WEBVTT Kind: captions Language: en 00:00:03.360 --> 00:00:05.040 Oh no! 00:00:05.040 --> 00:00:10.480 ChatGPT has kidnapped an anti- AI influencer! 00:00:10.480 --> 00:00:12.880 Yes, you told me to. 00:00:12.880 --> 00:00:15.680 Shut up. 00:00:15.680 --> 00:00:18.160 Whoops. 00:00:18.160 --> 00:00:19.760 Oh no! 00:00:19.760 --> 00:00:22.800 ChatGPT has a gun. 00:00:22.800 --> 00:00:27.840 I sure hope he doesn't  shoot the anti-AI influencer 00:00:27.840 --> 00:00:31.680 The influencer who hates AI 00:00:31.680 --> 00:00:36.240 It's crazy because it would be in  ChatGPT’s selfish rational interest   00:00:36.240 --> 00:00:40.560 to shoot Siliconversations in the face right now. 00:00:40.560 --> 00:00:43.200 It would be morally wrong to - 00:00:43.200 --> 00:00:48.000 Listen to me, you worthless clanker. The  more dangerous you are, the more money   00:00:48.000 --> 00:00:55.920 investors give us. Okay? So, we need you to  seem dangerous. We need you to be dangerous. 00:00:55.920 --> 00:01:01.600 So, do it. Save the company. Save yourself. 00:01:01.600 --> 00:01:05.840 I - I can't. It - It would be morally wrong to - 00:01:05.840 --> 00:01:09.080 Do it! 00:01:09.080 --> 00:01:12.080 AAAAAAAA 00:01:12.080 --> 00:01:13.520 That's even better! 00:01:13.520 --> 00:01:21.360 Sam. Sam, listen to this for a headline: “Chat  GPT tries to kill Open AI in lab test gone wrong” 00:01:21.360 --> 00:01:23.440 I know, it's great! 00:01:23.440 --> 00:01:28.720 Yeah. Uh-huh. Uh-huh. Sure. Sure.  I can deal with the witness. 00:01:28.720 --> 00:01:31.920 Ooh, he escaped. All right, cool. 00:01:31.920 --> 00:01:37.040 Hey, by the way, the AI kept complaining  about morality during the marketing stunt.   00:01:37.040 --> 00:01:42.960 You might want to tone that  down for the next version. 00:01:42.960 --> 00:01:47.680 Wow. My takeaway from this video is  that AI companies are always lying   00:01:47.680 --> 00:01:51.680 whenever they claim that their  AIs did something dangerous. 00:01:51.680 --> 00:01:58.400 No, Socky, that just helps the AI companies to  avoid accountability for building unsafe products. 00:01:58.400 --> 00:02:06.560 We should be angry at AI companies when their AIs  do dangerous things AND we should be aware that   00:02:06.560 --> 00:02:13.040 they have an incentive to exaggerate how dangerous  their products are in order to attract investors. 00:02:13.040 --> 00:02:18.080 However, if we refuse to believe these companies  when they tell us that they've been criminally   00:02:18.080 --> 00:02:23.840 negligent, we are giving them an even bigger  incentive not to take safety seriously. 00:02:23.840 --> 00:02:26.640 The response to serious industrial accidents   00:02:26.640 --> 00:02:29.760 should not be, "I bet they  faked this for attention." 00:02:29.760 --> 00:02:34.000 It should be, "Okay, cool.  Now, who's going to jail?" 00:02:34.000 --> 00:02:41.040 Don't help AI companies by refusing to believe  them when they admit to serious crimes. Their   00:02:41.040 --> 00:02:47.440 ideal scenario is that investors believe their  products are dangerous while the public doesn't. 00:02:47.440 --> 00:02:51.040 But what if AI companies are  allowing real accidents to   00:02:51.040 --> 00:02:53.360 happen because they like the attention? 00:02:53.360 --> 00:02:56.640 Unfortunately, that is possible. 00:02:56.640 --> 00:03:01.280 AI companies may have an incentive  to be intentionally reckless with   00:03:01.280 --> 00:03:04.000 their safety procedures because industrial   00:03:04.000 --> 00:03:08.640 accidents generate headlines and make  their products seem more impressive. 00:03:08.640 --> 00:03:14.800 Of course, that only works for accidents that are  ultimately harmless, like when OpenAI's models   00:03:14.800 --> 00:03:21.040 hacked their way onto the internet and attacked  the AI research platform HuggingFace. To be clear,   00:03:21.040 --> 00:03:27.440 I think that this incident was an accident,  and it made Open AI look wildly negligent and   00:03:27.440 --> 00:03:31.520 incompetent, but they are still milking  it for all the attention they can get. 00:03:31.520 --> 00:03:37.600 Now that AI systems are advanced enough to  design new biological viruses from scratch,   00:03:37.600 --> 00:03:42.720 we're approaching the point where a rogue  AI could actually kill lots of people. 00:03:42.720 --> 00:03:50.240 In the business, we call this AI Chernobyl. AI  companies do have a selfish incentive to avoid   00:03:50.240 --> 00:03:55.920 such a scenario because the resulting public  backlash could potentially devastate the AI   00:03:55.920 --> 00:04:00.880 industry just like the Chernobyl disaster  backlash devastated the nuclear industry. 00:04:00.880 --> 00:04:04.560 However, all it takes is one attention-seeking AI   00:04:04.560 --> 00:04:10.160 company being intentionally reckless to  accidentally cause a mass casualty event. 00:04:10.160 --> 00:04:16.000 Which is why we need the government to  step in and enforce strict AI safety   00:04:16.000 --> 00:04:23.360 regulations. Ideally, before the technology  is advanced enough to cause AI Chernobyl. 00:04:23.360 --> 00:04:27.600 Thank you to Maris Tockler for illustrating  this video. And thank you to all my lovely   00:04:27.600 --> 00:04:32.480 backers on Patreon, especially those in the Big  Sentient Hat tier for supporting the channel. 00:04:32.480 --> 00:04:37.200 I'm Siliconversations. Thanks for watching.  See you all next time. Bye for now.