Dan Mac's Weekly Podcast Roundup: October 13-20, 2025
Every week I swim through hours of podcast content so you don’t have to. This is what stuck with me.
Karpathy’s Decade Warning
Andrej Karpathy — AGI Is Still a Decade Away | Dwarkesh Podcast
Karpathy dropped a reality check on the AI hype cycle this week. While everyone’s calling this “the year of agents,” he says we’re actually in the decade of agents.
The gap between what current AI can do and what we need it to do is massive. Think about hiring an intern. You wouldn’t just hand them your entire codebase and say “fix it.” They need intelligence, multimodality, continual learning, the ability to remember what you told them yesterday. Current models have none of that.
What caught me is how he frames the problem. These aren’t unsolvable issues. They’re just hard. Really hard. The kind of hard that takes a decade to work through when you’re at the frontier.
He’s particularly skeptical of reinforcement learning as currently practiced. It’s noisy, inefficient, sucking supervision through a straw. Every token in a successful solution gets upweighted, even the wrong turns. Humans don’t work like that. We reflect. We synthesize. We don’t do hundreds of rollouts and hope the right answer emerges from the noise.
And here’s the kicker: models are actually worse at creating novel code than they are at everything else. They excel at pattern matching, at regurgitating what exists on the internet. But ask them to write something that’s never been written before? They collapse into the same narrow manifold of solutions.
This matters if you believe the fast takeoff story. If AI can’t automate AI research as easily as it automates CRUD apps, your timelines for superintelligence just got a lot longer.
Tyler Cowen on Oral Culture
Tyler Cowen | Tetragrammaton with Rick Rubin
Cowen made an observation that’s been gnawing at me: we’re shifting from written to oral culture faster than anyone realizes.
YouTube. TikTok. Audiobooks. ChatGPT as therapist. The sources of authority from 2000 still exist—they still do their conferences, get their funding, publish their papers—but they’re increasingly irrelevant.
Reading fosters analytical thinking. It forces convergence on truth through reasoned back-and-forth. Oral culture excites, motivates, involves. But it also encourages mood affiliation. You judge ideas by their vibe, not their substance.
This isn’t entirely bad. Millions of people using ChatGPT as therapist is probably net positive. But something’s being lost in the translation from print to voice.
What struck me most: Cowen thinks AI will make high-quality education available everywhere. Not fully automated—you still need tutors—but the constraint won’t be geography anymore. You could live in rural Montana and get world-class instruction.
That’s the optimistic case. The pessimistic case is we’re heading toward compute scarcity, where access to intelligence becomes the new class divide.
Also: stablecoins are more important than people think. Countries want dollars without Uncle Sam’s oversight. Europe might prefer stablecoins to the euro. This is a bigger deal than the crypto discourse suggests.
Brett Hall vs. AI Doomerism
Ep 248- AI and Philosophy of Science | ToKCast
Brett Hall went after the AI doom crowd with surgical precision.
His target: the whole edifice of thinking that treats AI as predictable and humans as prediction machines. Bostrom, Yudkowsky, the whole gang.
The core issue is knowledge creation. It’s inherently unpredictable. Not just unknown—unknowable. The future depends on knowledge we haven’t created yet. And we have no algorithm for creating explanations. If we did, we’d already have AGI.
But the doomers assume away this problem. They treat intelligence as scaling compute, as predictable extrapolation from current capabilities. They miss that LLMs are fundamentally language calculators, next-token predictors. Impressive, yes. But not explanatory.
Hall’s most damning critique: the doomers don’t actually believe their own predictions. Look at their affect. Real belief in imminent doom produces grave seriousness, not levity. Compare them to kids who’ve been told climate change will end the world—those kids cry. The AI doomers laugh. They’re on speaking tours, writing op-eds, advising governments.
They know they’re getting away with something.
OpenAI Goes Vertical
Episode 8 - OpenAI X Broadcom and the Future of Compute | OpenAI Podcast
Sam Altman and Greg Brockman laid out OpenAI’s compute strategy. The short version: they’re going full vertical integration.
From etching transistors to the token that comes out of ChatGPT. Chip design. Rack design. Networking. Algorithms. The entire stack, optimized end-to-end.
Why? Efficiency gains. Huge ones. Better performance, faster models, cheaper models. And as costs drop, usage explodes.
The scale is insane. OpenAI started the year at 2 gigawatts. They’ll end at 2 gigawatts. Recent partnerships will take them to 30 gigawatts. For context, that’s multiple data centers the size of small cities.
And even that won’t be enough. Brockman’s vision: everyone gets their own AI agent running 24/7. That’s 10 billion chips. We’re nowhere close.
The fundamental insight: intelligence is the driver of economic growth. AI amplifies everyone’s intelligence. As models improve, productivity increases exponentially. The demand is infinite.
What I found most interesting: the collaboration with Broadcom on custom inference chips. They’re not just buying off-the-shelf hardware. They’re designing the silicon specifically for their models, their workloads. Stacking chips in 3D. Integrating optics for 100 terabits of switching.
This is the biggest joint industrial project in human history. And it’s just getting started.
Michael Dell on Naivete
Michael Dell, Dell Technologies & David Senra | Founders Podcast
David Senra interviewed Michael Dell about the early days. One exchange stuck with me.
Dell’s method for understanding things: take them apart. He’d buy an Apple II, bring it home, immediately disassemble it. His parents would get frustrated. Sometimes the things still worked after. Sometimes not.
But the insight: you can’t understand something without understanding its internals.
Dell said something profound about entrepreneurship: you need a combination of naivete and confidence. Naivete because you don’t know enough to realize something won’t work. Confidence to try anyway.
The danger is when naivete tips into arrogance. That’s when you stop learning. When you think expertise means you’ve figured it out.
There’s a Henry Ford quote David shared that nailed it: “If I wanted to sabotage my competition, I’d fill their ranks with experts. Experts know so much, they’re so convinced they’re right, they get no work done.”
At the limit, nobody knows anything. Especially about the future.
Closing Thoughts
The through line in this week’s podcasts: underestimating complexity.
Karpathy says agents are a decade away because building them is harder than it looks. The doomers say AGI is imminent because they don’t understand knowledge creation. OpenAI is building the largest industrial project in history because the compute requirements are astronomical. Dell succeeded because he didn’t know enough to know it wouldn’t work.
In every case, the people doing the actual work are humbler about timelines and capabilities than the observers.
This is a pattern worth remembering.
The decade of agents is here. Not the year. The decade. Which means the real work is just beginning.
What matters now isn’t predicting what will happen. It’s deciding what we’ll build.
This roundup covers podcast episodes highlighted in my Obsidian vault from October 13-20, 2025. Want these delivered weekly? Subscribe below.