Speaker 1Just how close are we to AI being able to invent worlds on the fly? And as it gets closer, what are the implications for games and video and all sorts of different types of experiences that we have today? Those questions are back at the top of the discourse today as Google releases Genie 4, its latest world model and one that shows that where we are headed might be very different indeed than where we are right now. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. All right, friends, quick announcements before we dive in. First of all, thank you to today's sponsors, KPMG, CDATA, Granola, and Blitzy. To get an ad-free version of the show, go to patreon.com slash AI Daily Brief, or you can subscribe on Apple Podcasts. To learn more about sponsoring the show, send us a note at sponsors at ai-dailybrief.ai. Lastly today, if you haven't yet, please go check out How We Got Here. It is a new weekly video podcast that I am doing that is all about the history behind today's podcast. Biggest phenomenon, trends, anything that's happening in our world that we assume is inevitable, but which actually came to be through a set of very specific decisions. For those of you who like the more ponderous exploratory episodes of AIDB, that is a place to get even more of that. You can find a link in the show notes or go to howwegothere.com. With that out of the way, though, let's talk about AI creating worlds. Welcome back to the AI Daily Brief Headlines Edition, all the daily AI news you need in around five minutes. Anthropic is doubling down on biotech with a big one. A big acquisition of a very small company. On Tuesday, Anthropic published a blog post talking through their approach to AI-enhanced drug discovery. They wrote, Developing a drug involves many varied tasks. Doing early research to identify a target for a drug and the right type of drug to engage it. Designing a molecule with the right properties. Optimizing it to be manufactured reliably at scale. Drafting the regulatory filings needed to run clinical trials. Designing and administering those trials and much more. Our strategy to realize AI's medical benefits is based on having AI increasingly involved in all of these tasks. There is no single problem to solve. The bottlenecks are distributed all throughout the drug development process. The best way to identify and work through these bottlenecks is to get deeply involved in real drug programs across many different disease areas and stages of development. To that end, Anthropic said that they've been partnering with six organizations as well as launching drug discovery efforts of their own. They included several case studies showing how Claude had accelerated both the scientific and operational work required to develop promising drug candidates. In addition, they announced a new partnership with Bristol-Myers Squibb to use AI to address bottlenecks and enhance planning and early stage clinical trials. But the news that caught everyone's attention was another biotech acquisition. They wrote, "We've acquired Mithril Medicine, a one-person company that has been developing treatments for two rare forms of muscular dystrophy. Its founder, Cody Tranbarger, used Claude to organize contracts, logistics, and experiments with 11 external labs and delegate many other aspects of Mithril's operations to Claude. As a result, Mithril took a possible treatment for one form of muscular dystrophy, from an idea to promising results in mice, in less than 18 months. Anthropic hopes to have a drug ready for clinical trials by the end of 2028, and this would be a significant acceleration of the typical 4-7 year timeframe for preparing a new drug for human testing, largely due to a big reduction in the drug discovery phase. The thing that stood out most to many folks was the feasibility of a single-person biotech startup. News aggregator Andrew Curran commented, "Sam Altman's one-person company prediction is coming true. Anthropic just bought one. The founder of Mithril Medicine used Claude to take a muscular dystrophy antibody treatment from concept to proof of principle in mice in 18 months, with just himself, Claude, and part-time consultants." Yu Chenjin wrote, "The real one-person billion-dollar company, a solo, smart, high-agency CEO and AI can have an outsized impact on the world." Then again, not everyone has bought in yet. Activist/short seller Scorpion Capital wrote, "Founder of Mithril Medicine one-person drug company acquired by Anthropic is not a scientist or academic as we assume. But a Harvard MBA, ex-Goldman Sachs banking analyst and ex-hedge fund guy." Now, some did note that the founder has a master's in biotech, meaning they are not completely foreign to the science, but mostly this is a Rorschach test for whether you see the death of expertise or a demonstration of democratization that could lead to more innovation. Others questioned the entire premise of the one-person AI company. Peter Yang commented, "Some of these one-person companies look highly suspicious to me, to be honest." Lenny Rychitzky from Lenny's podcast agreed, adding, "I'm very skeptical of a one-person company reaching massive scale. Just the support time alone makes it impossible to handle, unless you don't count contractors helping you with support. Don't think AI will ever solve this for a one-person company." Elsewhere on X, people viewed this as a very shrewd marketing move from Anthropic ahead of their IPO. Fundamentally, this was a restatement of their thesis that AI can massively accelerate drug discovery and make research into low-frequency diseases more financially viable. Still, Anthropic hammered home that AI use in biotech isn't about full automation. It's about reducing friction and bottlenecks throughout the process. Commenting on the meta-strategy of promoting the IPO, Chubby on X wrote, "Drug discovery will be one of the most significant revenue streams in the future. The goal here is to set the right course early on. Most diseases will become treatable, a market encompassing 7 billion people. The timing is clearly intentional, announced just before the IPO, to signal to investors that the business will continue to thrive in the future. This will likely make it clear, even to those outside the tech nerd sphere, why AI is so important. It goes beyond just writing code. It solves problems for humanity." And picking up a theme that we've explored a bit in the past few weeks, for some, this is yet another example of an AI company diversifying their revenue streams outside of just selling tokens. Martin Varsavsky writes, "The tens of millions from the better molecule go to whoever owns the molecule. If AI labs want to be paid for value, they will have to own the outcome. Discover the drug, run the trial, hold the patent. That is why Isomorphic Labs, the drug discovery company alphabet spun out of DeepMind, matters more than any pricing page." Selling intelligence is becoming a commodity business. Owning what intelligence discovers is not. Boy, you can count on that to continue to be a debate, and one that gets nothing but more fierce when Anthropic IPOs. Speaking of which, the assessment that the market was generally positive on the Anthropic S1 seems to have been validated. The public reveal of Anthropic's IPO filing came after a three-week drawdown for tech stocks and a four-month bear market in semiconductors. So far this week, however, the AI trade has been nothing but green. The NASDAQ index is up 2% since last week. It was up 2% since trading opened on Monday. And the Philadelphia Semiconductor Index has rebounded by almost 6%. Certainly, the market has not assessed that Anthropic has no weaknesses in the business. But for investors who were expecting the filing would confirm no path to realizing a return on AI capex, the picture of a booming business with a realistic path forward was enough to get many of those investors back on sides. Hebby Chan from Vantage Global Prime said, "Anthropic's landmark IPO filing, alongside Meta's aggressive capex push, has reignited broader risk appetite, pulling capital out of its defensive crouch, and firmly restoring chip stocks back to the driver's seat." James Wee, the chief market strategist at Tiger Brokers, said it was all about Anthropic's cash pile, commenting, "With more than $112 billion in cash and cash equivalents as of June, even before any potential IPO proceeds, Anthropic appears to have considerable financial capacity to support its expansion." Meanwhile, the Wall Street Journal is reinforcing the point that I raised on Monday's show, that it's really in everyone in the AI industry's best interest for this IPO to go well. I argued that that could be a very good reason for Open AI to go down. Open AI to just sit back, mothball their next model release, and let Anthropic have the limelight for just a couple of weeks with Fable 5-5. The Journal took it a step further, pointing out that the IPO will say a lot about access to capital for the entire industry. If Anthropic can raise $100 billion with room to spare, which would be the largest IPO in history, then it bodes well for smaller companies with more pressing fundraising needs. In a separate article, the Journal grappled with how Wall Street will price existential risk. "The logic is simple," writes the Journal, "if there is a 10% chance that artificial intelligence could kill all humans in the next decade, as an Anthropic researcher said last month, there's a 90% chance it's really, really good, but designers managed to rein in its genocidal instincts. An AI smart enough to kill us all, but which prefers to help us, is supremely useful, and should generate healthy gains for its inventors." Now, I don't know exactly how firmly author James McIntosh's tongue is in his cheek, but it does appear that this week, at least, it looks like Wall Street has stopped focusing on P-Doom and is starting to price in P-Boom. Lastly, in a brief follow-up to yesterday's story about new scientific breakthroughs from OpenAI, a group of scientists are still unhappy with how OpenAI went about their disclosure. During the giant dump of novel math proofs earlier this month, the criticism was that OpenAI didn't do enough to verify results. This time around, OpenAI tried to address this criticism by appointing groups of scientific experts to shepherd the breakthrough results. These experts recommended a different approach for each result. The math proof was just released publicly as an academic paper, the computer science result disproving the exponential time hypothesis, had proofs in AI sessions released alongside, with a team of scientists appointed to do further work explaining the result and exploring its implications. However, the physics result, which was said to include an existence proof for Yang-Mills theory, was kept private while a group of physicists worked on verifying and formalizing the result. Following OpenAI's announcement, their own advisory group on mathematics and artificial intelligence publicly disagreed with this approach. In a public recommendation, the mathematics advisory group said they had great respect for the physicists appointed to shepherd the results, but that, quote, "this understanding should develop in the open, with everyone given the same access." Later on Tuesday, the group of physicists in question released a response, commenting, "The length and complexity of the OpenAI manuscripts are formidable, and some time will be needed to digest its methodology and meaning in both physics and mathematics. Before we release it, we would like to be convinced that it is correct. This is an experiment in how the academic community can deal with the rapid development of new knowledge by AI. We intend to proceed very carefully." Still, the frustration with the delayed publication seems to be playing out faculty lounges across the country. The Verge spoke with several physics professors, with many of them calling for an immediate release. Constantine Kogler from the Institute for Advanced Studies said, I'm fine with them releasing the problems as they did today, referring to releasing a smaller batch of three extremely important results. However, Kogler added, it doesn't make any sense that only a few distinguished scientists can look at them. Andreas Kark from the University of Texas Austin said, the fact that they handpicked the group is a disaster. The large labs have enough power as it is. They should not be picking winners and losers. Kark initially suggested a merit-based selection criteria, although he acknowledged that this would likely produce the same group that included Nobel laureates. Instead, he simply called on Open AI to release the proof, saying, this would democratize access while still showing that they care about making results meaningful and understandable. I don't know, man, I think on this one, Open AI simply can't win. If they release everything they have as soon as possible, they're criticized for creating a gigantic verification problem for academics to deal with. If they arrange their own verification working groups and hold back important results until they're fine, they're criticized for creating a gigantic verification problem for academics to The thing that's totally understandable is that people are uncomfortable with how much power these labs have. And that is not solely based on specific decisions that the labs make. It is based on the intractable reality of where they sit vis-a-vis this new intelligence. But when it comes to this sort of specific, they're pretty much damned if they do and damned if they don't. Stanford mathematician Jared Duker-Lichtman commented, my sense is that for any approach, there is going to be someone who is upset. So it's just a question of finding the least offensive option. I am glad that is not my job. And for now, that's going to do it for the headlines. Next up, the main episode. A new study from KPMG and the University of Texas at Austin found that when people work with AI, similar skills don't guarantee similar outcomes. Researchers studied more than 500 early career professionals and found that the best performers consistently amplified the value of AI by guiding, evaluating, and refining its outputs. These top performers, called AI amplifiers, weren't defined by what they knew alone, but by how they worked with AI. Learn more about what separates AI amplifiers from everyone else at kpmg.com slash US slash AI amplifiers. This episode is brought to you by CDATA. Companies rolling out AI keep hitting the same wall. Token spend is climbing, security is holding agents back, and no one knows how AI is truly being used. The root cause? AI is cut off from the systems and context your business runs on. That's why CDATA built Connect AI Gateway, one control point between your AI and your systems. Built on the connectivity layer powering Palantir, Google, Salesforce, and hundreds more, each agent gets exactly the data and tools it needs. It governs down to the record and lets you finally see your org's complete AI usage in one place. Accurate answers, a fraction of the token cost, and a gateway that learns your business with every interaction. See how at cdata.ai. You're in Back-to-Back, meetings all day. You're trying to stay present, but you're also worried you'll forget the decision, the action item, the important next step. This is a scenario that we absolutely live here at the AI Daily Brief, given how fast things move and change and how quickly we're evolving everything that we do. That's where Granola comes in. Granola is an AI-powered notepad for meetings. You jot down rough notes like you always do, and in the background, Granola transcribes and turns them into clear, useful notes when the meeting ends. No bots joining your calls, no distractions, just a clean notepad that helps you focus. During or after the call, you can chat with your notes, ask Granola to pull out action items, help you negotiate, write a follow-up email, or even coach you using recipes which are pre-made prompts. Once you try it on a first meeting, it is very hard to go without. Head to granola.ai slash ai-daily and use code ai-daily to get 100% off for three months for new users. That's granola.ai slash ai-daily and code ai-daily to get 100% off for new users for three months. Here's why most legacy modernization projects fail. The AI doing the work can't understand code bases at scale. It sees a small slice of context, examines syntax, and misses years of decisions distributed across the global application ecosystem. Blitzy solves this the way it solves everything. Grounded in your code before any migration begins, Blitzy's agents reverse-engineer the entire legacy system into a persistent knowledge graph. Every dependency, every constraint, every piece of tribal knowledge that used to live in one engineer's head. From that understanding, Blitzy autonomously executes language migrations, framework upgrades, and monolith-to-microservices transformations, all validated end-to-end. One Blitzy customer modernized a $10 million monolithic insurance stack in 16 weeks against a 137-week baseline with coding agents. That's 9x compression. Retire technical debt while accelerating your roadmap. See how at Blitzy.com. That's B-L-I-T-Z-Y.com. Welcome back to the AI Daily Brief. Quick note before we dive into today's episode, there is going to be a lot of visual content. As always, I will do my best to make it even if you are just listening, not watching, but this might be a good one to go check out either on Spotify with video or on YouTube. Apple continues to roll out video podcasting, but has not yet made it available for my particular host. Hopefully that will come soon. Today we are talking about Genie 4. It's the latest iteration of Google's world model, and it could have some big implications for robotics, gaming, and more. On the AI Daily Brief, we've been tracking the development of world models since they first began showing up in 2024. One of the reasons for that was because of the development of world models. AI luminaries like Yann LeCun and Fei-Fei Li believed that world models were a more viable path to AGI than large language models. The argument was basically that you had to have a different physical understanding of the world around to achieve AGI, not just be able to read a bunch of stuff. Obviously I'm being wildly overreducted, but you get the idea. The early experiments in world models were breathtaking tech demos. They showed us something completely different to anything we'd seen AI do before. At the time, generative video was still in its infancy, so the AI was still in its infancy. So the AI was still in its infancy. So the ability to create fully 3D worlds at all was a massive breakthrough. And yet even then it was clear that these demos really weren't about what was available now, but what was to come. Over the next year or so, Google made rapid progress with their Genie models. Genie 2, which was first revealed in December 2024, was very impressive. But when Genie 3 was previewed in January 2025, it was clear that Google had made a huge leap. The model was eventually served to the public in August of 2025 and allowed users to use the keyboard and a mouse for the first time. One of the most impressive technical breakthroughs was a level of persistence that we hadn't seen before. Users could change something about the world, like paint a symbol on a wall, look away, and then look back, and the symbol would be the same. At the time, generative video models were struggling mightily to achieve that level of consistency. Google said that this was an emergent property of how the models worked rather than a frame of video auto-regressively based on what had come before, so the model kept a record of the world as it was being generated. And yet still, we were very clearly in the tech demo stage of world models. Incredibly cool tech demos, but far, far too limited for practical use. Technologists suggested world models could help train robots or generate fully playable games in real time, but Genie 3 was limited to 60 seconds of generation because of the high compute requirements. Now, alongside all of this, Google has, of course, in the eyes of many, been languishing without a state-of-the-art language model to compete in the main areas of AI competition like agenda coding. And yet, for most people who were paying close attention, there was never any question of Google's proficiency in advanced multimodal AI. Now, Google has revealed Genie 4, and with it, a pathway to making world models far more practical and useful. With this iteration, DeepMind has made a huge change to the model's architecture. Rather than being a standalone model, Genie 4 integrates other Google models to provide things that Genie can't do by itself. Gemini Omni, Google's video model, quote, anchors the simulations in real-world appearance and behavior. Presumably, that means something like Omni creates a video clip for reference, which Genie then expands out into a fully playable world. In addition, Gemini 3.8 Flash provides a reasoning engine for physical interactions, giving the model reliable causality. On top of that, some unspecified combination of Google's audio models brings sound to generated worlds for the first time. The result is Google's entire AI ecosystem being leveraged towards creating the most realistic world simulations we've seen to date. In their announcement post, Google highlighted four key features that they viewed as the big advanced advances for Genie 4. First, users can now deliver prompts to guide the simulation in real-time. Previous iterations of Genie only allowed users to create the world with a prompt up front, then navigate whatever was generated as a stable world. Now, users can give live prompts to guide the camera or to modify the environment by changing the weather or introducing a new object or character. Second, Google highlighted the huge jump in realism since Genie 3. They wrote, built on an advanced physical understanding of the world, Genie 4 simulates real-world physics scenarios such as aerodynamics collisions and more. The most dramatic leap is how Genie 4 simulates people, creating characters that move naturally, respond organically to their environments, and can be interactively steered in real-time. The third highlight was generative audio, which, as I mentioned, we haven't seen in world models before. Now, Google's Omni video model featured impressive sound, as did their Lyria music model, so it's not all that surprising to find amazing audio in the new version of Genie, but the fact that the worlds can not only include music or human speech but also sound effects gives creators a whole different level of tooling than was available before. Check out, for example, the otherworldly sounds to match an abstract landscape, such as those that go alongside this stained glass wilderness scene that they used as an example. Now, the speech does feel a little behind the state of the art, still featuring that hollow and echoey style that was the hallmark of dialogue in generative video until about three months ago, such as this example of a couple talking while riding on a motorcycle. Finally, Google highlighted what adding a reasoning model in the background brings to the table. Genie can now simulate complex, multi-step actions with a much higher level of realism than we've seen before. Explaining how this works, Google wrote, To achieve this in real-time, Gemini 3.8 Flash serves as a guiding agent in the loop. As Genie 4 generates a world in real-time, Gemini continuously observes the scene, breaks down goals into a sequence of steps, and reviews incoming frames to track the progress and verify that each step succeeded. As the world shifts or the user intervenes, Gemini adapts its plan in the moment, unlocking worlds with reliable causality, procedural behaviors, and multi-step reasoning. The result is undeniably a huge step up in world models. We now have the realism of modern video models combined with the reasoning capabilities of modern LLMs baked into a fully playable world simulator. One immediate implication is a new paradigm in robotics. Google has been experimenting with using Gemini 4 Argon rather than Gemini 3.8 Flash as the backend reasoning model for cutting-edge robotics research, showing a demonstration of how the world works in real-time. Google has been experimenting with using Gemini 4 Argon as the backend reasoning model for cutting-edge robotics research, showing a demonstration of how the world works in real-time. Google has been experimenting with using Gemini 4 Argon rather than a robot playing chess inside a Genie simulation. They wrote, The general response from the community was jaws hit and floors. And I think it's worth noting since I've critiqued before the fact that everyone has an incentive to focus on 3D renderings and things like that because they're popular social media shares when a model first gets released. That's a little bit different in this case because that is the entirety of what this model does. In other words, it's not a model that you're going to use for agent decoding or writing essays that people are judging on the basis of its ability to make 3D renders. It's not a model that you're going to use for agent decoding or writing essays that people are judging on the basis of its ability to make 3D renders. and improve their spatio-temporal reasoning, similar to how humans can imagine a real-world scenario and think through what might happen before doing something. This gives agents a place to compare actions and learn from simulated outcomes before trying them in the real world, assuming these simulations are accurate enough. I think this could help with some of the spatial and physical reasoning problems models still struggle with. Humans can mentally rehearse what they're about to do, but LLMs are only able to do text-based reasoning and can describe physics and math possibilities in text, but an interactive world model gives AI a way to play them out. The physics and long-term consistency still need work, obviously, but this is a great step towards giving AI richer environments to interact with and learn from. Now, as you might imagine, one place that people are seeing very clear implications is for the world of gaming. In fact, since the earliest days of world models, most have assumed that that was where this technology would have its first big impacts. The idea is that one day, perhaps one day soon, gamers will be able to prompt their own personalized game using a world model and create a fully tailored generative gaming experience. In fact, the release of Genie 3 sent gaming stocks plummeting, although this Genie 4 release was less damaging to the sector. It seems investors have gotten a little bit more sophisticated in their understanding that the technology is still a little way off for being able to disrupt the gaming industry. Still, there's already a huge AI transition underway in that industry, driven by Opus 5.5's ability to decompile and repackage games. The first few weeks of that saw weird mashups, like dropping Mario into Dark Souls. Then there was the wave of ports, with people realizing Opus could make basically any game compatible with pretty much any hardware, after a tech report reported Tim Sweeney's first game from PC to the NES. The Epic Games CEO commented, "This is amazing and shows what would have taken a vast manual effort in the past can be done as a hobby project very efficiently. Truly a new age of gaming and software ahead." Now people are bringing some truly zany ideas to life. Lucas Martinich managed to turn Mario 64 into a mixed reality game, the gameplay is presented on a virtual screen inside of VR glasses, and the player can reach in and interact with Mario as he's running around, for example using their hand as an additional platform. The player can also grab Mario pull him out of the game and place him on their kitchen table, turning their real-world space into another world for Mario to run around in. In short, while the tech might not be ready for fully generative games, AI users are figuring out how to do some strange and incredible things with the technology we already have, which will inevitably shape what comes next. As with Genie 3, Google is making Genie 4 fully available for Gemini Ultra subscribers to try out. It's still only capable of 2 minute simulations, which is double the limits of Genie 3, but still short of a fully generative gaming experience. So, for now, this is just the beginning, and I hope you enjoyed this video. Thanks for watching, and I'll see you in the next video! If the AI frontier is jagged, DeepMind's output is perhaps the most jagged of all. They haven't been able to fully catch up on LLMs, but their progress on multimodal is still unrivaled. With Genie 4, Google is demonstrating that combining a series of disparate model architectures can produce entirely novel AI systems with huge amounts of potential. What's more, as an AI business, Google also continues to have some diverging bets that are paying off. When Google started ramping up TPU production last year and leaked discussions about selling the chips externally, Google was able to catch up on some diverging bets. Similarly, many semiconductor analysts scoffed at the idea, NVIDIA had a stranglehold on AI chips, and Google was starting from zero on sales and business development. Still, TPU sales have grown to become a significant line item in the revenue statement, albeit almost entirely through sales to Anthropic. This week, however, the information reports that Google has won two new customers, in OpenAI and Oracle. Terms weren't disclosed for the OpenAI deal, so it's difficult to know how many TPUs will be purchased and how they will be deployed, but Oracle is set to become the first large cloud provider, aside from Google, to rent GPUs. And in addition to their hardware sales, the other big thing to watch for Google in the coming months will be whether they can rapidly iterate on their LLMs. Even before Gemini 4 Argon was released, Google already had their follow-up model ready for internal testing. The model is known by the codename Gemini 4 Carbon, and it seems like it's getting positive reviews from the Googlers who've tried it. Critically, Gemini 4 Argon isn't just a new commercial model. It's also a new ultra-large base model that Google can train on top of. They don't need to compete with Google for the long pre-training run. They can apply relatively fast post-training on top of Argon. We've seen what that meant for Meta with the release of Muse Spark. The first version was pretty dismal, but they released three additional versions in five months and bootstrapped their way to a competitive mid-sized model. Reinforcement learning runs on an ultra-large model won't be quite as fast, but it still gives Google a clear pathway to rapid model improvement. There's also the fact that the creators and listeners of this podcast may be a little too deep in AI to understand how much existing distribution matters over even model capability. A week ago, Google's Logan Kilpatrick posted, In just the first 72 hours since Gemini 4 Argon launched, we processed around 25 trillion tokens. The demand for intelligence is incredible and still underappreciated. At the end of the day, Genie 4 is still not out of the tech demo stage when it comes to world models. But you can feel us getting tantalizingly close to that inflection, where tech demo becomes real world application. And in the case of world models and Genie specifically, that could be a big change indeed. For now, that's all I have for you. Thanks for watching. That's going to do it for today's AI Daily Brief. Appreciate you listening or watching, as always. And until next time, peace.
PodcastTranscript AI