For three days, everyone's been declaring the future. Today, somebody finally showed us the plumbing. Before we get to today's development, some context. Chinese open-weight AI was already a policy story before this a16z episode dropped. Kimi K3's release and Xi Jinping's World AI Conference speech turned model access into a strategic fight over governance. Then All-In brought the U.S. side into focus: open-source access, Chinese distillation, and alleged regulatory capture by Anthropic and OpenAI. This is Tech Podcast Podcast — and today, it's agents that actually run: Codex, Claude Code, plus Altman making his compute bet. Let's see which numbers survive contact. We start with Codex. For updates on this story — Chinese open-weight AI governance — tap follow so the next episode lands in your feed. Here's Akshay Nathan at Latent Space:
Less than two weeks after their July 9th launch, OpenAI said ChatGPT Work and Codex had reached 10M million users combined (as we cover in the pod, Codex now powers ChatGPT Work, so all ChatGPT Work users are now users of the Codex harness, even if they aren’t traditional engineers) — showing the early innings of what happens when you graduate from coding agents to knowledge work agents
All week, it's been big declarations, model wars, panic. Akshay Nathan on Latent Space brought a hard number: 10 million users on Codex and ChatGPT Work, less than two weeks after the July 9th launch. And here's what I keep rereading: Codex now powers ChatGPT Work. So those 10 million users aren't all engineers. Every knowledge worker in that group is running the coding harness, whether they know it or not. Right, and say that out loud: the coding agent and the knowledge-work agent are literally the same system now. Their framing is that roughly a hundred times more people use code than can write it. That's the audience they just walked into. I love that the reorg tell was buried in there too. Greg and Tibo sliding over to run ChatGPT — the org chart moved before the product did. You won't find that operating detail in the press release. And that gets at what's been bugging me all week. When Altman talks about AI improving itself without attaching a mechanism — this is the closest we've gotten to one. It doesn't prove the case, but it gives us a real system with a number and a date attached. Sure, one caveat — a usage-reset tweet hyping the next drop is marketing; nobody audited it. But 10x MAU since January is a specific enough claim that I'll take it seriously until someone shows me the churn. From Boris Cherny at Wave AI Podcast:
Yeah, there's a lot that goes into every new model and there's a lot of new capabilities that we teach and get the model to do. Whenever you do model training, you try to teach a whole bunch of different things and most often it doesn't work. But some subset of the things the model does learn.
Arc AGI 3 jumped to 30 percent. Boris puts the old ceiling somewhere between the low single digits and the low teens. Thirty percent blows through that. And Opus 5 shipped the day before they taped this. What I want from this one, though, is the security claim: Anthropic can't demonstrate a prompt injection on Opus 5 anymore. Can't demonstrate it, or can't be bothered to try that hard? Very different sentences. Right, and this is a YC show — a friendly room. If nobody pushes on 'can't,' all you've got is a pull quote. I would've killed for that follow-up here. I put more stock in the operating detail: the model running for days in auto mode. Put that next to the Codex-to-ten-million piece we just covered, and you can see both labs building the frontier model into the harness. The overlap today is the coding agent quietly becoming the work agent. It's 36 minutes, and I'd cue the security claim first. Barbell Insights writes:
We started calling it clouds. We started calling it chip fab. We started calling it energy providers, and everyone's like, you're totally crazy. This is impossible. No industry has ever moved like this. We've been around. There's these booms and busts. It's not going to go up in a straight line. This is reckless.
So Altman goes on Invest Like the Best and the big reveal is… fundraising's hard, most people say no, and all you need is one yes. That's startup wisdom from the guy building the most expensive infrastructure in human history. Framing it like a seed round is cute, but the number underneath is real: he's describing gigawatt data centers, each requiring roughly 10,000 construction workers for 18 months. That's the actual bet. Microsoft was the first yes, back when everyone said the compute demand was irrational. Now the list includes Oracle and NVIDIA, along with energy providers. Basically, he's narrating a straight line up and daring you to say booms and busts don't apply. That's where I get stuck. Every historical infrastructure buildout went through a bust, and his answer is, 'No industry has ever moved like this.' Okay, but he's still giving us a claim without evidence. Right, and it lands differently after the Codex episode we just covered — 10 million users in under two weeks. That's a mechanism you can point to. The Altman hour supplies the mood music: scaling laws, abundance, no ceiling. So queue it for the Microsoft origin story and the gigawatt math. Skip it if you want someone pushing back on the abundance sermon — Patrick isn't the guy stress-testing 'straight line forever.' From The Twenty Minute VC:
Matt Murphy is a Partner at Menlo Ventures, who just raised $3 billion in fresh capital, its largest pool ever. Matt's portfolio includes Anthropic, Lovable, Legora, OpenRouter, Chai Discovery, Axiom, OpenEvidence and more.
The guy who wrote Anthropic's first-ever check is now sitting on a fresh three billion dollars — Menlo's biggest fund ever. So when Matt Murphy says Triple Triple Double Double isn't venture-scale anymore, he's telling you where the LP-side bar sits now. Right. Tripling revenue two years running, then doubling twice, used to be the platonic hypergrowth story. Murphy says the bar has moved. I want to hear what replaced it. The bit I clicked for comes at minute 23: why open source won't derail Anthropic. Easy to say when Anthropic's your headline position. That's the one I'd stress-test. Chamath frames IP theft and distillation as a legal threat. Murphy's looking at the same question through a business lens — margins and moats — and he's the person closest to Anthropic's cap table. His 'do margins matter in AI' segment is where those arguments meet. This one's from The a16z Show:
Steven Sinofsky joins Theo Jaffee and Sofia Puccini for a conversation on AI regulation, open-source models, and what history can teach us about technological revolutions. Drawing on decades of experience leading products at Microsoft, Sinofsky argues that governments are rushing to regulate AI before they fully understand the technology, risking innovation in the process.
Sinofsky spends 29 minutes on a16z making a pretty simple case: governments are regulating AI before they understand it, and existing law probably covers most of it anyway. It's a real argument, and he's got the resume to make it. He spent decades running product at Microsoft and watching regulators show up late to every prior wave. The critique of the precautionary principle has some weight. Sure, but it's the a16z show telling me a16z's regulatory position. The messenger's credible; the venue is doing some of the arguing for him. Here's where it gets sharp for me, though. If even Sinofsky says current law already handles the risks, who exactly holds the off switch on an open-weight release? Once the weights are out, there's no recall. 'Existing law is enough' and 'nobody can pull it back' don't obviously coexist. Yeah, that's the gap. His argument is that open source has always accelerated innovation, and historically, fine. But frontier open weights don't behave like a Linux kernel. You can't sue your way to un-releasing model weights. We're back at the same governance fight, just in an a16z jacket. It's 29 minutes and worth it for the historical lens. I'd queue it right after someone who actually pushes back on him. Have feedback, a story idea, or a correction? Email us at techpodcastpodcast at lantern podcasts dot com. We’d love to hear from you.
You’ll find links to every story in today’s show notes, so take a look if anything caught your interest. That’s Tech Podcast Podcast for today. This is a Lantern Podcast.