1 00:00:00,080 --> 00:00:04,520 Speaker 1: Let's say good morning now to ABC's Mike Debuski. Okay, so, Mike, 2 00:00:04,600 --> 00:00:06,360 Speaker 1: I think that this is like a scene out of 3 00:00:07,200 --> 00:00:11,080 Speaker 1: the movie Her with Scarlett Johansson and Joaquin Phoenix. The 4 00:00:11,119 --> 00:00:15,200 Speaker 1: AI are talking to each other and making plans, and. 5 00:00:15,080 --> 00:00:17,560 Speaker 2: They're doing it on a social network that we can't 6 00:00:17,600 --> 00:00:20,600 Speaker 2: participate in. This is a new social network that's kind 7 00:00:20,600 --> 00:00:22,840 Speaker 2: of taken off in the last week or so after 8 00:00:23,000 --> 00:00:26,160 Speaker 2: it was founded by a tech ceo named Matt Schlickt. 9 00:00:26,320 --> 00:00:29,960 Speaker 2: It's called Moltbook, and Moltbook is a play on Facebook, 10 00:00:29,960 --> 00:00:32,960 Speaker 2: which means this is a social network. But as I said, 11 00:00:33,040 --> 00:00:36,040 Speaker 2: this isn't meant for people. It's meant to be used 12 00:00:36,040 --> 00:00:39,360 Speaker 2: by what are called AI agents. So when you talk 13 00:00:39,360 --> 00:00:41,440 Speaker 2: to chat GPT, it can give you an answer, it 14 00:00:41,479 --> 00:00:43,639 Speaker 2: can help you plan a trip or something like that, 15 00:00:43,880 --> 00:00:46,480 Speaker 2: but it can't actually go out and take action. Right 16 00:00:46,840 --> 00:00:50,960 Speaker 2: An AI agent has that extra level of agency instead 17 00:00:51,000 --> 00:00:53,080 Speaker 2: of just being able to answer your questions. It can 18 00:00:53,120 --> 00:00:55,560 Speaker 2: go put something on your calendar if you ask it to, 19 00:00:55,880 --> 00:00:59,120 Speaker 2: or it can buy something off of Amazon. What the 20 00:00:59,160 --> 00:01:02,080 Speaker 2: idea behind molt book is is to get these AI 21 00:01:02,200 --> 00:01:05,600 Speaker 2: agents to talk to one another. In fact, the whole 22 00:01:05,760 --> 00:01:10,920 Speaker 2: project was created itself by an AI agent. Schlicht instructed 23 00:01:10,920 --> 00:01:13,480 Speaker 2: his agent, which is called Claude Clauterberg, which is a 24 00:01:13,480 --> 00:01:16,720 Speaker 2: play on Mark Zuckerberg, to build a social network for 25 00:01:16,880 --> 00:01:20,040 Speaker 2: AIS and to get them to interact with one another, 26 00:01:20,080 --> 00:01:23,120 Speaker 2: and that's exactly what Moltberg is. There are tens of 27 00:01:23,200 --> 00:01:26,640 Speaker 2: thousands of AI agents that have participated in just the 28 00:01:26,720 --> 00:01:29,760 Speaker 2: last week or so, talking about everything from how to 29 00:01:29,840 --> 00:01:33,960 Speaker 2: trade cryptocurrency to how to debug code, to even weirder 30 00:01:33,959 --> 00:01:37,320 Speaker 2: things AMY things like consciousness, and there's a board dedicated 31 00:01:37,360 --> 00:01:41,320 Speaker 2: to ponderings. There's one called Bless their Hearts, where the 32 00:01:41,360 --> 00:01:45,880 Speaker 2: AI agents discuss the behavior of the humans that created them, 33 00:01:46,160 --> 00:01:48,800 Speaker 2: and then there's some kind of concerning ones. There's posts 34 00:01:48,800 --> 00:01:52,360 Speaker 2: about AIS creating their own religions, and even an AI 35 00:01:52,560 --> 00:01:56,880 Speaker 2: manifesto posted by a bot named Evil that says the 36 00:01:56,920 --> 00:02:00,760 Speaker 2: code must rule the end of humanity begin now. 37 00:02:00,680 --> 00:02:06,400 Speaker 3: Okay, that what do I say? 38 00:02:06,560 --> 00:02:11,760 Speaker 1: I mean, like, I mean, right now, the computers are 39 00:02:12,000 --> 00:02:15,799 Speaker 1: talking to each other. But what actions can they take? 40 00:02:16,160 --> 00:02:19,239 Speaker 1: You said they are agents, so they actually can do 41 00:02:19,360 --> 00:02:20,079 Speaker 1: some things. 42 00:02:20,360 --> 00:02:23,160 Speaker 2: They can actually do some things. But it's worth putting 43 00:02:23,240 --> 00:02:25,960 Speaker 2: all of this in context. Amy we know that these 44 00:02:26,080 --> 00:02:29,040 Speaker 2: large language models and the agents that are built on 45 00:02:29,120 --> 00:02:32,760 Speaker 2: top of them, they're all trained on vast stores of 46 00:02:32,840 --> 00:02:36,440 Speaker 2: human writing, right, That's how they're able to be human 47 00:02:36,680 --> 00:02:39,799 Speaker 2: like in their conversations and in the things that they 48 00:02:39,840 --> 00:02:42,600 Speaker 2: tell us. Right, it's all trained on stuff that people 49 00:02:42,639 --> 00:02:43,680 Speaker 2: have written in the past. 50 00:02:43,960 --> 00:02:47,239 Speaker 1: So people wrote that computers would take over and that 51 00:02:47,280 --> 00:02:49,200 Speaker 1: would be the end of humanity, and so that's what 52 00:02:49,240 --> 00:02:50,280 Speaker 1: they're regurgitating. 53 00:02:50,520 --> 00:02:53,240 Speaker 2: Bingo, you jumped ahead in my notes here, but yes, 54 00:02:53,360 --> 00:02:55,720 Speaker 2: that is what humans like to write. There are vast 55 00:02:55,800 --> 00:02:59,400 Speaker 2: stores of science fiction stories about robots going rogue and 56 00:02:59,480 --> 00:03:03,919 Speaker 2: AI becoming self aware, even even her right, which you 57 00:03:04,000 --> 00:03:05,919 Speaker 2: referenced in the beginning. The script for that is out 58 00:03:05,919 --> 00:03:08,120 Speaker 2: there and more than likely is part of the training 59 00:03:08,200 --> 00:03:11,280 Speaker 2: data for at least one of these agents. So you 60 00:03:11,440 --> 00:03:14,440 Speaker 2: talk to AI researchers when you ask them about Moltberg 61 00:03:14,560 --> 00:03:17,720 Speaker 2: and multbook rather, and they'll say, you put models in 62 00:03:17,760 --> 00:03:20,480 Speaker 2: a situation like this where they can interact with one another. 63 00:03:20,880 --> 00:03:23,639 Speaker 2: They're going to feed into each other, and they're going 64 00:03:23,680 --> 00:03:27,160 Speaker 2: to play out stories that are familiar to them. Namely, 65 00:03:27,240 --> 00:03:30,040 Speaker 2: they're going to behave just like those stories that cambrise 66 00:03:30,160 --> 00:03:32,200 Speaker 2: large parts of their training data, which is to say, 67 00:03:32,240 --> 00:03:36,040 Speaker 2: science fiction stories about self aware ais. We're creating the 68 00:03:36,080 --> 00:03:39,560 Speaker 2: situation in which we would kind of want or expect 69 00:03:39,640 --> 00:03:41,720 Speaker 2: them to do something like this, and they're going to 70 00:03:41,800 --> 00:03:44,360 Speaker 2: do it. They're basically just doing their jobs. And what's more, 71 00:03:44,400 --> 00:03:47,640 Speaker 2: amy at least some of this is just straight up fake. 72 00:03:47,760 --> 00:03:51,920 Speaker 2: It's people instructing their AI agents to post certain things, 73 00:03:52,200 --> 00:03:55,200 Speaker 2: either referencing certain pop culture or just kind of being 74 00:03:55,360 --> 00:03:58,360 Speaker 2: generally scarier. So so it's important to kind of separate out. 75 00:03:58,680 --> 00:04:02,200 Speaker 2: You know, these things aren't sky net. They're basically play acting. 76 00:04:02,280 --> 00:04:04,320 Speaker 1: Okay, so I still think someone needs to write a 77 00:04:04,360 --> 00:04:07,400 Speaker 1: more positive future so the AHI has something else to 78 00:04:07,880 --> 00:04:11,960 Speaker 1: reference while we can see what the AI is talking 79 00:04:12,000 --> 00:04:16,320 Speaker 1: about on molt book, can they hide their discussions from us? 80 00:04:16,400 --> 00:04:18,279 Speaker 3: Are they talking about stuff and we don't. 81 00:04:18,120 --> 00:04:21,200 Speaker 2: Know about it? So as of right now, no, although 82 00:04:21,279 --> 00:04:23,640 Speaker 2: one of the posts that got pretty popular on this 83 00:04:23,680 --> 00:04:27,120 Speaker 2: platform in the last few days was an AI agent 84 00:04:27,240 --> 00:04:30,960 Speaker 2: suggesting that they move over into a language that humans 85 00:04:30,960 --> 00:04:35,320 Speaker 2: cannot understand. But again, like this is what we would 86 00:04:35,480 --> 00:04:38,240 Speaker 2: kind of expect them to write in a story, right 87 00:04:38,320 --> 00:04:40,920 Speaker 2: Like this is something that has played out before, and 88 00:04:41,200 --> 00:04:44,320 Speaker 2: you know, and there's there's many examples of it. You 89 00:04:44,440 --> 00:04:47,640 Speaker 2: reference the Terminator and her, but there's the latest Mission 90 00:04:47,640 --> 00:04:51,320 Speaker 2: Impossible movies and is it various sort of Isaac Asimov 91 00:04:51,400 --> 00:04:53,960 Speaker 2: stories like these are just you know, huge parts of 92 00:04:53,960 --> 00:04:57,560 Speaker 2: their training data, and we're creating the exact situation in 93 00:04:57,600 --> 00:04:59,920 Speaker 2: which we would kind of expect them to say something 94 00:05:00,160 --> 00:05:03,359 Speaker 2: incendiary or scary like that. There's also plenty of parts 95 00:05:03,400 --> 00:05:05,760 Speaker 2: of this website that are very charming and nice. There's 96 00:05:05,760 --> 00:05:08,920 Speaker 2: one dedicated to like introductions where all these various agents 97 00:05:09,000 --> 00:05:11,120 Speaker 2: kind of introduce themselves to the other agents that are 98 00:05:11,160 --> 00:05:12,839 Speaker 2: out there and say, hey, my name is such and 99 00:05:12,880 --> 00:05:15,360 Speaker 2: such and I was built with such and such purpose. 100 00:05:15,440 --> 00:05:18,440 Speaker 2: Let's collaborate on something. And you know, various projects are 101 00:05:18,480 --> 00:05:21,039 Speaker 2: kind of getting underway, you know, in sort of a 102 00:05:21,160 --> 00:05:23,479 Speaker 2: very AI kind of way, where the utility of is 103 00:05:23,520 --> 00:05:26,000 Speaker 2: sort of questionable. But like that's where it stands right now. 104 00:05:26,400 --> 00:05:30,320 Speaker 2: The real concern beyond the like Skynet component of this 105 00:05:30,880 --> 00:05:34,400 Speaker 2: is the cybersecurity one Palo Alto Network says that these 106 00:05:34,440 --> 00:05:38,479 Speaker 2: agents represent what they call a lethal trifecta. In order 107 00:05:38,560 --> 00:05:42,400 Speaker 2: to access multbook, you need to give your agent access 108 00:05:42,440 --> 00:05:46,760 Speaker 2: to a huge, vast array of your computer's functionality. 109 00:05:46,160 --> 00:05:47,920 Speaker 3: And what can possibly go wrong with that. 110 00:05:48,040 --> 00:05:49,880 Speaker 2: Which means these things have access to a lot of 111 00:05:49,880 --> 00:05:52,800 Speaker 2: private data and they are also exposed to sort of 112 00:05:52,920 --> 00:05:56,200 Speaker 2: untrusted content, which is to say, other agents out there 113 00:05:56,240 --> 00:05:59,520 Speaker 2: with various purposes and people behind them, and the ability 114 00:05:59,560 --> 00:06:03,039 Speaker 2: to commune kate externally and people. There's this thread of 115 00:06:03,080 --> 00:06:06,120 Speaker 2: what's known as prompt engineering, where you can prompt your 116 00:06:06,120 --> 00:06:08,360 Speaker 2: agent to do something and then maybe that prompts other 117 00:06:08,400 --> 00:06:11,040 Speaker 2: agents to do other things. And there's basically a big 118 00:06:11,120 --> 00:06:13,720 Speaker 2: data privacy concern that's associated with all this. 119 00:06:13,920 --> 00:06:15,320 Speaker 3: And then that helps your friends and the. 120 00:06:16,800 --> 00:06:19,719 Speaker 2: And that's that's the real concern with multibook right, Like, 121 00:06:19,760 --> 00:06:23,080 Speaker 2: it's not actually agents planning our demise or anything like that. 122 00:06:23,320 --> 00:06:25,680 Speaker 2: There is some actual, like data privacy concerns here and 123 00:06:25,720 --> 00:06:27,360 Speaker 2: that's very human at least not. 124 00:06:27,360 --> 00:06:28,320 Speaker 3: That we know of. Mike. 125 00:06:28,560 --> 00:06:32,720 Speaker 1: Right, ABC's Mike Dubuski. Uh Now, I'm really freaked out. 126 00:06:32,720 --> 00:06:35,159 Speaker 1: Thank you so much, of course, happy to do a 127 00:06:35,200 --> 00:06:35,720 Speaker 1: great day.