GPT-5.6 Reactions, Phoebe Gates’ Phia Drama, Hollywood Chases Internet Horror | Diet TBPN

0:01John Coogan:SK Hynix is up. Successful IPO, up 14 % in the Wall Street debut. The South Korean memory chip giant, SK Hynix, made its NASDAQ debut today, raising$26.5 billion in the largest ever first-time U.S. share sale by a foreign company. Demand reportedly exceeded the available shares more than sevenfold. But to give a sense of the scale of this thing, normally when we hear about 7x oversubscribed IPOs, you might be expecting like 100 % pop, 80 % pop. But at this level of price and demand, you're just seeing a 14 % pop, but still a great debut nonetheless. The company's American depository receipts opened, the ADR shares, opened at 170, 14 % above the 149 they were offered at, giving SK Hynix a market cap of roughly$1.03 trillion.

1:00John Coogan:We got another one, another, another one, another trillion dollar company in the AI boom. The blockbuster listing caps an extraordinary run for memory manufacturers fueled by surging demand from AI systems, particularly for high bandwidth memory used alongside advanced AI accelerators. SK Hynix, Samsung, and U.S.-based Micron dominate the global memory market. And SK Hynix's Korean-listed shares are up over 600 % in the past year. Tyler, is that good? rushed to gain exposure to one of the AI boom's most critical bottlenecks. But it's been tricky to get shares and participate in SK Hynix because you've got to go over to the COSPI.

1:38John Coogan:You've got to go over to the South Korean index. Instead, now you can just head to the NASDAQ and get your ADR shares in SK Hynix. So congratulations to the whole team over there, a massively successful IPO. The Wall Street Journal asks a question. They say, why is SK Hynix trading under a temporary ticker symbol? They're currently trading under the temporary ticker SKHYV before reverting to a permanent ticker on Monday, SKHY. The V in the temporary symbol indicates the shares are trading on a when-issued basis. They haven't been fully issued yet, meaning the exchange is allowing investors to buy and sell the stock on its first U.S.

2:19John Coogan:trading day before the transaction actually settles. They're sort of being eased into the American markets. This is not the case with all ADRs, all American depository receipts. but rather a tool used to close the gap when a transaction has been announced, but the actual delivery of the shares hasn't finished. So all in process now to bring SK Hynix to the American markets. And that looks like a lot of red, but the stock is up. It just opened at 170. And so the fact that it is flat relative to where it opened does not reflect the 14 % pop or 16 % pop. What was it? 14 % pop from the actual issuance price.

2:55John Coogan:Most of the folks that bought shares in the IPO, that$26 billion that went into the company. Of course, situational awareness. There were a few others. Bailey Gifford all participated early, at least according to early news. I assume that those transactions closed. They are all sitting up 14 % today, and we'll be continuing to follow this. Kanner in the chat says CERN has only been off for a week, and John is dressing casual. Ooh, you think it's a conspiracy? I like it. I like it. 5.6 launched yesterday. 5.6 Soul and Soul Ultra. And people are having fun with it. More and more demos are hitting the timeline.

3:35John Coogan:There's a lot of controversy over the desktop app for Mac. We can get into that. But first, let's play a little bit of this video. Someone tried to recreate all of Interstellar, truly trying to infuriate Christopher Nolan the most, probably. This is the exact opposite of what he wants to see, I imagine. considering that he doesn't even use CGI very often. But this is an incredible demo, although I don't think this will be picking up a billion dollars at the box office anytime soon. But it is incredible that you're able to sort of go from prompt to 3D model to rendering so quickly. Very, very cool demo.

4:19John Coogan:And it feels like a new meme, a new art form. You have to be creative with it. merely recreating interstellar is not going to do it it's not going to do anything other than just be an impressive demo of the technology but uh when you look at what people have done with blender you can imagine that with creativity added to this in a new storyline and something that leans into the rough edges of this particular well there's also gonna there's also gonna be a whole meta of like i recreated interstellar but i cast conor mcgregor as yeah you know and i imagine that This gets even more powerful when you use style transfer and video models on top.

4:58John Coogan:I've been seeing a lot of people block out AI videos, either just using their iPhone. So they'll just have, like, you're playing an actor who's playing, like, an old character. And then you're reading the lines. You're giving the delivery. And then that is used as the driving video for some of these AI video models. And you get much better results with that. And I've also seen people block out little examples of driving scenes or different mechanical scenes in Blender and then use an AI video model to take it to high fidelity. And because the video model has an underlying reference that is deterministic and physically accurate, even though it's low fidelity and very blocky, you wind up with a better AI video at the end of the project.

5:45John Coogan:So you can imagine someone taking an idea, puppeteering a 3JS model like what's happening here, and then adding basically a style transfer layer on top with AI video to get to something that's a lot more photo real and more striking. But again, you wind up in this situation where you can go and recreate Interstellar perfectly, but at the end you just have Interstellar, which already exists and is available to stream. And so you didn't really get anything in the process. you have to bring something special to the table. But I'm excited for the inevitable mashups of this, the Conor McGregor starring in Interstellar or something like that.

6:21John Coogan:Also, there's a broccoli farmer that's running his farm on GPT 5.6, apparently. This was in the demo video, but Daniel Tenryo shares a very funny meme showing the, well, I don't even know what aesthetic this is. This is the Share Zone aesthetic. That was the account that shared all of this. the extreme skeleton with the two revolving pistols. Very fun. Daniel Tan Riro shares some more information. He says, we were promised flying cars, and instead we got a broccoli farmer running his farm with GPT 5.6. I don't know. It's still relevant to show a demo that is as far afield. Everyone knows how these models can be used for research, how these models can be used for coding.

7:03John Coogan:That gets a little boring. I want to imagine something completely different that I can then pull into my daily life. What is the equivalent task that I'm doing in my life that I can take? We sort of differ on this sometimes. You think that a lot of these things are self-explanatory. They don't. This is Joe Weisenthal's position, too. It says, like, AI has no learning curve. I disagree. I think that it doesn't necessarily have a learning curve, but there is an educational process that happens when you're presented with a blank box. You can type anything. People need to understand the fringes of their imagination.

7:37John Coogan:What is possible? and then they can go and reach for that tool. But the tools are getting sharper and sometimes they get sharper without the box actually changing because it's just a box of what do you want to build or what do you want to do? Yeah, not everyone has a tiler where I tend to prompt tiler. Yes, yes, yes. It goes into codex. Yes. There was another video that was shaking up the timeline. People were going back and forth on how fast this model was building in, is this Blender? Blender. This is Blender, right? Yeah. So it's building a 3D canon, modeling it and adding different primitives, different spheres and toruses.

8:17John Coogan:And the video is not sped up. And Chris says 5.6 soul and what I assume is 750 tokens per second in the wild. That is the preview of Cerebris. But this apparently was not actually. Yeah, it's Soul Ultra on fast mode. So we should be seeing even faster rollouts. Although every time I see this, I think. And this is just computer use. Yeah. Like this is not like an integration. Yeah, yeah. Yeah, it is an interesting experience watching a machine use your machine at a rate that would be completely impossible. unless it was like your entire bit online was like, I just use software really fast. Yeah.

9:09It is, yeah, it's an interesting sort of feel the AGI moment.

9:15John Coogan:Anyway, Paki McCormick is having fun with 5.6 Soul. He says it one-shotted this on its own. He didn't ask for a new logo. It's just like, look, bro, you need a new logo. I actually think you should adopt this. It looks cool. I like it. He says it's part of a web design based on a link to a Cosmos. Not perfect, but way better than any other time I've tried. And that's been the big talking point is like they were behind on, or the GPT models were sort of behind on front end in design. And this is a big leap forward in that. So I'm excited to see. Nick on the Codex team is asking, is this image gen?

9:47I think it might be.

9:48John Coogan:Well, I mean, that's the thing about Sol and Codex. But it's been generating, now if you're asking it to build a site, it will generate you a preview image. Image. And so I think this is the preview image for the site. Yep. But the really confusing thing is that the image model now has tool use and can scaffold things out in HTML if it needs to. And so like all of these different models are sort of blending together at this point. Michael had a very funny interaction where he got banned for open AI for cyber abuse. He had no idea what he did. He pasted the ban notice into Codex, which I don't know how he was still using Codex if he was banned.

10:28John Coogan:but he said he asked it to figure out what triggered the ban. Codex found that it had asked for an API key to his own server. Codex wrote the appeal, submitted the appeal. A few minutes later, the appeal was auto-approved by an AI at OpenAI. So it was banned by AI, convicted by AI, defended by AI, and pardoned by AI in about 10 minutes. There's a bunch of pushback on the new ChatGPT desktop app because it is the super app. It unifies codex. The new ChatGPT app is just the codex app. And when you install it, you can keep the old ChatGPT app as ChatGPT classic. But a lot of people don't like that because the codex app is written in Electron, which means it's a bunch of like web views, web browsers effectively, as opposed to the ChatGPT app, which was the native app and had a bunch of integrations that people had sort of come to know and love.

11:23John Coogan:And so this is like a change in the user experience if you're deeply embedded in a particular workflow. The flip side is like, I just don't know how many people were daily driving ChatGPT as a Mac desktop app. We've heard from someone on the team that was effective for cheating and we won't dock who it was. Well, so I would sometimes use it because there's like a very nice shortcut. Yeah. So it was like a control plus or command plus is the normal like main search on Mac. But it was like a very easy way to interact with the models. Yeah. I wonder if that can just be, can that just be, like, can that shortcut just be created with the new codex or ChatGPT work?

12:05John Coogan:Yeah, probably. Where you just bind that same key shortcut. But, I mean, it is making the user do work, which is always a tricky thing. But I feel like for most people's workflow, at least for mine, it was the desktop app that I used was Codex. And then when I wanted to interact with ChatGPT in a chat flow, I would do a new tab. And then on my phone, I would have the ChatGPT app. And then the ChatGPT app on my phone could talk to Codex on the desktop. But I never actually needed to open the ChatGPT desktop app. But maybe I'm just weird. But I don't know. We'll see as this rolls out. Same flow for me.

12:41The one thing that is...

13:11It feel like Codex is going away over time. Absolutely not our intention. We love Codex, and it is here to stay. Long story short, usage limits are getting reset twice today, so people can keep experimenting.

13:23John Coogan:Hammer it. Plan out some sites. There's also some confusion over the difference between ChatGPT Work and ChatGPT Codex, and I wonder when people will go to one or the other. I just want one place to do things. Yeah, the Omni box that can do anything for you. And I think that's obviously where it goes. But there's a walk, crawl, run. There's integration points. And people have ingrained experiences and workflows and habits. And there's all sorts of different tradeoffs here. There is no enterprise Google search. Like you just use Google at work. You use Google at home. It's all one box that gives you answers.

14:03John Coogan:And there was never like a dividing line between these two things. And so that's probably where all of this goes ultimately. Let's tell you about Codex. There we go. Codex is a powerful workspace for getting things done with AI agents, whether you're writing code, analyzing data, creating content, or automating business workflows. Codex helps move projects forward from start to finish. People are saying it's a horse for the mind. A horse for the mind. So the MetaBull case. SemiAnalysis wrote a long piece about the future of MSL, what's going on there. Very interesting contrarian take. I think a lot of people were confused by the expansion of activities and sort of an entirely new business line with B2B sales.

14:48John Coogan:And we've discussed the bull cases here, the benefits. Obviously, there's a ton of benefits with the ad model. They have the cash flow to draw down on. There's a lot of different things that we can go through. But let's read some of the reactions first, and then we can go into the semi-analysis piece. So Boring Business says, People don't realize that Zuckerberg is all in on AI because he refuses to pay a 30 % tax to platform owners for the rest of his life. Meta caught the shift toward mobile back in the early 2010s, but they were never able to create their own hardware platform to capture users in their ecosystem.

15:21John Coogan:As a result, they got stuck paying a 30 % tax to Apple on in-app purchases, which has cost Meta billions of dollars in profit. That makes no sense to me. no one buys ads via in-app purchases every advertiser i know uses the web-based also also i add ads i believe count as like a real world service so there's not like there's there's a lot of things you order in the that you order on mobile apps like groceries kind of a digital good but yes i i agree with you that it would be yeah but but it's like eyeballs right it's like impressions uh which like eyeballs yes human attention is not a digital good it is like a real world thing yes so if i open up the amazon app on my iphone and i buy a t-shirt apple does not get 30 of that because that's a real world or doordash good example or a movie ticket yes But if I go to Kindle or Netflix and I subscribe or even like the X subscription premium, Apple will take 30 % if I buy Fortnite V-Bucks or Roblox or any digital good.

16:34John Coogan:And I could see Apple saying, look, if you want to be able to, because in the meta platforms, you can be an Instagram and you can just say, I want to boost this post. And you can put down your credit card and pay for that. And you can get started on the advertising flow as a very small business, I think, through effectively in-app purchases. Yeah. So maybe Billions is right. The bigger thing is just like you have this one of the largest companies in the world. Yeah, yeah. And you have to constantly be playing nice and sort of sucking up to Apple, who is sort of actively who doesn't like you like Apple that you can imagine.

17:12What do Apple executives think about Meta if you were to ask them over coffee off the record? And it would be like, we don't like Meta. And so that's fundamentally like a very, very uncomfortable position to be in. And so he doesn't want to be in a position running his company where he has to ask for permission for things. And so an example with that, somebody brought it up. I think Andrew Metta lost billions from Apple because of the permissions prompt. Yes, that's true. You know, making it harder to just run their core advertising business. Yeah. And also there's so much that's like very unpredictable about where AI is going.

17:49Like it's possible that we get to the point where you don't go to a bunch of different social media platforms and you're just like talk. Your agent is just kind of surfacing content for you on the fly. Like there's so much like potential disruption. I think the stance is like I would rather participate in this massive technology cycle and have a shot at being a big part of it. Then like just like hope nothing bad happens to me. The other thing I've always felt is like, you know, with the glasses, you're just going to be able to, like he sees a world, like the glasses are more of a new platform bet right now than AI.

18:29And I just think he wants to be in a world where you're wearing your very, very cheap pair or maybe even free pair of glasses around. They act as your phone and your computer. And you can look at stuff and say, buy me that. And he's getting a cut of it. you're walking by a restaurant and it pops up an offer and it says like free boba for any orders over 20 and he's getting a cut of that right so like that will be like some type of like always on intelligence so yeah it all it all totally makes sense even if it is like deeply sort of like chaotic and at times confusing yeah uh at least over the last year the funny thing about

19:07John Coogan:the glasses as a platform is AI spending is so huge that he can probably bring back a lot of the reality lab spending and sort of just tuck it in a bucket and people won't notice if there's an extra eight. I'm not leaving. I'm not leaving. I'm not leaving.

19:31John Coogan:Yeah. I mean, if there's an extra eight billion spent on the next version of the meta ray-ban displays or something people will be like ah it's a rounding error compared to the 200 billion on data centers or whatever. Ari says the tax meta pays to Apple is the fact that they can't track users cross-platform. Yeah yeah totally the other interesting thing is like I do think the uh the screen time is a is at least like a flashing warning light like I've personally noticed as an early adopter of LLMs that my screen time has switched much more towards ChatGPT and LLMs and deep research reports, just because if I see something interesting, oh, there's a car video on Instagram, we'll often switch over and then start prompting to actually learn more.

20:18John Coogan:Whereas before that would have required a whole bunch of Google searches. You might have to sit down and collect a document. You're not actually going to do that. Now I've been noticing that I'm much more in the hour a day on ChatGPT versus hour a day on Instagram. And so I don't know if I'm just like the early adopter that will never do that, but you can certainly see that people are spending a lot of time with AI systems. And as you've put it before, like wherever consumers, wherever people are spending time is an important place to have a foothold at least. Semi-analysis said at the simplest level, there are three things you need to build a true frontier model, data, talent, and compute.

20:54We believe Meta is the only hyperscaler slash neolab on track to be world class at all three and therefore has the best chance of catching up with anthropics slash open ai we'll explain why in full detail below and so alex wang uh quoted this uh in a funny uh in a funny way uh he said compute daddy dylan patel has spoken um and yes the stock is up six percent wang uh quote quoted uh poly quote tweeted Polymarket Money, which said meta jumps over 6 % of the open as investors react to yesterday's MuseSpark 1.1 release. Again, I don't really think that they were reacting to MuseSpark as much as SemiAnalysis saying there's a chance.

21:43So yeah, Alex Wang said nice with the thumbs up. He's pumping the stock. The crazy thing is that SemiAnalysis just completely leaving out SpaceX

21:55John Coogan:and they didn't leave out SpaceX and Google. They said, but, but, but saying, they said SpaceX is selling$26 billion a year worth of GPUs to Anthropic slash Google. And that's why they're not in the race. Even though it is odd because Meta might wind up selling a lot of GPUs. It's sort of unclear what the breakdown will be, but they're at least laying out a narrative that if Meta picks it up, would differentiate them from SpaceX AI in some ways, I think. Yeah, I agree with you. But they're essentially saying Gemini doesn't have, because Gemini has compute and data. They're effectively saying Gemini does not have talent.

22:40Yeah.

22:41John Coogan:Right? They're saying you need all three. Meta took their data efforts to another level in late May by announcing a new applied AI engineering org as part of their most recent round of layoffs restructuring. 3 ,000 engineers, which includes 70 % of their new grads and a significant number of seniors, will now be making RL tasks and environments full-time. We think this is an extremely underappreciated advantage for MSL. The question is, if this becomes the standard playbook, you would imagine that Google DeepMind would be able to do the same thing. They have the same structural workforce, but maybe it's a Band-Aid that Google's not willing to rip off, but Zuck is willing to rip it off at Meta.

23:21John Coogan:That's maybe the thesis at a deeper level. Anthropic has been the most aggressive lab by far when it comes to buying coding data from RL environment startups. And it's one important reason why their models are so good at coding today. Mercor recently disclosed that they logged 2.5 million expert hours on their platform in Q2 of 2026. That's equivalent to 5 ,000 people working 40 hours a week. Meta is already in the same ballpark and their average quality is likely higher. Additionally, they have another 70 ,000 people to pull into that organization if they need to, if it winds up being as valuable as semi-analysis thinks it will be.

23:55John Coogan:It's also worth briefly dispelling this myth that 3 ,000 meta-engineers will be doing mindless low-level data labeling. The days of undereducated contractors from third-world countries drawing boundary boxes or classifying texts as NSFW are gone. At this point, the models are sufficiently smart that creating a good piece of training data is a real intellectual challenge. So they're sort of taking the contrarian stance that being an RL environment designer is not soul-crushing and, in fact, can be intellectually rigorous. They also say that the top expert contractors at every data company, they're all making seven figures a year, which is pretty crazy since people thought of data labeling as a$10 an hour task.