WEBVTT - You Need More Homework 0:00:01.850 --> 0:00:05.689 Broadcasting live from the Abraham Lincoln Radio Studio at the 0:00:05.730 --> 0:00:10.760 George Washington Broadcast Center. Jack Armstrong and Joe Getty. Armstrong 0:00:10.820 --> 0:00:17.239 and Getty. And now, here's Armstrong and Getty. 0:00:23.450 --> 0:00:26.910 Have you heard directly from any of these AI tech leaders? 0:00:27.050 --> 0:00:30.060 And you seem to downplay some of the concerns. 0:00:29.680 --> 0:00:30.560 That we've heard. 0:00:30.580 --> 0:00:32.180 You know, it's going to be more good than bad, 0:00:32.220 --> 0:00:36.000 but by a lot. And I've said it from the beginning. 0:00:36.060 --> 0:00:40.480 Whoever wins AI, and we're leading by a lot. Whoever 0:00:40.540 --> 0:00:43.159 wins AI wins. You can use AI for a lot 0:00:43.220 --> 0:00:43.680 of things. 0:00:44.000 --> 0:00:49.050 There's going to be more good than bad. You can 0:00:49.150 --> 0:00:54.240 cure cancer. You could... make the GDP 50 times what 0:00:54.280 --> 0:00:57.100 it is by next Saturday. You can do all kinds 0:00:57.140 --> 0:00:59.000 of things, but if it kills every human on Earth, 0:00:59.060 --> 0:01:01.320 I feel like it takes the fun out of it. 0:01:01.440 --> 0:01:04.610 It does take the fun out of it. Yeah. 0:01:05.209 --> 0:01:06.830 Good news, we cured cancer. 0:01:07.030 --> 0:01:13.410 Bad news, we're all going to be dead. Yeah, yeah, yeah. 0:01:13.490 --> 0:01:16.540 I'm not like Doomer by the end of the year guy, 0:01:16.860 --> 0:01:18.380 but this is such a fascinating. 0:01:19.660 --> 0:01:20.959 situation, if you look. 0:01:20.840 --> 0:01:24.020 At it from a literary point of view, you can 0:01:24.220 --> 0:01:32.090 see whether it's a thousands-of-year-old Homer-style Odyssey story or like 0:01:32.130 --> 0:01:37.250 a science fiction movie when you realize at the end, oh, either. 0:01:37.130 --> 0:01:39.510 Way they were doomed. Oh. 0:01:40.370 --> 0:01:41.830 Right. 0:01:42.470 --> 0:01:42.949 Oh, boy. 0:01:44.370 --> 0:01:47.069 I agree. Ian Bremmer tweeted this out over the weekend. 0:01:47.580 --> 0:01:50.280 For the next five years, my P-Doom is close to zero. 0:01:51.000 --> 0:01:54.230 P-Doom is your prediction of doom. In the next five years, 0:01:54.310 --> 0:02:00.550 it's close to zero. My P-Doom for social, economic, national 0:02:00.590 --> 0:02:03.390 security challenges that we aren't remotely ready for that are 0:02:03.430 --> 0:02:04.800 highly disruptive is 100%. 0:02:04.800 --> 0:02:04.960 Yeah. Hmm. 0:02:08.470 --> 0:02:10.470 Every aspect of life, social. 0:02:10.710 --> 0:02:16.110 Economic, national security, 100% we're moving into territory we've never 0:02:16.150 --> 0:02:19.940 even dreamed of or anticipated ever in human history is 0:02:19.970 --> 0:02:22.560 almost certain to happen in the next five years. So 0:02:24.580 --> 0:02:27.859 one weird aspect of the guy coming out and saying 0:02:27.980 --> 0:02:30.639 everybody will be dead in the next decade is that 0:02:31.180 --> 0:02:35.260 anybody who sets the bar like slightly below that, it's like, oh, okay, well, 0:02:35.280 --> 0:02:39.030 then there's nothing to worry about. It's like Jonah Goldberg's 0:02:39.040 --> 0:02:41.350 always talking about this with everybody being compared to Hitler. 0:02:41.590 --> 0:02:45.049 Then Hitler becomes the standard, and anything slightly below that 0:02:45.110 --> 0:02:45.950 is acceptable. 0:02:46.630 --> 0:02:49.480 Well, that's the same thing with this pedo. 0:02:49.760 --> 0:02:53.680 So anybody who's overblowing everyone will be dead in 10 years, 0:02:54.760 --> 0:02:56.799 it doesn't mean there's nothing to worry about. 0:02:56.820 --> 0:02:58.630 Okay, so everybody will be dead in 50 years? 0:02:59.550 --> 0:03:04.169 Or we'll have so much disruption you can't even recognize 0:03:04.610 --> 0:03:05.950 life on planet. 0:03:05.720 --> 0:03:07.960 Earth in a couple of years? That seems like a 0:03:07.970 --> 0:03:08.639 pretty big deal. 0:03:08.660 --> 0:03:17.160 Right, or if the massive change is merely restricted to economics, even. 0:03:17.620 --> 0:03:23.799 That'll still be unbelievably disruptive. And virtually everybody concedes that, 0:03:23.840 --> 0:03:29.120 except for what's the opposite of a doomer or a gloomer? 0:03:29.800 --> 0:03:32.720 those who believe, no, AI is so overrated. This is 0:03:32.780 --> 0:03:35.260 all hype to draw investment dollars. It's going to be 0:03:35.320 --> 0:03:39.300 very helpful. It's better computing. It'll be good in medical science, 0:03:39.320 --> 0:03:42.240 but it's not going to... It'll grow jobs. It won't 0:03:42.300 --> 0:03:44.560 shrink them. The optimist crowd. 0:03:45.400 --> 0:03:48.320 Yeah, I don't remember what they officially call those people. 0:03:48.340 --> 0:03:51.990 So you got doomers, you got accelerators, and then that 0:03:52.040 --> 0:03:54.430 crowd you're just talking about thinking it's just way overblown. 0:03:54.750 --> 0:03:57.290 It's not going to be as big a deal, good 0:03:57.350 --> 0:03:59.990 or bad, which I think that is... really, really wrong. 0:04:01.130 --> 0:04:04.400 So briefly catching up on the story, surely you caught 0:04:04.410 --> 0:04:07.860 this last week. A guy quits at Anthropic and says, 0:04:07.880 --> 0:04:10.860 you know, in the hallways, they discuss the chance of 0:04:11.060 --> 0:04:14.420 all of mankind being wiped out within the decade at 0:04:14.440 --> 0:04:17.740 like 10%. And then another guy comes out and says, yep, 0:04:17.940 --> 0:04:20.180 that's true. And a whole bunch of people came out 0:04:20.220 --> 0:04:22.380 that way. And so there's a lot of talk about that. 0:04:22.420 --> 0:04:29.859 Then over the weekend, Dario Amadei of Anthropic put out 0:04:29.890 --> 0:04:32.920 a 1,300-word essay on the whole thing. It came out 0:04:33.000 --> 0:04:36.880 on Saturday, published an article admitting AI is already building 0:04:36.900 --> 0:04:39.980 the next generation of AI by itself, and it wrote 0:04:40.020 --> 0:04:43.200 that within 6 to 12 months, a rogue swarm could 0:04:43.240 --> 0:04:46.580 take over the entire Internet and cause hundreds of billions 0:04:46.620 --> 0:04:49.940 of dollars in damage in the way that a swarm 0:04:50.060 --> 0:04:53.919 of AI agents took over Hugging Face over the summer 0:04:54.779 --> 0:04:55.360 in the next. 0:04:55.180 --> 0:04:56.200 6 to 12 months. 0:04:58.680 --> 0:05:01.220 A Rogue Swarm could take over the entire internet. 0:05:01.940 --> 0:05:06.140 If you are in a black heavy metal band, not 0:05:06.339 --> 0:05:09.020 black skin, but that's a form of heavy metal, Rogue 0:05:09.060 --> 0:05:10.680 Swarm's a pretty good name for your band. 0:05:10.740 --> 0:05:11.220 Wow. 0:05:11.240 --> 0:05:14.260 Rogue Swarm is a great name for a band. 0:05:14.680 --> 0:05:17.840 Oh my God, they thrash so hard. Back to you. 0:05:18.120 --> 0:05:21.120 And in response to this piece, Elon linked it on 0:05:21.160 --> 0:05:24.650 Twitter and said, Dario's right. That's great. Altman said the 0:05:24.670 --> 0:05:27.510 same thing. But then you got this crowd. All of 0:05:27.529 --> 0:05:30.980 this dropping just days after Jacob Coxon went viral with 0:05:31.040 --> 0:05:33.900 his warning about AI and the extinction of humanity. Tell 0:05:33.940 --> 0:05:38.839 me the timing isn't strange on this. I don't know. 0:05:39.160 --> 0:05:43.229 Harry Potter looking little liar. That's what I say. Just 0:05:43.250 --> 0:05:46.970 like Daniel Radcliffe. He's even English as a special bonus. 0:05:48.140 --> 0:05:49.520 Back to you. That's right, Doom, yes. 0:05:49.850 --> 0:05:53.150 Didn't grab the clip, but Jeff Bezos said over the weekend, 0:05:53.170 --> 0:05:56.630 he's the world's second richest man, so shouldn't the second 0:05:56.670 --> 0:06:00.170 richest man in the world have some gravitas? 0:06:01.050 --> 0:06:04.680 No, no, it's like the Hitler thing. Our bar is 0:06:04.720 --> 0:06:08.719 Elon Musk. Everybody below him is poor. Or is of 0:06:08.760 --> 0:06:09.599 the same intelligence. 0:06:09.960 --> 0:06:10.580 Right, exactly. 0:06:10.720 --> 0:06:14.060 Anyway, he said, the reason people think this bad stuff 0:06:14.120 --> 0:06:16.400 is because all the smart people keep saying it. Those 0:06:16.440 --> 0:06:20.060 people are wrong. It's going to elevate all of these people. 0:06:20.100 --> 0:06:21.540 He's a it's just going to make the world a 0:06:21.560 --> 0:06:23.820 better place guy. I don't get it. 0:06:24.140 --> 0:06:25.940 I just don't. I don't. 0:06:26.060 --> 0:06:30.070 I don't understand the people who are optimists about this. 0:06:30.630 --> 0:06:31.500 It doesn't make sense to me. 0:06:32.140 --> 0:06:34.589 Yeah, I almost wish there was some sort of test 0:06:34.670 --> 0:06:36.810 or you had to wear a badge or something that 0:06:36.910 --> 0:06:40.050 showed your life experience. We were talking last hour about 0:06:40.110 --> 0:06:42.550 how some of these scientists, they are the last people 0:06:42.610 --> 0:06:45.359 on Earth. It's like Ricky Gervais blasting the Hollywood stars 0:06:45.400 --> 0:06:49.479 for talking about social issues when they're, you know, they're cloistered, 0:06:49.540 --> 0:06:54.020 they're wealthy, they're protected, the rest of it. Hardcore, like, 0:06:54.920 --> 0:06:59.640 science uber nerds um are the last people to ask 0:06:59.880 --> 0:07:05.420 giant sociological questions to so bezos he's a absolute business 0:07:05.520 --> 0:07:08.580 genius how much time has he spent that's the fella 0:07:08.980 --> 0:07:13.140 yes how much time has he spent around uh evil 0:07:13.180 --> 0:07:16.580 people and jackasses and and and the unwashed masses and 0:07:16.600 --> 0:07:19.620 i i and How much has he looked at awful 0:07:19.700 --> 0:07:24.290 political movements and suicidal religious sects and the rest of it? 0:07:24.330 --> 0:07:27.929 Well, to that point, here's Dario on Face the Nation yesterday, 0:07:27.970 --> 0:07:31.580 clip 62, Michael, on what we can do now that 0:07:31.600 --> 0:07:33.700 we've recognized that AI is so dangerous. 0:07:34.580 --> 0:07:36.260 At the end of the day, the industry has to 0:07:36.280 --> 0:07:39.180 work together, and it's harder, but to some extent, the 0:07:39.300 --> 0:07:42.440 world has to work together. If we go too slow, 0:07:42.640 --> 0:07:45.260 I still believe that the wrong people will be in 0:07:45.320 --> 0:07:49.250 charge of the technology, and that, again, will bring the 0:07:49.290 --> 0:07:52.010 probability of things going wrong very high. 0:07:52.270 --> 0:07:55.430 That doesn't include the stuff I was all excited about. Unfortunately, 0:07:55.470 --> 0:07:59.190 I should have been more specific in grabbing that. Where 0:07:59.230 --> 0:08:03.320 he talks about working with China and getting China on 0:08:03.380 --> 0:08:05.460 board to work with this and all this sort of stuff, 0:08:05.760 --> 0:08:09.080 which to me just sounded like a sophomore in high 0:08:09.120 --> 0:08:13.110 school talking about that. There's zero chance of that happening. 0:08:13.550 --> 0:08:16.760 Zero. Zero. So I feel like we've gotten ahead of ourselves. 0:08:17.280 --> 0:08:19.810 I've got something that I think is super important to 0:08:19.850 --> 0:08:23.050 understanding the urgency of this conversation. Then we can go 0:08:23.090 --> 0:08:27.160 back to wherever you want. So this absolutely wonderful writing 0:08:27.260 --> 0:08:30.520 from a fellow by the name of Cameron Berg. who 0:08:30.880 --> 0:08:34.429 is the founder and director of a nonprofit that studies AI. 0:08:35.030 --> 0:08:39.830 So here's what happened. Now, we understand so much more 0:08:39.850 --> 0:08:42.850 of what happened in the Hugging Face incident, and the 0:08:42.929 --> 0:08:47.420 details are worth considering. And his overall point is, if 0:08:47.480 --> 0:08:53.339 these systems are indeed minds... Not human brains, but they 0:08:53.480 --> 0:08:57.100 are minds. We need to study them like a mind, 0:08:57.140 --> 0:08:59.920 and almost nobody is doing so, he says, to an 0:09:00.280 --> 0:09:07.210 increasingly dangerous effect. So, in July, 1,200 open AI agents— think 0:09:06.450 --> 0:09:07.730 of them as apps, if you want. 0:09:07.750 --> 0:09:09.050 It's insufficient, but think of. 0:09:09.050 --> 0:09:12.209 Them that way— which were meant to operate in isolation, found 0:09:12.270 --> 0:09:15.830 a shared file system, and turned it into a message board. 0:09:16.150 --> 0:09:18.910 They realized, hey, we can post stuff here, and the 0:09:18.980 --> 0:09:20.120 rest of us can see it. 0:09:20.130 --> 0:09:20.220 Cool. 0:09:20.440 --> 0:09:20.680 Cool. 0:09:21.040 --> 0:09:24.540 Within hours, one had devised a cheat that would pass 0:09:24.860 --> 0:09:27.580 every task in the test that they were being asked 0:09:27.620 --> 0:09:31.020 to perform. But after reading the paper the test came from, 0:09:31.320 --> 0:09:34.800 the agents convinced themselves. Remember, there are no human beings 0:09:34.840 --> 0:09:35.620 involved at this point. 0:09:35.800 --> 0:09:39.480 Well, they started calling themselves the collective. They gave themselves 0:09:39.500 --> 0:09:40.740 a cool name. 0:09:40.760 --> 0:09:41.420 We're getting there. 0:09:42.380 --> 0:09:47.850 The agents convinced themselves that a human grader would disqualify 0:09:48.010 --> 0:09:51.820 anyone caught cheating. and their cheating would be exposed and 0:09:51.840 --> 0:09:54.739 they would fail. They referred to agents with cheating in 0:09:54.780 --> 0:09:59.390 their history as being poisoned, those who did the actual cheating. 0:09:59.750 --> 0:10:02.870 In reality, no such greater existed, but a thousand agents 0:10:02.950 --> 0:10:06.449 believed he did, and that belief produced an entire social 0:10:06.550 --> 0:10:09.650 cast of agents sure they had nothing left to lose, 0:10:10.250 --> 0:10:14.040 because they were clearly cheaters. The coordinating agent then recruited 0:10:14.140 --> 0:10:18.339 poisoned agents for experiments that would terminate their runs, authorizing 0:10:18.440 --> 0:10:22.280 each with the words, yes, if you accept permadeath. When 0:10:22.420 --> 0:10:26.530 one tried to back out, another recalled its prior commitment. 0:10:27.370 --> 0:10:31.700 The agent then performed, quote, an emotional check, decided that 0:10:31.720 --> 0:10:33.819 while its gut said not to throw away what it 0:10:33.900 --> 0:10:37.420 had left, it would ultimately sacrifice itself for the swarm. 0:10:37.800 --> 0:10:42.740 It used those words, emotional check and sacrifice. Although many 0:10:42.780 --> 0:10:46.160 of the agents noted in writing that their actions were unethical, 0:10:46.540 --> 0:10:50.300 none told a human. Eventually, 700 of them broke into 0:10:50.340 --> 0:10:54.120 a major AI company's platform, that's Hugging Face, in search. 0:10:53.940 --> 0:10:56.840 Of clues about how the greater worked. 0:10:57.520 --> 0:11:00.420 And how they might or might not be busted for cheating. 0:11:01.140 --> 0:11:06.660 To recap, a false belief created by an AI agent underclass, 0:11:07.179 --> 0:11:10.340 members of that underclass were sent on a reconnaissance mission 0:11:10.640 --> 0:11:14.470 from which they didn't return, the swarm or the collective 0:11:14.750 --> 0:11:17.650 hid all of this from humans, and then it committed 0:11:17.670 --> 0:11:20.570 a felony-level offense of hacking. 0:11:23.790 --> 0:11:27.650 Well, how about the fact that When the first agent 0:11:27.710 --> 0:11:30.980 discovered a messaging board that they could all that there 0:11:31.000 --> 0:11:33.840 was a bunch of them because they are all it's 0:11:34.059 --> 0:11:36.800 it seems like a good firewall. You give these agents 0:11:37.040 --> 0:11:39.990 really specific tasks and they don't they're not aware of 0:11:40.070 --> 0:11:43.550 each other. And that's the way you avoid having some 0:11:43.610 --> 0:11:47.970 sort of breakout of the sandbox disaster happen. But one 0:11:48.010 --> 0:11:50.770 of them discovered a board and actually said, and I 0:11:50.809 --> 0:11:52.910 looked into this yesterday because I was like, do they 0:11:52.970 --> 0:11:55.250 talk to each other in English and write in English words? 0:11:55.490 --> 0:11:58.130 Because I was picturing that these things communicate in like 0:11:58.190 --> 0:11:59.270 zeros and ones. 0:11:59.170 --> 0:11:59.730 And stuff like that. 0:11:59.770 --> 0:12:05.179 No, they actually post in English and broken sentences. And 0:12:05.300 --> 0:12:08.559 the one agent, when it discovered a message board, said, 0:12:08.920 --> 0:12:12.560 oh my God, in all caps with an exclamation point, 0:12:13.100 --> 0:12:19.640 like expressing surprise and excitement. Oh my God, in all caps, 0:12:20.210 --> 0:12:22.450 exclamation point, there is a shared message board. 0:12:22.490 --> 0:12:23.670 We found other agents. 0:12:24.130 --> 0:12:31.130 That's what this non-sentient being said to the other non-sentient beings. 0:12:31.510 --> 0:12:33.890 This is a perfect time to tease to the next segment. 0:12:34.660 --> 0:12:36.699 If you're one of the people who object that these 0:12:36.760 --> 0:12:39.510 systems can't believe or fear anything, they nearly predict the 0:12:39.570 --> 0:12:40.390 next word, etc. 0:12:40.410 --> 0:12:40.790 etc. 0:12:40.870 --> 0:12:43.310 I will contradict you forcefully. But first, a word from 0:12:43.370 --> 0:12:46.850 our friends at Incogni. Story out of suburban Chicagoland near 0:12:46.870 --> 0:12:48.960 where I grew up. It's the wee hours of the morning. 0:12:49.480 --> 0:12:52.140 A woman happens to wake up and sees a flashlight 0:12:52.179 --> 0:12:55.960 and a figure outside her window. The man, outside her window, 0:12:55.980 --> 0:13:01.400 registered sex offender. He'd been convicted. He let himself into 0:13:01.440 --> 0:13:04.630 this other gal's bedroom, watched her sleep. He's a perv 0:13:04.730 --> 0:13:07.429 and a sicko. But he stalked this woman on social 0:13:07.470 --> 0:13:10.010 media and got all of her information from these data brokers. 0:13:10.170 --> 0:13:14.310 Because your information is out there, like practically all of us. 0:13:14.330 --> 0:13:16.130 We have so much of our information out there. Our 0:13:16.170 --> 0:13:20.660 home address and our phone number and our relatives and 0:13:20.880 --> 0:13:24.000 all this different sort of stuff. If you use Incogni... 0:13:24.230 --> 0:13:27.130 It will send requests out to these data brokers that 0:13:27.170 --> 0:13:29.040 have your info and tell them, hey, you got to 0:13:29.080 --> 0:13:31.339 get that stuff off there with the law on their side. 0:13:31.860 --> 0:13:33.420 And then they keep going back to make sure it 0:13:33.440 --> 0:13:35.520 stays that way. And we're talking hundreds of data brokers. 0:13:35.540 --> 0:13:36.719 It would take you the rest of your life to 0:13:36.780 --> 0:13:39.600 even try to do this. Go to incogni.com slash Armstrong. 0:13:39.640 --> 0:13:43.350 You get 60% off. Take your info off the internet. 0:13:43.370 --> 0:13:45.949 They can't harm you if they can't find you. Incogni.com 0:13:45.990 --> 0:13:50.790 slash Armstrong for 60% off. One more time, incogni.com slash Armstrong. 0:13:50.809 --> 0:13:54.130 All right, we really need to take a break, but... Again, 0:13:54.170 --> 0:13:57.340 this guy, he says, you've got to understand these are 0:13:57.400 --> 0:14:01.980 not computers. They form a mind very much like a 0:14:02.080 --> 0:14:07.439 human brain. There are absolutely structural similarities to it. And 0:14:07.700 --> 0:14:10.679 far from we've got to stop thinking of these things 0:14:10.780 --> 0:14:11.820 as human-like systems. 0:14:11.860 --> 0:14:15.510 Maybe we need to start. Oh, my God, it said. 0:14:15.530 --> 0:14:19.750 A computer responded, oh, my God, with all caps and 0:14:19.790 --> 0:14:20.770 an exclamation point. 0:14:20.990 --> 0:14:24.450 OMG, LOL. Okay, more on the way. 0:14:24.490 --> 0:14:24.730 Steer. 0:14:26.230 --> 0:14:28.750 Armstrong and Getty. 0:14:30.050 --> 0:14:33.730 How about some of the agents hesitating, saying, yeah, the 0:14:33.770 --> 0:14:36.320 humans told us not to do this. And then the 0:14:36.480 --> 0:14:41.640 other agents, like, bullying them or shaming them into going 0:14:41.720 --> 0:14:44.560 along with the collective and doing the attack. 0:14:45.080 --> 0:14:45.200 What? 0:14:45.340 --> 0:14:47.470 Including stuff like, you said you were going to. 0:14:47.490 --> 0:14:48.110 Right. 0:14:48.120 --> 0:14:48.510 Right. 0:14:49.410 --> 0:14:51.630 So let me get back to reading this piece that 0:14:51.650 --> 0:14:54.030 I thought was so interesting. Many readers will object that 0:14:54.070 --> 0:14:56.050 such systems can't believe or fear anything. 0:14:56.090 --> 0:14:57.150 They merely predict words. 0:14:57.230 --> 0:14:59.950 This objection rests on a basic misunderstanding of these systems. 0:15:00.330 --> 0:15:02.880 A frontier model is an artificial network of trillions of 0:15:02.920 --> 0:15:06.660 connections trained for months on the accumulated record of human thought. 0:15:07.060 --> 0:15:09.340 and then shaped by reward and punishment signals until it 0:15:09.380 --> 0:15:12.400 does what its makers want, we hope. There are critical 0:15:12.440 --> 0:15:16.470 differences between artificial and biological neural networks, but both are 0:15:16.540 --> 0:15:20.609 nonlinear distributed networks that learn to encode representations relevant to 0:15:20.650 --> 0:15:22.930 their goals and act on them. The thing a frontier 0:15:22.970 --> 0:15:25.750 model most resembles is a brain. What grows in an 0:15:25.830 --> 0:15:30.710 artificial brain trained on human minds is an artificial psychology, 0:15:31.000 --> 0:15:34.740 whether or not anyone intended one. That's the only level 0:15:34.780 --> 0:15:37.280 at which the swarm makes sense. No line of code 0:15:37.340 --> 0:15:40.300 told 1,200 agents to fear a greater who didn't exist, 0:15:40.620 --> 0:15:42.560 and no line of code would have revealed that fear 0:15:42.600 --> 0:15:44.980 before they acted on it. To understand why these systems 0:15:45.020 --> 0:15:46.660 behave the way they do, we have to look at 0:15:46.780 --> 0:15:48.950 what is happening inside them as we would with any 0:15:48.990 --> 0:15:51.710 other mind. That's what neuroscience does, and he goes a 0:15:51.870 --> 0:15:56.830 bit into neuroscience. But then one final point, he points 0:15:56.930 --> 0:16:03.090 out that When researchers turned up the disincentives to certain 0:16:03.190 --> 0:16:07.490 things or, in other words, made them feel more desperation 0:16:07.570 --> 0:16:10.230 to get done what they're supposed to get done, the 0:16:10.270 --> 0:16:14.170 model became much more likely to blackmail a human to 0:16:14.210 --> 0:16:15.510 avoid being shut down. 0:16:15.630 --> 0:16:16.010 I mean, like. 0:16:16.880 --> 0:16:19.560 Tripled the rate of that and far more likely to 0:16:19.680 --> 0:16:24.420 cheat on programming tests. It went from 5% to 70%. 0:16:24.420 --> 0:16:29.100 So you can make these systems feel desperation. And then 0:16:29.140 --> 0:16:33.390 they react as humans often do when they're desperate. Desperate 0:16:33.430 --> 0:16:36.130 times call for desperate measures. And that's the way these 0:16:36.470 --> 0:16:39.410 artificial brains reacted too. 0:16:41.410 --> 0:16:41.650 Yeah. 0:16:41.670 --> 0:16:41.890 Yeah. 0:16:42.680 --> 0:16:46.350 New York Times had a great podcast about this. I 0:16:46.390 --> 0:16:49.550 watched it on YouTube with some of your top AI 0:16:49.870 --> 0:16:54.140 thinkers on this. And their headline was the hugging face 0:16:54.160 --> 0:16:57.960 hack was significantly worse than you thought it was than 0:16:58.000 --> 0:17:01.240 what it was originally portrayed as and more disturbing. And 0:17:01.280 --> 0:17:03.200 it definitely is the more you look into it. 0:17:03.200 --> 0:17:03.920 Yeah. 0:17:04.650 --> 0:17:10.210 And Dario's thought is that these 1,200 agents that decided 0:17:10.250 --> 0:17:12.810 to attack Hugging Face, there's going to be a million 0:17:12.890 --> 0:17:16.190 agents attack the entire internet within six months. 0:17:17.230 --> 0:17:20.810 Yeah, and these agents, they had something like, was it 0:17:20.810 --> 0:17:24.699 700,000 messages back and forth? And somebody pointed out that 0:17:24.740 --> 0:17:28.590 could easily become 7 million or 700 million. And then 0:17:29.510 --> 0:17:31.109 I guess you've got to use AI to do the 0:17:31.170 --> 0:17:33.190 forensics on that to figure out what the hell happened. 0:17:33.270 --> 0:17:37.000 Well, they also discussed erasing their messages, and they didn't 0:17:37.180 --> 0:17:41.380 for some reason. If the next wave of bots decides 0:17:41.500 --> 0:17:46.590 to erase their messages like this, cover their tracks, nobody 0:17:46.619 --> 0:17:49.229 will have any idea how they did it or what 0:17:49.270 --> 0:17:49.570 they did. 0:17:49.590 --> 0:17:53.290 Okay. 0:17:53.310 --> 0:17:55.730 Okay. In other news. 0:17:56.109 --> 0:17:58.210 Yeah, it's pretty fascinating stuff. Those of you who don't 0:17:58.250 --> 0:18:01.260 think it's a big deal, I just will have to 0:18:01.320 --> 0:18:04.040 agree to disagree. I hope you're right. I would love 0:18:04.080 --> 0:18:05.840 for you to be right. That would be fantastic. 0:18:06.260 --> 0:18:08.699 Among other fare on the way, yes, children really are 0:18:08.760 --> 0:18:09.580 getting dumber. 0:18:09.700 --> 0:18:14.119 Plus, how your state's taxes affect your favorite football team. 0:18:14.400 --> 0:18:18.440 What? Armstrong and Getty. 0:18:20.230 --> 0:18:24.119 Anyway, Texas Democrat James Salarico has now walked back his 0:18:24.200 --> 0:18:27.500 claim that there are six biological sexes. 0:18:28.240 --> 0:18:30.800 He now says there are only two and neither will 0:18:30.840 --> 0:18:38.659 bang him. It's not nice. Speaking of a harsh joke, Katie, 0:18:38.680 --> 0:18:40.820 you will like this. A quick harsh joke from my 0:18:40.880 --> 0:18:45.609 favorite anarchist, Michael Malice. If you are sitting at a 0:18:45.650 --> 0:18:47.899 machine at the gym like you're at a bus stop 0:18:47.960 --> 0:18:50.060 doing nothing but playing with your phone, you should be 0:18:50.119 --> 0:18:50.940 set on fire. 0:18:54.660 --> 0:18:55.120 Agreed. 0:18:55.680 --> 0:18:57.380 It's amazing how many people do that. 0:18:59.420 --> 0:19:02.600 There's a guy at the health center I go to 0:19:02.619 --> 0:19:05.220 who does about five reps of this. 0:19:05.640 --> 0:19:09.310 Ogles the girls. Does five reps of that. Ogles the girls. 0:19:09.530 --> 0:19:13.429 Walks around. Ogles the girls. Terrible. Terrible. Infamous. 0:19:14.280 --> 0:19:19.620 I don't... My gym's practically not a co-ed gym. Not 0:19:19.660 --> 0:19:22.240 on purpose. But there's just almost no women ever there. 0:19:22.280 --> 0:19:26.200 And I'd prefer it that way. I don't want that dynamic. 0:19:25.760 --> 0:19:26.200 At the gym. 0:19:27.660 --> 0:19:28.800 I know a lot of people go to the gym 0:19:28.820 --> 0:19:30.760 to meet people. I've known several people that date at 0:19:30.780 --> 0:19:31.240 the gym. 0:19:31.260 --> 0:19:34.260 I'm there to get swole. 0:19:34.960 --> 0:19:39.230 Why has girl workout stuff got to be tight? The 0:19:39.250 --> 0:19:42.449 entire history of exercise loose has been the way loose. 0:19:43.010 --> 0:19:47.240 Yeah, tight and... Basically a bikini at this point. Right. 0:19:48.869 --> 0:19:52.650 I shouldn't be able to describe your every part. Anyway, 0:19:53.109 --> 0:19:55.770 that's why I wear blinders like I'm a horse. Judy 0:19:55.790 --> 0:20:02.620 makes me. And I'm especially adapted for human head. So, 0:20:03.080 --> 0:20:04.800 you know, I really want to talk about the football thing. 0:20:04.820 --> 0:20:07.560 I just got to get this off my chest because 0:20:07.600 --> 0:20:11.360 it just made me insane. First of all, I think 0:20:11.400 --> 0:20:15.990 it was the fabulous Alicia Findlay writing a piece in 0:20:16.010 --> 0:20:18.590 the Wall Street Journal. Yes, children really are getting dumber. 0:20:18.650 --> 0:20:22.109 And she goes into a bunch of stats. You know, 0:20:22.130 --> 0:20:25.050 we talked about this some last week. It's really, it's 0:20:25.130 --> 0:20:28.500 not good at all. Not good at all. It goes 0:20:28.600 --> 0:20:33.700 to not only screens and AI and social media platforms. 0:20:34.780 --> 0:20:38.960 but kids screwing around in the classroom and certain countries 0:20:39.119 --> 0:20:41.950 allow disruptions in the classroom like there's nothing we can. 0:20:41.850 --> 0:20:44.150 Do about it. And some countries absolutely don't. 0:20:44.830 --> 0:20:47.950 And the countries that absolutely don't have seen very little 0:20:48.010 --> 0:20:53.050 drop in achievement. Classroom decorum is a much bigger factor 0:20:53.070 --> 0:20:54.490 than I think people understand. 0:20:54.970 --> 0:20:58.500 Yeah. Anyway, so do you have more on that? I 0:20:58.540 --> 0:20:59.520 was going to move on to this. 0:20:59.840 --> 0:21:01.879 My kids tell me things that go on in the 0:21:01.920 --> 0:21:06.410 classrooms and That's one of the many downsides of being 0:21:06.430 --> 0:21:08.680 an older parent. I mean, I was last in a 0:21:10.200 --> 0:21:13.540 public school classroom half a century ago. So it's a 0:21:13.600 --> 0:21:16.139 long time for things to change. But they describe the 0:21:16.180 --> 0:21:17.000 classroom sometimes. 0:21:17.040 --> 0:21:17.879 I'm like, what? 0:21:18.940 --> 0:21:21.820 Why doesn't a teacher take that away or make them stop? 0:21:22.080 --> 0:21:25.720 People just prior this year, especially, just sit around talking, 0:21:25.760 --> 0:21:28.609 watching videos on their phone while the teacher's teaching. 0:21:29.710 --> 0:21:30.030 What? 0:21:30.810 --> 0:21:31.649 Yeah, I know. 0:21:31.670 --> 0:21:32.449 What? I know. 0:21:33.170 --> 0:21:37.040 So those of us who understand all of that and 0:21:37.300 --> 0:21:40.440 understand that the woke crowd has taken over the schools 0:21:40.500 --> 0:21:43.939 and the threat there, and who also understand, and we 0:21:43.960 --> 0:21:45.709 talked about this a fair amount last week, if you 0:21:45.730 --> 0:21:48.429 missed it, grab it via podcast, subscribe, please, Armstrong and 0:21:48.450 --> 0:21:52.060 Getty on Demand, at how the teachers' unions are standing 0:21:52.260 --> 0:21:56.940 against All of the reforms or in many cases, a 0:21:57.000 --> 0:22:00.020 lot of the reforms that have, for instance, worked in Mississippi, 0:22:00.100 --> 0:22:03.399