing at such speed? We've gone through several 113 00:06:27,400 --> 00:06:31,000 Speaker 2: big iterations of technological development. I'm thinking about the rise 114 00:06:31,040 --> 00:06:33,400 Speaker 2: of social media, for example. Are there any lessons to 115 00:06:33,440 --> 00:06:35,760 Speaker 2: be learned about how we could do better this time around? 116 00:06:36,760 --> 00:06:39,440 Speaker 5: I think so, But I think governments need to be 117 00:06:39,640 --> 00:06:43,600 Speaker 5: much more involved than they currently are. And the difficulty 118 00:06:43,720 --> 00:06:47,880 Speaker 5: is with guardrails is that you say that to the 119 00:06:47,920 --> 00:06:51,359 Speaker 5: European Union, or the United States government, or the British government, 120 00:06:51,760 --> 00:06:55,599 Speaker 5: and that's like anathema to them, because actually the big 121 00:06:55,680 --> 00:06:59,440 Speaker 5: fear among governments isn't so much about the harmful consequences 122 00:06:59,440 --> 00:07:03,359 Speaker 5: of AI. It's about being left behind economically, as China 123 00:07:03,839 --> 00:07:08,200 Speaker 5: and other nations capitalized on artificial intelligence to make their 124 00:07:08,240 --> 00:07:10,840 Speaker 5: economies much more productive. So what we're actually seeing in 125 00:07:10,920 --> 00:07:15,360 Speaker 5: terms of government policies around AI is let's upskill ten 126 00:07:15,440 --> 00:07:19,120 Speaker 5: million British people to become better at AI so that 127 00:07:19,160 --> 00:07:22,400 Speaker 5: we can really harness the productivity benefits of this technology 128 00:07:22,880 --> 00:07:27,160 Speaker 5: and not so much rhetoric or policies around let's put 129 00:07:27,280 --> 00:07:30,520 Speaker 5: institutes or safety measures in place or guardrails in place 130 00:07:30,560 --> 00:07:33,080 Speaker 5: on this technology because there's this fear they're going to 131 00:07:33,200 --> 00:07:36,440 Speaker 5: hold companies back from actually exploiting it. 132 00:07:36,800 --> 00:07:40,000 Speaker 2: How does the euse the AI Acts, for example, measure 133 00:07:40,120 --> 00:07:43,040 Speaker 2: up when we're thinking about the potential for regulation, because 134 00:07:43,040 --> 00:07:45,360 Speaker 2: you know, tech regulations something the EU has done a 135 00:07:45,400 --> 00:07:45,800 Speaker 2: lot on. 136 00:07:46,840 --> 00:07:48,960 Speaker 5: Yeah, and the EU in a lot of ways is 137 00:07:49,000 --> 00:07:51,840 Speaker 5: a step towards that and not entirely a bad step. 138 00:07:52,160 --> 00:07:54,560 Speaker 1: There's a lot of debate about how effective. It is. 139 00:07:54,680 --> 00:07:57,520 Speaker 1: It's very broad, it's quite vague, it's. 140 00:07:57,400 --> 00:07:59,400 Speaker 5: Not specific where it's meant to be, or it's two 141 00:07:59,480 --> 00:08:02,200 Speaker 5: specific where it shouldn't be specific. And that's in part 142 00:08:02,240 --> 00:08:05,960 Speaker 5: because the people who drafted that law did so in 143 00:08:06,000 --> 00:08:08,360 Speaker 5: a rush right after Chad GBT came out. 144 00:08:08,360 --> 00:08:09,440 Speaker 1: They were sort of scrambling. 145 00:08:09,680 --> 00:08:12,480 Speaker 5: And there's an argument that actually, EU opin union has 146 00:08:12,520 --> 00:08:15,240 Speaker 5: other laws like the Digital Services Act the Digital Markets 147 00:08:15,280 --> 00:08:19,400 Speaker 5: Act that address some of the potential unintended consequences of AI. 148 00:08:19,600 --> 00:08:22,520 Speaker 5: But even so, you know, the AI Act from the 149 00:08:22,520 --> 00:08:24,440 Speaker 5: EU gets a lot of flack. But I still think 150 00:08:24,480 --> 00:08:27,520 Speaker 5: it was a good step forward by policy makers. There's 151 00:08:27,680 --> 00:08:29,880 Speaker 5: no intention, as far as I understand in the UK 152 00:08:30,280 --> 00:08:34,480 Speaker 5: to have any kind of similar legislation, but rather existing 153 00:08:34,520 --> 00:08:37,200 Speaker 5: acts like the Online Safety Act that we have that 154 00:08:37,280 --> 00:08:40,000 Speaker 5: will address some of the potential harmful consequences of AI. 155 00:08:40,400 --> 00:08:43,800 Speaker 5: I think perhaps one way we could see some progress 156 00:08:44,040 --> 00:08:47,280 Speaker 5: on this is the AI Security Institute here in London, 157 00:08:47,400 --> 00:08:49,600 Speaker 5: which again produced that report on mythos. If we can 158 00:08:49,679 --> 00:08:52,960 Speaker 5: see more activity from them, and more similar organizations like 159 00:08:52,960 --> 00:08:55,200 Speaker 5: that in other nations or other regions, I think that 160 00:08:55,240 --> 00:08:55,959 Speaker 5: would be a good thing. 161 00:08:56,520 --> 00:08:59,320 Speaker 2: Okay, parme Els and Bloomberg Opinion columnists thank you, and 162 00:08:59,360 --> 00:09:01,600 Speaker 2: you can read more are from parme also at Bloomberg 163 00:09:01,640 --> 00:09:05,280 Speaker 2: dot com forward slash opinion. For more explanations like this 164 00:09:05,400 --> 00:09:08,000 Speaker 2: from our team of three thousand journalists and analysts around 165 00:09:08,000 --> 00:09:12,520 Speaker 2: the world, go to Bloomberg dot com slash explainers. I'm 166 00:09:12,520 --> 00:09:15,520 Speaker 2: Stephen Carol. This is here's why. I'll be back next 167 00:09:15,520 --> 00:09:17,400 Speaker 2: week with more. Thanks for listening.