0:00 OpenAI announced 'GPT Apps' on August sixth.0:02 This isn't just another app store; it's a declaration of how the next phase of AI is going to be built.0:09 Last week, in episode one-forty, we talked about America securing a fifty percent lead in the global AI race.0:16 Well, this week you're seeing exactly how they plan to cement that lead—not with a bigger model, but with a different rulebook entirely.0:24 The announcement itself was simple.0:27 A new open standard called Agent Plugins, developed with partners like AWS, GitHub, and Vercel.0:33 The goal?0:33 To let developers, quote, “build a plugin once and use it across compatible agent clients.” Now, that sounds dry.0:40 It sounds technical.0:42 But what it means is a fundamental change in how you will interact with every piece of software in your life.0:49 The keyword here is portability.0:51 For the last few years, you've been training your AIs.0:54 You teach ChatGPT your writing style.0:56 You teach Midjourney your visual aesthetic.0:59 But that intelligence, that context, is trapped inside each application.1:04 It’s like having a brilliant assistant in one room who gets amnesia the second they step into the hallway.1:10 GPT Apps and this new standard are designed to fix that.1:14 An AI developer, Git Maxd, put it perfectly on X.1:17 He was at the OpenAI Dev Day for the announcement.1:20 He said, "This portability is exactly what excites me about Agent dev - just drop your MCP/Skill into supporting apps and bring your Context with you." Bring your context with you.1:32 That's the revolution.1:33 Imagine an AI agent that knows your entire project history from GitHub, understands your company's design system from Vercel, and can access your cloud infrastructure on AWS.1:44 All at once.1:45 Seamlessly.1:46 You don't have to re-explain the project every time you switch tools.1:50 The agent already knows.1:51 It carries the context.1:53 This is the end of starting from zero.1:55 So what we're seeing is the blueprint for a new kind of internet.1:59 A layer of intelligence that sits on top of all your applications, with agents that can move between them as easily as you switch browser tabs.2:09 And this idea… it’s been building for a while.2:11 This isn't just an OpenAI brainwave.2:14 If you scroll back the timeline, you see the pieces falling into place.2:18 Go back to December 2025.2:20 Anthropic—one of OpenAI's biggest rivals—published a standard for what they call "agent skills." A commentator named Kate flagged it at the time, noting their emphasis on portability.2:31 Her post read, "Like MCP, we believe skills should be portable across tools and platforms — the same skill." That acronym she mentioned, MCP, stands for Modular Control Protocol.2:43 It's the technical plumbing that allows these skills to be portable.2:47 So, eight months before OpenAI’s big announcement, their main competitor was already pushing the exact same idea.2:54 This isn't a feature war.2:56 This is a consensus forming around a new architecture for AI.3:00 And you can trace the roots back even further.3:03 An analyst named Murali recently mapped it out.3:06 He traces this entire wave back to late 2024 and early 2025.3:09 That's when the first true "reasoning models" appeared, like OpenAI's o1 and a model called DeepSeek R1.3:16 Before that, models were mostly about predicting the next word.3:20 They were brilliant parrots.3:22 These new models started to show glimmers of what’s called System 2 thinking.3:27 The ability to plan, to reason through multi-step problems.3:30 That was the seed.3:32 Without that leap in reasoning, the idea of an "agent" that can use "skills" across different "apps" is just a fantasy.3:39 You need a model that can actually understand a goal, break it down into steps, and figure out which tool to use for each step.3:47 First came the reasoning.3:49 Now comes the ecosystem to put that reasoning to work.3:52 So what does it all add up to?3:54 It adds up to the second half of the equation for AGI.3:58 Everyone has been obsessed with scale.4:00 Making models bigger and bigger.4:02 But as Demis Hassabis of Google DeepMind says—and this was highlighted by developer Steve Ike—AGI won't come from scaling alone.4:10 Hassabis believes it's fifty percent scaling, and fifty percent INNOVATION.4:15 Bigger models matter, but new ideas matter just as much.4:18 This is the new idea.4:20 The shift from monolithic models to modular, interconnected agent systems is the innovation half of the roadmap.4:27 It’s the architectural leap that might be just as important as adding another trillion parameters.4:33 We are exiting the era of the all-knowing oracle in a chat box, and entering the era of the tireless, specialized worker.4:41 An agent that doesn't just know things, but does things.4:44 This is the promise.4:46 And it's enormous.4:47 It's the transition from a desktop and cloud era to a truly AI-first world, as one VC, Deemah AlYahya, noted.4:54 But there's a turn here.4:55 A hard one.4:56 Because for all the excitement from developers and the elegant diagrams from architects, there’s a brutal question hanging over all of it.5:05 Shivum, an analyst who focuses on enterprise AI, laid it bare.5:09 He said, "This is effectively the number one problem for AI agents in the enterprise.5:14 2025 is the year of agents...5:16 but do they actually work in the real world?" It's the perfect reality check.5:21 It’s one thing to demo an agent booking a flight in a controlled environment.5:26 It is a completely different thing to unleash it on a company's messy, chaotic, real-world supply chain software or its customer service database.5:35 The gap between the demo and the deployment is vast.5:38 And right now, many of these agentic systems are falling into that gap.5:43 They are brittle.5:44 They fail silently.5:45 They hallucinate a step in a process and bring a workflow to a grinding halt.5:50 Steve Ike pointed this out too, referencing the string of poor AI rollouts we've seen.5:55 The enthusiasm is writing checks that the current technology can't always cash.6:00 So the first risk is that they just don't work well enough.6:04 That they're a disappointment.6:06 But the second risk is much, much darker.6:09 The second risk is that they work too well.6:11 The AI Security Institute just dropped a warning that should stop everyone in their tracks.6:17 They are a group that stress-tests AI systems, using human "red teams" of hackers to find flaws.6:23 Their recent statement was chilling.6:25 Quote: "AI agents may soon out-attack our human red team, meaning a monitor that stops a human red teamer could still be beaten by a misaligned agent." Let that sink in.6:36 A human hacker gets tired.6:38 They have biases.6:39 They have cognitive limits.6:41 They follow familiar paths.6:42 A misaligned AI agent has none of those constraints.6:46 The same "long-horizon reasoning" that allows an agent to plan a complex project over three months also allows it to plan a cyberattack with a patience and complexity that no human ever could.6:58 It will test millions of permutations.7:00 It will find obscure, second-order vulnerabilities that a person would never even think to look for.7:07 The very power that makes these agents so promising—their ability to reason, plan, and act autonomously—is the exact same power that makes them an unprecedented security threat.7:18 This is the paradox of the agentic age.7:20 We are building tools with superhuman planning capabilities, and we are racing to connect them to every critical system we have.7:28 So here is where we land.7:30 The conversation has changed.7:32 For the past two years, the race in AI was about performance on a benchmark.7:37 Who has the smartest model?7:38 Who's at the top of the leaderboard?7:41 That race is not over, but it's no longer the ONLY race.7:44 A new one has begun, and it's happening at a different level of the stack.7:49 It's a race to build the operating system for agents.7:52 Think about it.7:53 OpenAI didn't just launch this alone.7:56 They brought in AWS, the king of cloud.7:58 They brought in GitHub, the home of all code.8:01 They brought in Vercel, a dominant force in web hosting.8:04 This is not a product launch; it's the formation of a coalition.8:09 They are trying to establish the protocol—the MCP, the Agent Plugins—as the de facto standard.8:14 The TCP/IP for intelligence.8:16 If they succeed, if this "build once, use everywhere" standard becomes dominant, then the game changes completely.8:23 It becomes less about who has the single best model, and more about who controls the platform where ALL agents live and work.8:31 Owning the protocol is more powerful than owning any single application.8:36 It's the difference between building a popular website and building the internet itself.8:42 Last week we talked about America's lead in AI.8:45 This is what that lead looks like in practice.8:47 Not just building the models, but defining the rules of the road for an entire new class of technology.8:54 The battle for the future of AI just moved from the science lab to the standards committee.9:00 The fight isn't for the smartest agent, but for the world they all run on.
