WEBVTT - Monologue: OpenAI Is Getting Desperate

0:00:02.160 --> 0:00:08.119
<v Speaker 1>Zone Media. Hello you, it's your better offline monologue and

0:00:08.160 --> 0:00:17.920
<v Speaker 1>I'm your host ed ZiT trum. Now next week I'm

0:00:17.920 --> 0:00:19.400
<v Speaker 1>going to have a two part of that digs into

0:00:19.400 --> 0:00:21.840
<v Speaker 1>Microsoft Stata senter pull back an open ai is shaking

0:00:21.920 --> 0:00:25.200
<v Speaker 1>you funding situation. But this week's monologue focuses an open

0:00:25.200 --> 0:00:28.880
<v Speaker 1>AI's new model, GPT four point five. You may be

0:00:29.040 --> 0:00:31.880
<v Speaker 1>wondering what it does differently to GPT four oh, or

0:00:31.920 --> 0:00:34.159
<v Speaker 1>claudes on at three point seven, or any number of

0:00:34.200 --> 0:00:36.760
<v Speaker 1>other large language models, And if I'm honest, I have

0:00:36.800 --> 0:00:40.800
<v Speaker 1>absolutely no idea. Thankfully neither does Sam Ortman, CEO of

0:00:40.800 --> 0:00:43.480
<v Speaker 1>open Ai, who said, and I quote, the GPT four

0:00:43.479 --> 0:00:47.120
<v Speaker 1>point five was the first model that feels like talking

0:00:47.159 --> 0:00:51.560
<v Speaker 1>to a thoughtful person to him, which makes me wonder

0:00:51.680 --> 0:00:55.000
<v Speaker 1>what the other models have been like. So I went

0:00:55.040 --> 0:00:57.680
<v Speaker 1>back and look. So compare this to the launch of

0:00:57.720 --> 0:01:01.520
<v Speaker 1>GPT four oh, which Altman called open AI's best model ever,

0:01:01.640 --> 0:01:05.679
<v Speaker 1>saying that it was fast, smart, natively multimodal, referring to

0:01:05.720 --> 0:01:07.880
<v Speaker 1>the ability to accept text as well as audio and

0:01:07.959 --> 0:01:11.360
<v Speaker 1>video and photos as well, and available to all chat

0:01:11.400 --> 0:01:14.520
<v Speaker 1>GPT users, including on their free plan, adding that it

0:01:14.640 --> 0:01:18.400
<v Speaker 1>was a very good model, especially at Coding. By contrast,

0:01:18.440 --> 0:01:22.160
<v Speaker 1>Alltman summarized GPT four point five as a giant, expensive model,

0:01:22.480 --> 0:01:25.520
<v Speaker 1>one that required hundreds of thousands of GPUs to launch

0:01:25.520 --> 0:01:28.760
<v Speaker 1>beyond chat GPT pro. It started there. It's still not

0:01:28.880 --> 0:01:32.440
<v Speaker 1>as I record this out to plus users or free users. Now.

0:01:32.520 --> 0:01:35.160
<v Speaker 1>GPT pro of course is open AI is two hundred

0:01:35.160 --> 0:01:39.960
<v Speaker 1>dollars a month subscription, and it's unclear when plus, which

0:01:40.000 --> 0:01:41.640
<v Speaker 1>is the twenty buck a month one will get it,

0:01:41.680 --> 0:01:44.920
<v Speaker 1>but apparently it's in the next few days. Aortman also

0:01:44.920 --> 0:01:47.880
<v Speaker 1>added that GPT four point five isn't a reasoning model

0:01:47.960 --> 0:01:50.680
<v Speaker 1>and won't crush benchmarks on account of it being a

0:01:50.720 --> 0:01:53.480
<v Speaker 1>different kind of intelligence that has and all of these

0:01:53.520 --> 0:01:56.880
<v Speaker 1>are quotes magic to it that Sam Mortman had not

0:01:57.000 --> 0:02:02.360
<v Speaker 1>felt before. Yeah, just you know, shit's not doing well

0:02:02.800 --> 0:02:06.120
<v Speaker 1>when you have to just be like it's magic, It's

0:02:06.160 --> 0:02:09.240
<v Speaker 1>it's literally magic. I made magic. Now what does the

0:02:09.240 --> 0:02:13.519
<v Speaker 1>magic do? I'm really not sure. In fact, it's pretty

0:02:13.520 --> 0:02:15.960
<v Speaker 1>difficult to find exactly what it is that GPT four

0:02:16.000 --> 0:02:18.120
<v Speaker 1>point five does differently, or what it's good at, or

0:02:18.160 --> 0:02:21.240
<v Speaker 1>indeed really anything about him. BENJ. Edwards over At asked

0:02:21.240 --> 0:02:24.639
<v Speaker 1>Tenneker had one developer called it a lemon. GPT four

0:02:24.680 --> 0:02:27.960
<v Speaker 1>point five costs an incredible seventy five dollars per million

0:02:28.000 --> 0:02:30.880
<v Speaker 1>input tokens prompts and data pushed into a model, and

0:02:30.919 --> 0:02:33.840
<v Speaker 1>one hundred and fifty dollars per million output tokens is

0:02:33.840 --> 0:02:36.200
<v Speaker 1>in the thing it creates. A token is like zero

0:02:36.240 --> 0:02:39.920
<v Speaker 1>point seven. I think one token is maybe three words.

0:02:40.160 --> 0:02:42.600
<v Speaker 1>Someone will get up my ass for this. Nevertheless, this

0:02:42.680 --> 0:02:45.240
<v Speaker 1>seems like a lot. It isn't when you're running a company.

0:02:45.600 --> 0:02:48.080
<v Speaker 1>And by the way, this is roughly three thousand percent

0:02:48.160 --> 0:02:51.080
<v Speaker 1>more expensive for input tokens and fifteen hundred percent more

0:02:51.120 --> 0:02:53.680
<v Speaker 1>expensive for output tokens than GPT four to zero for

0:02:53.800 --> 0:02:57.680
<v Speaker 1>results that open ai co founder Andredge Carpathy described as

0:02:57.800 --> 0:03:00.960
<v Speaker 1>a little bit better and awesome, but also not exactly

0:03:00.960 --> 0:03:04.200
<v Speaker 1>in ways that are trivial to point to. That translates

0:03:04.200 --> 0:03:06.960
<v Speaker 1>to it's a little bit better, but I can't really

0:03:07.000 --> 0:03:10.399
<v Speaker 1>tell you why. And yes, you're gonna hear me say

0:03:10.400 --> 0:03:13.120
<v Speaker 1>something similar in next week's episode, because the larger picture

0:03:13.160 --> 0:03:16.119
<v Speaker 1>for open ai right now is pretty fucking dire considering

0:03:16.120 --> 0:03:18.440
<v Speaker 1>their main backer, soft Bank, has to borrow billions of

0:03:18.440 --> 0:03:21.960
<v Speaker 1>dollars to fund them. Nevertheless, back to four point five

0:03:22.360 --> 0:03:24.639
<v Speaker 1>Since launch, which was for some reason, on the day

0:03:24.639 --> 0:03:27.760
<v Speaker 1>that Sam Ortman's child was being born in the hospital,

0:03:29.040 --> 0:03:32.280
<v Speaker 1>He's been posting some really weird shits since, though. A

0:03:32.280 --> 0:03:35.080
<v Speaker 1>few days after launch, Aortman claimed that GPT four point

0:03:35.080 --> 0:03:38.080
<v Speaker 1>five was the first time people had been emailing with

0:03:38.120 --> 0:03:41.440
<v Speaker 1>such passion, asking the open AI promised never to stop

0:03:41.480 --> 0:03:44.320
<v Speaker 1>offering a specific model or even replace it with an update,

0:03:44.600 --> 0:03:47.240
<v Speaker 1>at which point I assume everybody in the room started

0:03:47.280 --> 0:03:50.760
<v Speaker 1>clapping and they saluted Sam Mortman and said, thank you,

0:03:50.840 --> 0:03:53.280
<v Speaker 1>sir for making this happen. And by the way, what

0:03:53.360 --> 0:03:55.360
<v Speaker 1>I'm suggesting is that no one's ever done this, or

0:03:55.400 --> 0:03:58.560
<v Speaker 1>like one freak did, or maybe Aortman emailed it to himself,

0:03:59.080 --> 0:04:04.440
<v Speaker 1>who just shut up? Just your company burns five billion

0:04:04.480 --> 0:04:06.400
<v Speaker 1>dollars a year and the best you've got is this

0:04:06.480 --> 0:04:10.360
<v Speaker 1>warmed up dog shit about people marine todding you over

0:04:10.440 --> 0:04:13.400
<v Speaker 1>your model and never taking it away. Has opening I

0:04:13.520 --> 0:04:17.960
<v Speaker 1>ever even taken away a model? Jesus fucking these companies anyway.

0:04:18.520 --> 0:04:21.400
<v Speaker 1>A few days later, Altman posted a conversation where he

0:04:21.480 --> 0:04:24.280
<v Speaker 1>asked GPT four point five if it believed it was real,

0:04:24.600 --> 0:04:26.560
<v Speaker 1>leading to a series of bullet points with things like

0:04:26.640 --> 0:04:29.400
<v Speaker 1>what do we mean by real only for GPT four

0:04:29.400 --> 0:04:31.719
<v Speaker 1>point five, saying that it believed that it was not

0:04:31.800 --> 0:04:35.159
<v Speaker 1>an independent consciousness, but rather a structured experience happening within

0:04:35.480 --> 0:04:38.680
<v Speaker 1>your consciousness, referring to Sam Altman, which is the kind

0:04:38.680 --> 0:04:40.919
<v Speaker 1>of shit that's only impressive if you're an imbecile or

0:04:40.920 --> 0:04:43.279
<v Speaker 1>so stoned you've texted date of your friends the question

0:04:43.520 --> 0:04:45.680
<v Speaker 1>what if the joker was Batman? And by the way,

0:04:45.720 --> 0:04:47.800
<v Speaker 1>the answer to that is called The Batman Who Laughs

0:04:47.839 --> 0:04:50.280
<v Speaker 1>and it's one of the worst comics ever written. If

0:04:50.279 --> 0:04:52.200
<v Speaker 1>you want to talk to me about DC Metal, please

0:04:52.240 --> 0:04:55.800
<v Speaker 1>email me. It's easy. That's e z or z. If

0:04:55.800 --> 0:04:59.240
<v Speaker 1>you're Canadian or British at Better Offline dot com I

0:04:59.320 --> 0:05:02.320
<v Speaker 1>a really if you are working for DC Comics right

0:05:02.320 --> 0:05:04.640
<v Speaker 1>now and you had anything to do with Death Metal

0:05:04.760 --> 0:05:06.920
<v Speaker 1>or The Batman Who Laughs, you and I have a grievance.

0:05:07.240 --> 0:05:09.920
<v Speaker 1>You and I need to talk. Sorry what this is

0:05:09.960 --> 0:05:13.680
<v Speaker 1>a tech pop us right back to open AI. More worryingly,

0:05:13.960 --> 0:05:17.520
<v Speaker 1>Samultman posted an idea for paid plans where your twenty

0:05:17.520 --> 0:05:20.919
<v Speaker 1>dollars plus subscription converts the credits you can use across

0:05:20.920 --> 0:05:24.080
<v Speaker 1>speech just like Deep Research O one. GPT four point five, Sora,

0:05:24.200 --> 0:05:26.800
<v Speaker 1>and so on, with no fixed limits per feature, and

0:05:26.839 --> 0:05:29.080
<v Speaker 1>you choose what you want. If you run out of credits,

0:05:29.360 --> 0:05:31.960
<v Speaker 1>you can buy more. This, to be clear, is an

0:05:32.000 --> 0:05:34.840
<v Speaker 1>attempt to raise prices without actually raising them by attempting

0:05:34.839 --> 0:05:38.240
<v Speaker 1>to limit usage of open AI's more expensive models. Chat,

0:05:38.279 --> 0:05:41.040
<v Speaker 1>GPT plus and other subscriptions give you a limit, for example,

0:05:41.080 --> 0:05:43.560
<v Speaker 1>a limit of eighty messages every three hours on GPT

0:05:43.680 --> 0:05:46.320
<v Speaker 1>four to oh, but using one doesn't limit your use

0:05:46.320 --> 0:05:49.679
<v Speaker 1>of other products. Here, open ai is trying to create

0:05:49.680 --> 0:05:52.040
<v Speaker 1>a rent seeking model where power users have to pay

0:05:52.200 --> 0:05:54.600
<v Speaker 1>for more credits if they want to use, say open

0:05:54.640 --> 0:05:56.720
<v Speaker 1>ai is more expensive models like Sora and O one,

0:05:56.880 --> 0:05:59.040
<v Speaker 1>and I imagine any situation like this will be one

0:05:59.040 --> 0:06:01.479
<v Speaker 1>where they hope that people simply won't use their credits

0:06:01.680 --> 0:06:03.520
<v Speaker 1>or overuse them and have to pay for top ups.

0:06:04.040 --> 0:06:06.839
<v Speaker 1>This is, of course all theoretical, but it heavily suggests

0:06:06.839 --> 0:06:09.920
<v Speaker 1>that open ai is getting desperate. And now the information

0:06:10.040 --> 0:06:12.880
<v Speaker 1>is reporting that open ai executives have told some investors

0:06:12.880 --> 0:06:16.200
<v Speaker 1>that they will be charging two thousand dollars per month

0:06:16.400 --> 0:06:19.080
<v Speaker 1>for their low end agent product. And yes that's a

0:06:19.160 --> 0:06:24.080
<v Speaker 1>quote sold to And again I quote high income knowledge

0:06:24.080 --> 0:06:27.560
<v Speaker 1>workers with supposed mid tier agents for software development costing

0:06:27.720 --> 0:06:31.400
<v Speaker 1>possibly ten thousand dollars a month, with supposed PhD level

0:06:31.440 --> 0:06:34.440
<v Speaker 1>research agents costing twenty thousand dollars a month, and I

0:06:34.440 --> 0:06:36.960
<v Speaker 1>will tell you the PhDs I know would probably do

0:06:37.000 --> 0:06:39.560
<v Speaker 1>it for half, and they'd even work for an annoying

0:06:39.560 --> 0:06:42.560
<v Speaker 1>asshole like Sam Altman. Now you may wonder what any

0:06:42.600 --> 0:06:44.640
<v Speaker 1>of these things do, and the answer is that neither

0:06:44.839 --> 0:06:47.200
<v Speaker 1>I nor the Information know. As of right now. The

0:06:47.240 --> 0:06:50.680
<v Speaker 1>only operational or agent, open ai has his operator, open

0:06:50.720 --> 0:06:53.800
<v Speaker 1>AI's agent that sometimes successfully uses a web browser search

0:06:53.839 --> 0:06:57.240
<v Speaker 1>for something in minutes, which would usually take you seconds.

0:06:57.920 --> 0:07:00.240
<v Speaker 1>The Information attempted to suggest that the two thousand a

0:07:00.240 --> 0:07:03.159
<v Speaker 1>month agent would be some sort of thing that could

0:07:03.240 --> 0:07:05.839
<v Speaker 1>sort through and rank sales lead But I'm sorry, do

0:07:05.920 --> 0:07:08.040
<v Speaker 1>I really have to read this shit with a straight face?

0:07:08.120 --> 0:07:10.560
<v Speaker 1>Twenty thousand dollars for a PhD level agent? What the

0:07:10.600 --> 0:07:13.040
<v Speaker 1>fuck does that mean? What would it do? Why do

0:07:13.120 --> 0:07:15.960
<v Speaker 1>these companies? I get emails every week having to justify

0:07:16.080 --> 0:07:19.920
<v Speaker 1>my fucking cynicism, But these shittheads, they're allowed to just

0:07:20.160 --> 0:07:22.080
<v Speaker 1>make up stuff and it leak it to the information

0:07:22.160 --> 0:07:24.560
<v Speaker 1>The Information publishes. It were all meant to be impressed.

0:07:24.600 --> 0:07:28.960
<v Speaker 1>What the fucking what the fuck I'm allowed to rent

0:07:28.960 --> 0:07:31.000
<v Speaker 1>on these? They're allowing me to rant on these. It's

0:07:31.120 --> 0:07:33.760
<v Speaker 1>just it sickens me. I have had this week at

0:07:33.840 --> 0:07:37.240
<v Speaker 1>least five people email me and be like, well ed,

0:07:37.280 --> 0:07:39.760
<v Speaker 1>what would it take to change your mind about this stuff?

0:07:39.800 --> 0:07:42.320
<v Speaker 1>Why do I have to fucking do it? Why do

0:07:42.400 --> 0:07:45.680
<v Speaker 1>I The multi billion dollar companies do a dogshit job

0:07:45.720 --> 0:07:48.920
<v Speaker 1>of actually explaining this stuff or selling it. They lose

0:07:48.960 --> 0:07:50.880
<v Speaker 1>billions of dollars. But I'm the guy who has to

0:07:50.960 --> 0:07:54.200
<v Speaker 1>justify myself. Oh well, I'll keep doing it. Nevertheless, NES's

0:07:54.240 --> 0:07:56.000
<v Speaker 1>old At the bottom of this article's a far more

0:07:56.040 --> 0:07:59.560
<v Speaker 1>obvious payale horse. Open ai is planning to charge twenty

0:07:59.600 --> 0:08:02.400
<v Speaker 1>percent to thirty percent of pro customers the two hundred

0:08:02.400 --> 0:08:04.880
<v Speaker 1>dollars a month subscription that loses the money every time,

0:08:05.200 --> 0:08:07.880
<v Speaker 1>a higher price because of how many research queries they're

0:08:07.920 --> 0:08:11.160
<v Speaker 1>doing with Altmann, according to the information, suggesting some sort

0:08:11.200 --> 0:08:14.000
<v Speaker 1>of hey guessoir a la carte or pay as you

0:08:14.040 --> 0:08:18.520
<v Speaker 1>go approach. I want to be clear about something. This

0:08:18.640 --> 0:08:21.040
<v Speaker 1>is not a company that's cooking. This is not a

0:08:21.080 --> 0:08:25.120
<v Speaker 1>company that's worked out anything. Open Aiye is unprofitable, unsustainable,

0:08:25.160 --> 0:08:27.920
<v Speaker 1>and deeply, deeply lost. These are the actions of a

0:08:28.000 --> 0:08:32.080
<v Speaker 1>desperate company run by a desperate man. You have only

0:08:32.120 --> 0:08:34.320
<v Speaker 1>Sam Morman had a thoughtful friend to talk about all

0:08:34.360 --> 0:08:35.800
<v Speaker 1>these problems too,