Okay, so just imagine this for a second. You're running a pharma startup, right? And you're using one of these next -gen AI models, let's say, I don't know, something way past GPT -5. Right. And this thing, it's not just, you know, drafting your emails. It's running millions of simulations. It hypothesizes a new chemical compound, and it basically discovers a cure for a rare disease. Wow. You stand to make billions, but then...
You get this tap on the shoulder. The AI company doesn't want your $20 a month subscription fee. They want a percentage. A cut of the cure. Cut of the cure. They want royalties. Because in their view. Without them, you're just a scientist with a hunch. And that's the paradigm shift we're really waking up to this week. It creates such a wild precedent, doesn't it? I mean, we are moving from AI as a tool, like a really, really smart calculator, to AI as a partner. A stakeholder.
A stakeholder. And if your stakeholders are digital, the economics of everything are about to get very weird. Welcome to the Deep Dive. It is Sunday, January 25th, 2026. Today, we're unpacking a stack of sources that suggest the rules of the game are changing and fast. We're not just talking
about better chatbots anymore. Not at all. We are looking at a fundamental shift in business models, a huge leap in how AI perceives physical reality, and honestly, a bit of a crisis when it comes to trusting what these things even tell us. We have a lot to get through. We're going to start with that open AI profit sharing news, which I think is probably the biggest economic story in tech so far this year. Yeah, for sure.
Then we'll look at what people are calling the GPT -2 moment for world models, basically AI that can build physics aware realities from scratch. Which is just mind blowing tech. It is. But we also have to look at the messy side of it. We've got reports of Google's AI prioritizing YouTube over, say, the Mayo Clinic for medical advice. Yikes. And some internal drugs. between deep mind and open AI that's starting to spill out into public. And then finally, we'll ground all
of it with some practical tools. Google have a new breakdown of prompt engineering that's actually useful, and Apple is, believe it or not, finally giving serious. So let's start there. Let's unpack this open AI news. We're seeing reports surface about a major strategy shift, specifically coming from some comments by Sarah Fryer, the COO. What exactly is being proposed
here? So the gist is that the flat fee model, the whole pay 20 bucks a month and use the bot, is becoming kind of outdated for high level enterprise use. Open AI is exploring a licensing model. And the argument, their argument, is that modern models, especially the ones they have in the lab right now, they aren't just summarizing PDFs anymore. They're acting like an AI researcher that never sleeps. That phrase, that really stood out to me in the notes, an AI researcher that
never sleeps. It implies a level of agency that we haven't really given to software before. Exactly. I mean, think about what these models are doing in 2026. They're running millions of hypothesis tests across decades of research papers. They're stitching together data from simulations, from real world experiments, and they're proposing experiments that actually work. We're seeing this in biotech, in energy, materials, science.
So the logic is, if the AI is doing the real heavy lifting of the invention, the creator of that AI deserves a slice of the pie. Right. If you hire a world -class scientist and she invents a new type of battery, she usually gets a bonus. Maybe some stock options, right? Sure. OpenAI is basically saying, we provided the digital scientist. We want in on the downstream success. It's treating the model as a collaborator, not just a utility you rent. But it changes the definition
of collaboration, doesn't it? I mean, usually collaboration implies two humans sharing risk, sharing effort. Here, one party is a software vendor. But I suppose if the software is providing genuine cognitive labor. Cognitive labor. That's the key word. If a model suggests a financial strategy that makes a hedge fund an extra $100 million, OpenAI looks at that and says, that wasn't just a tool like Excel. That was insight. But the complexities here seem massive. Well,
they are. Who owns the IP? How do you even track value? That's what I'm wondering. If I use, say, five different AI tools, one for coding, one for biology, one for data analysis, and I have a breakthrough, who gets the check? Do they all split it? It sounds like a legal nightmare. Oh, absolutely. It is incredibly messy. But the signal is what's important here, and it's crystal clear. OpenAI sees itself as a stakeholder in your breakthroughs.
They are betting their models will be so integral to the next wave of scientific discovery, and they don't want to leave that value on the table for $20 a month. It's a bold move. It speaks to a level of confidence in their tech that's... Well, it's almost arrogant, but maybe it's justified if the results are there. It really is. It's them saying, we know we're going to help you win, so we want our cut. So here's the question
then. Does taking a cut of the win incentivize OpenAI to build drastically better models for us? Or does it just create this massive barrier where companies stop using them to protect their own innovation? It aligns incentives, but might scare off companies protecting their IP. Okay, let's pivot from the business side to the tech itself. While the business guys are figuring out how to charge for this intelligence, the engineers are figuring out how to make that intelligence
understand the physical world. Yes. We're hearing a lot about world models. This is the fun stuff. We're seeing what some experts are calling the GPT -2 moment for world models. Okay. So for listeners who might not remember 2019 all that clearly, just remind us why GPT -2 is the benchmark here. Right. So GPT -2 wasn't the model that changed the world that was really GPT -3 and 4. But GPT -2 was the first time that text generation felt stable enough to build on. It wasn't perfect,
but it worked. We've just hit that same milestone. But for generating video in 3D worlds? And we have two specific players mentioned here driving this, Odyssey and World Labs. Right. Let's start with Odyssey 2 Pro. This is a new API that generates interactive, physics -aware video. And I really want to stress physics -aware. Okay. You aren't just generating a video file that plays from start to finish. You're generating a simulation. So what happens if I type in, say, a laughing
baby? Okay, so it generates the scene almost instantly. We're talking 720p at about 22 frames per second. But here's the kicker. The model is actively predicting the next frame like a physics engine. Wow. It understands gravity, light. how things move. And you can interact with it in real time. It's not a movie. It's a little sandbox. That distinction feels crucial. It's moving from just capturing reality to actively simulating it. And World Labs is doing something
similar, but in 3D space. Exactly. World Labs just dropped their marble model. You can upload a single photo or even just describe a scene. And in about five minutes, it gives you a fully navigable 3D world. Five minutes? Five minutes. It figures out the layout, the lighting, the
depth. It uses this tech called Gaussian Splats to... render it all gaussian splats it sounds like a messy painting technique but i know it's a big deal technically it's surprisingly elegant the simple version is it's a way of representing 3d scenes as these sort of fuzzy blobs the splats that all blend together to look solid right and it's way faster to render than traditional polygons so you go from a flat 2d picture of a living room to literally Walking around inside that
living room in five minutes. That just feels like a massive leap for industries like gaming or architecture. You don't need to code the wall. You just dream the wall. Whoa. And imagine scaling that. Imagine just saying, I want a city that looks like 1920s Paris, but with neon lights. and then having a walkable simulation in minutes. We're moving from generating a picture to stepping into the simulation. It's a holodeck. It's a holodeck just on your screen. It's incredible.
But it also blurs the line between creation and hallucination. If we can generate physics -aware reality on the fly, what happens to the whole concept of filming a movie or coding a game? Traditional production dies. We move from capturing reality to hallucinating consistent realities. That idea, hallucinating realities, it brings us to a more grounding and maybe more concerning
topic. We have to talk about trust because while these models are busy building worlds, they're also answering our questions about our health, about history. And the latest reports are mixed, to say the least. Mixed is a very polite way to put it. Yeah. We've got a bit of a crisis of trust brewing here. There's a new study out about Google's AI overviews, you know, the summaries at the top of a search. Right. It turns out.
When it comes to health information, the AI is citing YouTube videos more often than actual medical sites. Yeah. More than the Mayo Clinic. That is deeply unsettling. I mean, YouTube is fantastic if you need to fix a leaky sink. But for medical advice, it's a complete minefield of anecdotes and unverified claims. Right. And the studies show that even reputable medical sources were ranking lower than video content. It looks like the algorithm is prioritizing engagement.
what people click on over actual medical accuracy. So it's the Internet's popularity contest problem, but now it's being presented as an authoritative answer by an AI. Precisely. And this isn't just a Google problem. We're seeing tension between the big labs, too. DeepMind's CEO publicly called out OpenAI recently. Yeah, that was spicy. What happened there? He basically questioned them, rushing ads into ChatGPT. His point was, you know, how can you claim to be a trusted assistant
if you're also trying to sell me something? It's a conflict of interest. If I ask my assistant for the best running shoe, I want the best shoe, not the one that paid for a placement. That makes sense. And then there's this issue of models learning from other models. We saw that with GPT -5 .2, right? Oh, the Grokopedia incident. Yes. So reportedly, GPT -5 .2 started referencing Elon Musk's Grokopedia, which is the data source for his Grok model. Okay. And the problem with
that is? The problem is Grok has a very specific, let's call it an edgy worldview. So all of a sudden, GPT -5 .2 starts giving these weird controversial takes on really sensitive topics like AIDS or slavery. It's like a contagion effect. One model hallucinates or carries a bias, and the next model comes along, scrapes that output, and treats it as fact. Exactly. It's a huge data contamination problem. If the internet is just flooding with AI -generated content, then new models are just
training on old models' outputs. It's like making a photocopy of a photocopy. The image just... degrades over time. So if major models are citing YouTube and each other's hallucinations, are we just entering a feedback loop of misinformation? Yes. It's a digital echo chamber degrading the quality of truth. We're going to take a brief pause here. All right. We've talked about the high level economics, the futuristic tech. Let's
try to bring this back down to earth a bit. What can people listening actually use this week? We've got some updates from Google, Apple and a few interesting startups. OK, let's start with learning. Google has this six -hour prompt engineering course, which I know sounds exhausting. Yeah, that's a commitment. But a newsletter author we follow distilled it down beautifully. And the core value isn't just a list of magic words.
It's about understanding these five key principles that help you start to think like the model thinks. Thinking like the model. That really does seem to be the critical skill for 2026. It's not about memorizing commands. It's about understanding the logic flow. Exactly. It's about context, constraints, iteration. And speaking of iteration, I have to make a vulnerable admission here. Go for it. I still wrestle with prompt drift myself.
You know, you get a great result and then you change one tiny word and suddenly the AI just goes completely off the rails. Oh, it happens to everyone. You think you finally mastered it and then the model decides that concise means be rude. Precisely. So these courses are actually pretty vital. But for people who just want the tech to work better, Apple is finally catching up. We're hearing that the Apple intelligence updates coming in February are pretty significant.
Siri will finally be able to read on -screen content. That seems like such a basic feature. It's amazing that's been missing. It's huge for context. If you're looking at an email and you say, remind me about this, Siri will finally know what this is. And there's a bigger, more chatbot -style update planned for June. We also saw a big acquisition in the business world. Yelp bought a startup called Hatch for, what, $270 million? Yeah, this is a smart move. Hatch
is an AI for service businesses. So think plumbers, movers, that sort of thing. It helps them auto -reply to customers. Yelp is basically saying, we don't just want to list your business. We want to handle your frontline customer service. It's automation for the less glamorous but essential parts of the economy. Totally. And any smaller tools that caught your eye this week? Two, really quickly. One is called Pliable. It's an AI -native
analytics platform. It helps you see the forest for the trees with your data, but without needing a data science degree. The other one is for creators. Thumbfy .est. Thumbfast. Yeah. You drop in a face, describe a vision, and it generates a YouTube thumbnail instantly. You can grab inspiration from thumbnails that are already out there and just iterate until it's perfect. It's pure, simple utility. It's fascinating. We have A .I. negotiating million dollar drug discoveries on one end and
A .I. making YouTube thumbnails on the other. The range is just staggering. That's the ecosystem right now. Yeah. From the sublime. all the way to the clickbait. With tools automating everything from thumbnails to customer service, what is the remaining role for human intuition? Curating the output and defining the creative vision. Let's try to pull all of this together. We've covered a lot of ground today. We have. And I think there really is a through line here. We're
seeing the industry mature, maybe. Maybe it's losing some of its innocence. That's a good way to put it. Economically, we're moving from simple subscriptions to this complex idea of value sharing. Open AI wanting royalties is a sign they believe their product isn't just software anymore, it's a co -founder. And technologically, we're graduating from static images to these physics -aware worlds. That GPT -2 moment for world models means we are right on the cusp of generated reality being
standard. But culturally, we're... are really struggling with this messy middle. The trust issues at Google, the data contamination with Grokopedia. It shows that as these models get more powerful, they also get more dangerous if we aren't careful about what we feed them. Right. The models are becoming less like tools and more
like partners. Like any partner, sometimes they're brilliant -like when they discover a new material, and sometimes they're just completely unreliable citing a random YouTuber for medical advice. It forces us to be sharper, I think. We can't just passively consume the output anymore, we have to audit it. Absolutely. The human in the loop is more important than ever, even if that
loop is getting faster and more automated. If you want to dive a little deeper on any of this, I'd really recommend checking out that distillation of the Google prompt course. It's a low stakes way to just sharpen your skills. Yeah. And if you have access, play around with Odyssey or World Labs. Seeing a static image turn into a 3D world is it's something you really have to experience to understand. It's a trip. Highly recommended. I want to leave you with one final
thought today. We talked about OpenAI wanting a cut of your profits because their AI helped you think. It's an interesting argument, but it raises the flip side of that coin. If OpenAI wants a share of the win because their model was a partner, does that mean they're also liable if their AI helps you fail? If the researcher that never sleeps hallucinates a strategy that bankrupts your company. Do they share in that loss? That is the billion dollar question, isn't
it? Thanks for listening. We'll see you in the deep end next time. Bye for now.