WEBVTT - Monologue: The Agony Of GPT-5 0:00:01.840 --> 0:00:07.560 Also media hell, and welcome to this week's better Offline Monologue. 0:00:07.560 --> 0:00:17.000 I'm your host, led Zeitron, and this has been a 0:00:17.000 --> 0:00:19.799 tough week. I'm not gonna lie. Monday, we recorded a 0:00:19.880 --> 0:00:23.040 wonderful episode with Victoria's song and Ashmun Rodriguez, which got 0:00:23.040 --> 0:00:25.640 lost due to a technical four. And then I recorded 0:00:25.720 --> 0:00:28.920 this in time monolog, which then got lost to a 0:00:29.000 --> 0:00:33.600 completely different different computer, different room, different place, technical error. 0:00:34.000 --> 0:00:36.320 I love making this show for all of you. It's 0:00:36.320 --> 0:00:38.400 been kind of a pisser, but it's important we get 0:00:38.440 --> 0:00:41.680 on top of this goddamn subject. So check the episode notes, 0:00:41.720 --> 0:00:44.760 buy a challenge coin, read the newsletter. A's a premium 0:00:44.840 --> 0:00:47.320 version unrelated to this show. I'd love if you'd subscribe, 0:00:47.320 --> 0:00:50.280 and if you don't, I won't feel anything. Don't worry 0:00:50.280 --> 0:00:52.760 about it, but let's get to it. Last week open 0:00:52.800 --> 0:00:55.680 Ai launched GPT five, a new flagship model of some 0:00:55.760 --> 0:00:58.240 sort that's allegedly better at coding and writing, but in 0:00:58.240 --> 0:01:00.840 reality is much more of the same. Another model is 0:01:00.840 --> 0:01:04.320 interterminately better at benchmarks built specifically for large language moddles, 0:01:04.360 --> 0:01:07.720 because they can't do actual work. The Wall Street Journal 0:01:07.760 --> 0:01:11.080 reported late last year that it took multiple half billion 0:01:11.160 --> 0:01:13.880 dollar training runs to get GPT five off the ground, 0:01:13.959 --> 0:01:16.360 and Altman him said said in the podcast with theovon 0:01:16.440 --> 0:01:19.000 of all people at GPT five scared him and made 0:01:19.080 --> 0:01:21.440 him say, what have we done? And that's a good 0:01:21.480 --> 0:01:24.880 bloody question, Sammy. According to open Ai, GPT five is 0:01:24.920 --> 0:01:27.640 a unified system with a smart, efficient model that answers 0:01:27.680 --> 0:01:30.280 most questions, a deeper reasoning model, and a real time 0:01:30.319 --> 0:01:32.840 router that quickly decides which model to use based on 0:01:32.880 --> 0:01:36.200 conversation type, complexity, tool needs. And you're explicit in tent. 0:01:36.680 --> 0:01:38.240 I read all of that out because I wanted you 0:01:38.280 --> 0:01:40.680 to hear how convoluted GPT five is and how much 0:01:40.720 --> 0:01:43.000 effort open Ai has had to put in to create 0:01:43.040 --> 0:01:46.279 something that, based on all reports is fine. To quote 0:01:46.319 --> 0:01:51.720 Simon Willison, it just does stuff. Wowie zowie. In simpler terms, 0:01:51.800 --> 0:01:54.840 Chad GPT's version of GPT five takes the user's prompt 0:01:54.880 --> 0:01:56.960 and decides which modele to use as a result, using 0:01:56.960 --> 0:01:59.760 one of a few submodels GPT five Regular, Mini or 0:01:59.840 --> 0:02:02.600 now I know, and then spits out an output. And 0:02:02.640 --> 0:02:04.240 there are rate limits, by the way, so if you 0:02:04.320 --> 0:02:06.920 use it too much you get kicked down to many automatically. 0:02:07.160 --> 0:02:09.440 If you ask it to think about something, it will 0:02:09.520 --> 0:02:11.560 choose to engage the reasoning part of the model. These 0:02:11.600 --> 0:02:14.440 things do not think, by the way. They are probabilistic models, 0:02:14.639 --> 0:02:17.520 so reasoning is kind of like you get a prompt 0:02:17.520 --> 0:02:20.600 and then it reads the prompt with another model and says, Okay, 0:02:20.639 --> 0:02:23.320 what would the steps be to execute this? It has 0:02:23.600 --> 0:02:26.640 some returns, but recent papers suggest that it doesn't work 0:02:26.639 --> 0:02:29.480 that well anyway. Using this, you mostly have to trust 0:02:29.480 --> 0:02:31.600 the open ai will choose the best model for the job, 0:02:31.600 --> 0:02:34.200 as opposed to the cheapest one for open ai to serve, 0:02:34.240 --> 0:02:36.680 which is what I think they're actually doing. As part 0:02:36.720 --> 0:02:38.920 of the launch, open ai has also killed access to 0:02:38.919 --> 0:02:41.040 all other models that at least is planning to, and 0:02:41.080 --> 0:02:45.079 truncated user access to two choices GPT five or GPT 0:02:45.120 --> 0:02:48.239 five thinking with legacy models like four four many and 0:02:48.280 --> 0:02:51.640 their associated rate limits gone immediately for most, though some 0:02:51.680 --> 0:02:53.880 of them are back, and in sixty days of people 0:02:53.960 --> 0:02:56.919 paying two hundred dollars a month, you'll work out why 0:02:56.960 --> 0:03:00.800 I'm dithering in a second. This enraged the chat gp subreddit, 0:03:00.840 --> 0:03:03.240 with users claiming that GPT five was and I quote, 0:03:03.280 --> 0:03:05.720 wearing the skin of their dead friend, referring to GPT 0:03:05.760 --> 0:03:08.280 four oh, and another saying that GPT four point five 0:03:08.360 --> 0:03:10.639 genuinely talk to them and as pathetic as it sounds, 0:03:10.720 --> 0:03:13.800 was their only friend. And it must be clear, we 0:03:13.840 --> 0:03:15.840 can make fun of these people if we want, but 0:03:15.960 --> 0:03:19.120 this is actually genuinely sad. There is something going on 0:03:19.240 --> 0:03:21.680 here where people are so lonely that they want to 0:03:21.720 --> 0:03:24.160 talk to a chat bot. Mock them if you want, 0:03:24.840 --> 0:03:26.720 and some of you will, and I don't know if 0:03:26.760 --> 0:03:30.200 I even want to. But something is happening here and 0:03:30.240 --> 0:03:33.120 it isn't brilliant. But after a few days, Clammy Samuel 0:03:33.280 --> 0:03:36.960 and restored access to GPT four overpaid users, and then 0:03:37.120 --> 0:03:39.160 this only matched the stem the type briefly, with one 0:03:39.240 --> 0:03:41.480 user saying that their baby was back, that they cried 0:03:41.520 --> 0:03:43.400 a lot, and that they were crying as they wrote 0:03:43.400 --> 0:03:46.080 the post, ending by saying love you. I assumed to 0:03:46.120 --> 0:03:50.400 GPT four zero. Here's the problem. Though, users are now 0:03:50.440 --> 0:03:52.840 doubting that the four O that open Ai is restored 0:03:52.880 --> 0:03:55.280 is actually the same model. One post claims that four 0:03:55.360 --> 0:03:57.440 roh has lost its soul, and another says that fourro 0:03:57.560 --> 0:04:00.280 is lobotomized all the way down in one three. One 0:04:00.360 --> 0:04:02.920 user said that four row had gotten markedly worse suddenly, 0:04:03.120 --> 0:04:05.360 and another said it's definitely not the same, though others 0:04:05.360 --> 0:04:07.880 in the thread claimed that it was. Another post said 0:04:07.960 --> 0:04:11.240 legacy GPT four row is GPT five in cosplay, even 0:04:11.240 --> 0:04:13.760 as others pleaded with them and said it was exactly 0:04:13.760 --> 0:04:16.640 the same and that people were experiencing some strange phantom 0:04:16.680 --> 0:04:19.720 love for placebo effect. And I think that's actually what's 0:04:19.800 --> 0:04:22.920 going on, writ large. Chat GPT was never a success 0:04:22.960 --> 0:04:26.159 based on its actual abilities or outputs or things it 0:04:26.200 --> 0:04:28.839 could do, but a global marketing campaign perpetuated by it 0:04:28.880 --> 0:04:31.640 taking business media asleep at the wheel, or worse still, 0:04:31.960 --> 0:04:35.000 that wanted these companies to win and help them by lying. 0:04:35.160 --> 0:04:37.960 I have done a comprehensive evaluation of the last three 0:04:38.040 --> 0:04:41.440 years of press around chat, GPT and GPT itself, and 0:04:41.520 --> 0:04:43.400 you will look in the things in twenty twenty three 0:04:43.400 --> 0:04:45.680 and there's shit that's just fucking made up. There's a 0:04:45.720 --> 0:04:49.200 whole go and look up task Grabbit GPT four. There 0:04:49.200 --> 0:04:51.280 are so many I'm a link one in the fucking notes. 0:04:52.080 --> 0:04:55.039 There are people that claim that GPT four ordered a 0:04:55.160 --> 0:04:58.039 task grab it to complete a capture. Now, on top 0:04:58.080 --> 0:04:59.960 of this not being a thing that task grabbit generally 0:05:00.000 --> 0:05:02.799 it does. This is from the system card of GPT 0:05:03.000 --> 0:05:07.240 four and it claims that it hired a taskrab, except 0:05:07.240 --> 0:05:08.920 when you look at just said it messaged them. It's 0:05:09.040 --> 0:05:11.880 very clearly made up. But everyone reported it as agents 0:05:11.960 --> 0:05:15.520 existing in twenty twenty three. Aha. Every time I read 0:05:15.560 --> 0:05:19.039 this stuff, I feel little goddamn insane. But anyway, because 0:05:19.080 --> 0:05:22.120 these models do not have obvious replicable ways outside of 0:05:22.200 --> 0:05:24.800 benchmarks of testing what they can do, each user is 0:05:24.800 --> 0:05:27.159 effectively in a constant vibe check with the models, and 0:05:27.160 --> 0:05:29.920 the sycophantic qualities of GPT four oh were clearly enough 0:05:29.920 --> 0:05:35.920 to endear them to the platform. People using GPT four Oh, 0:05:36.080 --> 0:05:38.479 they couldn't tell you why it's different to GPT five 0:05:38.520 --> 0:05:41.000 other than it feels less human or doesn't do the 0:05:41.040 --> 0:05:43.320 same things, even if those things are kind of hard 0:05:43.360 --> 0:05:46.159 to define. This is what happens when you build a 0:05:46.200 --> 0:05:48.920 following for a product based on specious hype and vague 0:05:48.960 --> 0:05:51.839 promises and lines of inference, of course, and then allow 0:05:52.080 --> 0:05:54.279 users to make up the reasons that they care. You 0:05:54.360 --> 0:05:57.159 begin engaging with the gamer mindset of vibes based fandom 0:05:57.200 --> 0:05:59.920 that's completely unbreakable unless you make one subtle change that 0:06:00.040 --> 0:06:02.480 you could never see coming that breaks the illusion, leading 0:06:02.480 --> 0:06:05.839 to gamer like distrust and anger. You see, I theorized 0:06:05.839 --> 0:06:08.200 that the vast majority of chat GPT users do not 0:06:08.360 --> 0:06:10.440 know where they used it in the first place. Three 0:06:10.520 --> 0:06:12.800 years of immediate pressure to use AI that AI was 0:06:12.839 --> 0:06:15.600 the future, their boss is saying AI is important and 0:06:15.640 --> 0:06:17.720 that you would be left behind if you didn't use AI. 0:06:17.839 --> 0:06:19.880 Mean that people come to chat GPT to work out 0:06:19.920 --> 0:06:21.680 why they're using it in the first place, which has 0:06:21.760 --> 0:06:24.760 led to all sorts of bizarre emotional attachments, kind of 0:06:24.800 --> 0:06:27.080 like one's attachments to a live service game and shout 0:06:27.080 --> 0:06:29.440 out to Catharsis twenty three on Blue Sky, who made 0:06:29.440 --> 0:06:33.240 this observation. As a result, users were incredibly sensitive to 0:06:33.320 --> 0:06:36.120 changes like removing or changing a model because their association 0:06:36.200 --> 0:06:38.760 with chat GPT was based on However, GPT four zero 0:06:38.800 --> 0:06:41.680 works and sounds by ripping it out and replacing it 0:06:41.680 --> 0:06:45.080 with GPT five, users immediately felt jilted and swindled by 0:06:45.120 --> 0:06:47.800 open AI, and much like a dying live service game, 0:06:47.839 --> 0:06:50.240 any changes that have been made as a result were 0:06:50.240 --> 0:06:54.000 met with paranoia and confusion. Clamiel Altman's attempts to paper 0:06:54.000 --> 0:06:56.560 over the problem by boosting rate limits on GPT five's 0:06:56.600 --> 0:06:59.159 thinking and restoring access to GPT four to ZH and 0:06:59.200 --> 0:07:01.799 other models for pay maid users were not enough because, 0:07:01.800 --> 0:07:03.960 in a very real sense, many of those users could 0:07:04.000 --> 0:07:05.440 not tell you what they liked four to oh to 0:07:05.480 --> 0:07:08.480 begin with four oh wasn't good so much as it 0:07:08.520 --> 0:07:11.360 was an investment of time. By showing that open ai 0:07:11.520 --> 0:07:14.160 is willing to cut things arbitrarily, users can no longer 0:07:14.160 --> 0:07:17.000 trust that this investment of time is worthy, especially as 0:07:17.040 --> 0:07:19.680 many complain that it launch GPD five deleted a bunch 0:07:19.680 --> 0:07:22.880 of conversations. Now open ai sits in an odd spot 0:07:22.920 --> 0:07:27.120 where they're supposedly huge Manhattan Project level launch has been 0:07:27.160 --> 0:07:30.559 met with either apathy or agony. While they've plicated users 0:07:30.560 --> 0:07:32.720 in the short term, it's very clear that the vast 0:07:32.720 --> 0:07:35.600 majority of users this, like GPT five and power users, 0:07:35.640 --> 0:07:38.520 don't seem particularly impressed with it either. This was meant 0:07:38.560 --> 0:07:40.400 to be the big launch that changed things for open 0:07:40.400 --> 0:07:42.560 ai forever, but it's turned into something of a mass 0:07:42.600 --> 0:07:45.400 betrayal or just kind of a mass letdown. And because 0:07:45.400 --> 0:07:47.720 it's based on vibes rather than its actual ability to 0:07:47.760 --> 0:07:50.080 do something, there's very little one could do to fix 0:07:50.160 --> 0:07:53.560 this problem. It's unclear how all of this affects the 0:07:53.560 --> 0:07:56.440 company long term, but things do not seem good. Sam 0:07:56.480 --> 0:07:58.640 Wortman has already said that open ai is having to 0:07:58.680 --> 0:08:02.080 reallocate capacity for the couple of months prioritizing paying Chat 0:08:02.120 --> 0:08:05.119 GPT users of an API demand up from the current 0:08:05.160 --> 0:08:08.360 allocated capacity and commitments that they've made to their customers 0:08:08.400 --> 0:08:11.080 code for those who do not want to pay priority processing, 0:08:11.320 --> 0:08:14.800 which is now available for any developer. It's also unclear 0:08:14.840 --> 0:08:18.240 what happens next. GPT five is not the future. Open 0:08:18.280 --> 0:08:20.520 ai is running out of capacity, and their product, despite 0:08:20.560 --> 0:08:23.280 the fanfare, has no capabilities or reasons to adopt it 0:08:23.560 --> 0:08:27.160 that are really new or interesting. Years of allowing the 0:08:27.160 --> 0:08:30.200 media to spin out ridiculous narratives about what AI can 0:08:30.320 --> 0:08:33.400 or could do using vague pablem that kind of suggests 0:08:33.400 --> 0:08:35.679 they're more powerful than they are has created a pr 0:08:35.760 --> 0:08:38.200 campaign for a product that does not exist, and the 0:08:38.240 --> 0:08:41.320 seven hundred million weekly active users of chat GPT have 0:08:41.400 --> 0:08:44.880 clearly arrived there without much guidance. The attachment born of 0:08:44.920 --> 0:08:48.600 compulsion and societal pressure rather than any real use cases. 0:08:49.360 --> 0:08:52.200 When you allow people to define an indeterminately powerful tool 0:08:52.200 --> 0:08:54.679 by any standard they like, with no interest in correcting them, 0:08:54.920 --> 0:08:57.520 with no interest in guiding them, with no interest in 0:08:57.559 --> 0:08:59.600 actually showing them what it was that they were paying 0:08:59.640 --> 0:09:03.040 for others. Then it can generate staff, You'll create an 0:09:03.040 --> 0:09:05.600 attachment to it that devis any real ability you have 0:09:05.640 --> 0:09:09.000 to control things. Open ai was never forced to productize 0:09:09.040 --> 0:09:11.880 and at scale. It's a very real possibility that people 0:09:11.880 --> 0:09:14.840 have pressure by the media and society itself force themselves 0:09:14.840 --> 0:09:18.480 to find meaning in LMS somehow, even if it feels 0:09:18.559 --> 0:09:20.600 kind of stupid. And what I mean by this is 0:09:20.840 --> 0:09:22.640 you get to the product and there really isn't that 0:09:22.679 --> 0:09:25.360 much guidance. Go on an open AI's website and have 0:09:25.400 --> 0:09:27.360 a look at the chat GPT page and look at 0:09:27.360 --> 0:09:29.960 what it tells you to do. It's quite vague. You 0:09:30.000 --> 0:09:33.200 can look at it can analyze data, right, it can 0:09:33.600 --> 0:09:37.120 generate stuff. It helps you do ideas. Is that good? 0:09:37.360 --> 0:09:40.719 Am I smart for using this? Everything else you look 0:09:40.720 --> 0:09:42.640 at in the software world will tell you what it 0:09:42.679 --> 0:09:45.720 is you're using it for any consumer driven software at 0:09:45.720 --> 0:09:49.839 the very least. Yeah, chat GPT's never had to because 0:09:49.840 --> 0:09:52.080 the media for three years or two years, I guess 0:09:52.080 --> 0:09:55.600 with chat GPT has kind of just sat there doing 0:09:55.600 --> 0:09:57.960 the work for them, telling them, oh, yeah, you can 0:09:58.600 --> 0:10:01.440 use it as a powerful person a system assistant. How 0:10:01.679 --> 0:10:05.240 to assist me with what? Nasty? Kevin Russe creepy Kevin 0:10:05.840 --> 0:10:07.920 in the New York Times a few weeks ago, when 0:10:07.960 --> 0:10:12.080 he did Everyone's Using AI piece with Casey Newton, he 0:10:12.200 --> 0:10:15.200 said that it's a powerful assistant. It's like, for what 0:10:15.880 --> 0:10:18.760 you can't do that? This shit can't control my calendar. 0:10:18.960 --> 0:10:21.240 I don't want to touch in my emails. I don't 0:10:21.240 --> 0:10:23.800 think most people do either. So you've just got people 0:10:23.800 --> 0:10:26.000 to use it as a shit, our search engine, an 0:10:26.040 --> 0:10:30.040 online companion, and a brainstorming thing, which is a natural 0:10:30.080 --> 0:10:32.760 way to get people kind of addicted, but addicted to 0:10:32.760 --> 0:10:35.360 a product you don't truly control and a product you 0:10:35.400 --> 0:10:38.200 don't truly understand, one that can be swiped by just 0:10:38.240 --> 0:10:42.800 about anyone. I actually think we're on the the kind 0:10:42.840 --> 0:10:46.120 of downward spiral for this shit. I am winking like 0:10:46.200 --> 0:10:49.800 a pig and squawking like a bird watching this happen. 0:10:49.880 --> 0:10:52.520 Even core Weavers crashing as I record this. They're down 0:10:52.600 --> 0:10:55.520 seventeen point nine one percent. Still up too much, though, 0:10:56.520 --> 0:10:58.599 And I really do think something has shifted thanks to 0:10:58.679 --> 0:11:02.280 GPT five, I'm a excited and like Doctor Stones once said, 0:11:02.400 --> 0:11:05.160 get excited because I think the next few months, like 0:11:05.280 --> 0:11:09.560 next year, is going to be chuckle heavy's. We're going 0:11:09.640 --> 0:11:12.520 to be whimsy pilled as we go through the remainder 0:11:12.640 --> 0:11:14.640 of the AI Boom, and I'll be here to guide 0:11:14.679 --> 0:11:16.280 you through it. Thanks for listening.