0:00 AMD just demonstrated a six-fold increase in CPU throughput for AI agent workflows.0:06 That might sound like a technical detail, but it’s a signal flare for a fundamental change in what AI is even for.0:14 Last week, we talked about ChatGPT's shareable profiles making AI more reusable.0:20 That was about software.0:21 This week, we're seeing the hardware consequences.0:25 The race isn't just about who has the biggest model anymore.0:29 It's about who can get an AI to do the most work, on its own.0:34 And that changes everything, right down to the silicon.0:38 So here's the lay of the land this week.0:41 The big story is the shift to AI that doesn't just talk, but acts.0:45 We'll get into the deep dive on that in a minute.0:49 But first, the headlines you need to see.0:52 The biggest move comes from OpenAI.0:54 They're teasing a new model, the "O" model, and a feature called "Always On Agents".1:00 The pitch is simple, and it's a big one.1:03 What if you could hand an AI a job, go to bed, and wake up with it finished?1:09 That’s the promise.1:10 This isn't about asking better questions.1:13 It's about delegating entire projects.1:16 This is the software that the new hardware from AMD is being built to run.1:21 And OpenAI isn't thinking small.1:24 The AI Frontier Weekly Brief just noted that the company is scaling its efforts to as many as ten thousand AI agents.1:32 Ten.1:32 Thousand.1:33 That's not a research project.1:35 That’s an industrial-scale operation to have AI start producing genuinely new knowledge.1:42 It’s a move from AI as a tool you wield, to AI as a workforce you manage.1:47 Of course, there's a reality check.1:49 A thread from the developer Burkov is making the rounds, and it’s a necessary dose of skepticism.1:56 He notes that while these agents are getting incredibly good at complex, long-running coding tasks...2:04 they are still struggling to sustain genuine scientific research.2:08 An agent can write code for a week to solve a defined problem.2:13 But can it manage a multi-year research project with ambiguous goals and unexpected turns?2:19 The answer, for now, is no.2:21 There's a frontier here, and we haven't crossed it yet.2:25 The ability to execute a task is not the same as the ability to have a breakthrough.2:31 This growing power, even with its limits, is forcing a very old debate back to the surface: who should control this?2:40 An account called Sentient AGI put it bluntly this week, with a mission statement that's also a warning.2:47 Their goal is "To ensure that Artificial General Intelligence is open-source and not controlled by any single entity." As these agents become more autonomous and more capable, the question of whether they belong to a corporation, to the public, or to no one at all becomes THE central question of governance.3:10 It’s a battle for the soul of the next digital age.3:13 And just a few quick hits to ground us.3:16 A post from Kevin Bryan is a good reminder of the fundamentals.3:21 He says, and I’m quoting here, "LLMs DON'T 'STORE' UNDERLYING TRAINING TEXT.3:26 It is impossible—the parameter size of GPT-3.5 or 4 is not enough to losslessly encode the training set." So when you hear people worry that these models are just giant copy-paste machines, remember that.3:41 The process is transformation, not storage.3:44 It's a distinction that matters.3:46 Finally, a bit of perspective from Dave, a developer at OpenAI.3:51 He just posted, "OpenAI is nothing without its people." In a week where we're talking about autonomous agents and silicon and scaling to ten thousand bots...4:02 it’s a good reminder.4:04 Behind all of this, there are still people.4:07 Making decisions.4:08 Writing code.4:09 And trying to steer this whole incredible, terrifying, world-changing enterprise.4:15 So that’s the field.4:17 A push toward autonomous agents from OpenAI.4:20 New hardware from AMD to run them.4:22 A reality check on their current limits.4:25 And a burning question about who gets to be in charge.4:29 Now.4:29 Let's connect the dots.4:31 Okay.4:31 Let’s go deeper.4:32 The real story this week isn't just one of these threads.4:37 It's the way they all weave together.4:39 It's the connection between AMD's new processors and OpenAI's new agents.4:45 This is a story about where the "brain" of AI is moving.4:49 For the last few years, the AI story has been simple.4:52 A massive model, like GPT-4, lives in the cloud.4:56 It sits on thousands of hyper-specialized GPUs.4:59 You send it a prompt from your laptop or your phone.5:03 It does the thinking up there, and sends you back an answer.5:07 Your device was just a dumb terminal.5:10 The real work happened somewhere else.5:13 That's changing.5:14 And it's changing because of what we're asking AI to do.5:18 The shift is from simple inference to what are called "agentic workflows." This is the key.5:25 Answering a question is inference.5:27 An agentic workflow is about achieving a goal.5:30 Think about it like this.5:32 Asking "What's the capital of Mongolia?" is an inference task.5:37 A single round trip.5:38 You ask, it answers.5:40 Ulaanbaatar.5:41 Done.5:41 But what if you tell an AI, "Figure out the cheapest way to ship a ten-pound package from my office to Ulaanbaatar, book the shipment, and email me the tracking number." That is NOT a single task.5:55 That is a project.5:56 The AI agent has to:
One: Understand the goal.6:00 Two: Break it down into steps.6:02 It needs to search for shipping carriers, compare their prices and times, maybe read their API documentation.6:10 Three: It has to execute those steps.6:12 This means running code.6:14 It might need to read a file from your computer with your office address.6:19 It might need to run a web scraper.6:22 It might need to call a booking API.6:25 Four: It probably needs to coordinate sub-agents.6:28 Maybe one agent is good at finding data, and another is good at filling out forms.6:34 The main agent has to manage them.6:37 Five: It has to check its work, handle errors, and finally, report back to you.6:42 Do you see the difference?6:44 It's the shift from a search engine to a project manager.6:48 From a calculator to an intern.6:51 And this is exactly what OpenAI is talking about with its "Always On Agents." The "go to bed and wake up with it finished" dream.7:00 But here's the catch.7:01 Where does all that work happen?7:04 All that file reading, program running, web scraping, sub-agent scheduling?7:09 It doesn't happen on a GPU in the cloud.7:12 It happens on a CPU.7:13 This brings us to the AMD announcement.7:16 For years, the GPU got all the glory in AI.7:19 Nvidia became one of the biggest companies in the world selling the shovels for the AI gold rush.7:26 The CPU was seen as...7:28 old news.7:29 Important, sure, but not where the action was.7:32 Now, the CPU is making its comeback.7:35 AMD put out a statement that is so clear, and so important.7:39 "The CPU remains the core that connects the overall workflow...7:43 CPU's multi-core computing power, single-thread performance, and scheduling efficiency will directly affect tool execution, program compilation, local data processing, application response speed, and overall task completion time." Let me translate that.8:02 The GPU is the brilliant, specialized mathematician who can solve one kind of problem incredibly fast.8:09 The CPU is the general contractor.8:12 It's the thing that reads the blueprints, calls the plumber, schedules the electrician, and makes sure the supplies arrive on time.8:21 In the old model of AI, we only needed the mathematician.8:25 In the new world of AI agents, you desperately need the general contractor.8:31 And this is why AMD's benchmark number is so significant.8:35 That six-fold increase in throughput wasn't just a theoretical number.8:40 They tested it by running six—SIX—ChatGPT 5.5 High agents simultaneously on a single laptop.8:47 Each agent was working on a complex coding project, doing things like analyzing code structure, compiling it, running tests, and querying databases.8:57 This is the new paradigm.8:59 Not one big brain in the cloud, but a swarm of smaller, specialized brains running locally, coordinated by a powerful CPU.9:08 The work is coming down from the cloud and landing right on your desk.9:13 Or, more accurately, right on your laptop's processor.9:17 This explains OpenAI's strategy.9:19 The "O" model and its Always On Agents aren't just a software update.9:24 They are the demand signal for a new kind of hardware.9:28 They are creating the killer app that makes you need a laptop with a next-generation, AI-optimized CPU.9:36 It's a classic technology two-step.9:38 The software pushes the boundaries, creating a need that the hardware then races to fill.9:44 So what does it all add up to?9:47 You're seeing the architecture of the next decade of computing being laid out in real time.9:53 It's a shift away from centralized cloud intelligence and toward distributed, localized, autonomous execution.10:01 The intelligence isn't just in the model's parameters anymore.10:06 It's in the system's ability to act in the world.10:09 To run programs, to read files, to complete tasks.10:13 This is a much bigger deal than just faster chatbots.10:17 It's about turning our computers from passive tools we operate into active partners that work for us.10:24 And the companies that build the best "general contractors"—the best CPUs for managing these agentic workflows—are going to be the ones who own the next phase of the AI revolution.10:37 Just a quick note here: as we talk about these technologies and companies, remember this is for informational purposes only.10:46 It is not financial or investment advice.10:49 The landscape is moving fast, and you should always do your own research.10:54 The bottom line is, the quiet, technical-sounding news from AMD this week is the other half of the loud, exciting promise from OpenAI.11:04 One is building the engine.11:06 The other is building the car.11:08 And together, they're about to take us somewhere entirely new.11:13 So where does this leave us?11:15 What does this week set up?11:17 The AI race has a new finish line.11:19 For a while, it was about who had the most parameters.11:23 Then, as we talked about last week, it shifted to who had the most useful, shareable AI constructs.11:30 Now, a new axis of competition is clear: autonomy.11:34 The central question is no longer just "How smart is your AI?" or "How big is your model?".11:40 It's becoming "How much can your AI get done without you?" This week, we saw the blueprints.11:47 OpenAI is designing the autonomous software agents.11:51 AMD is designing the local, client-side hardware to run them.11:55 The entire stack, from silicon to software, is being re-architected around this idea of giving AI a to-do list and walking away.12:04 The implications are huge.12:06 It changes the economics of knowledge work.12:09 It changes our relationship with our own devices.12:13 And it recenters the importance of the computer right in front of you, not just the massive data center hundreds of miles away.12:22 The next quarter, the next year...12:25 it won't be defined by a chatbot that can write a better poem.12:29 It will be defined by the first widely-adopted agent that can reliably book your travel, file your expenses, or debug your code while you sleep.12:39 This week wasn't just another series of announcements.12:43 It was the moment the industry picked a new direction.12:47 The race is on.12:48 Not for better answers, but for better agents.12:52 The next big question isn't 'what can your AI say?'.12:55 It's 'what can your AI do?'.12:58 And the answer to that will define the next generation of everything.
