0:00 On April fourth, Google released a new model that runs with the speed of a small model but the intelligence of one six times its size.0:10 The model, Gemma 4 26B, is now live on Cloudflare's Workers AI platform, and it fundamentally changes the economics of deploying high-end artificial intelligence.0:22 This isn't an incremental update.0:25 It's a structural break in how we balance cost and capability.0:29 Elsewhere, the money is following similar shifts in the application layer.0:35 Daydream, an agency built to handle SEO for the new age of generative search, just secured fifteen million dollars in a Series A led by WndrCo.0:46 Their stated goal is to build agents that can think and execute like a senior SEO practitioner, because, in their words, most companies are optimizing for a version of search that no longer exists.1:01 In the physical world, Anvil Robotics raised six-point-five million in seed funding to build a modular, plug-and-play platform for robotics.1:11 The idea is to create a common infrastructure—hardware, controls, data pipelines—to stop robotics teams from constantly reinventing the wheel.1:22 It's an attempt to do for physical AI what cloud computing did for software.1:28 And finally, a new platform called GetDocs launched to bring AI-driven document recognition and workflow automation specifically to law firms, another sign of AI moving from a general tool to a vertical-specific solution.1:45 Let's go back to that new Google model, Gemma 4.1:48 The specification sheet says it has twenty-six billion parameters, which would normally mean it's powerful but slow and expensive to run.1:59 The catch is its architecture.2:01 It's a Mixture-of-Experts model, or MoE.2:04 Think of it as a team of specialists.2:07 When a query comes in, the model only activates the most relevant experts—in this case, about four billion parameters' worth.2:16 So you get the quality that comes from a large, twenty-six-billion-parameter knowledge base, but the performance and cost of a much smaller four-billion-parameter model.2:29 This breaks the traditional trade-off.2:32 For years, developers had to choose: cheap and fast, or smart and slow.2:38 MoE architecture, when done right, lets you have both.2:42 Add in a quarter-million token context window—enough to analyze entire books at once—and you have a model that isn't just better, but enables entirely new, more complex applications at a practical cost.2:57 Now, look at that Daydream funding.3:00 Fifteen million dollars for an SEO agency sounds high, until you realize they aren't selling traditional SEO.3:08 Their entire premise is that generative AI hasn't just given SEO new tools; it has invalidated the entire playbook.3:17 When search engines provide direct answers instead of just a list of links, optimizing for clicks on those links becomes an obsolete strategy.3:28 Daydream is building AI agents, paired with human strategists, to operate in this new environment.3:35 This isn't about using AI to write blog posts faster.3:39 It's about using AI to redesign the entire marketing function from first principles.3:45 The signal here is that the value is moving.3:49 As foundational models like Gemma 4 become powerful and commoditized utilities, the real opportunity shifts up the stack.3:58 It moves to the companies that can apply that raw intelligence to a specific, high-value problem—whether it's navigating the new world of search, or finally creating a common language for robots to speak.
