All week it's been capability, moat, ROI — and today the bill showed up. Companies are scrambling to stop burning tokens. If you're just joining, agent-native development has been moving from coding assistants into full company operating models. The last big case was Gusto — CTO Eddie Kim said a five-person team used Claude Code to ship Gusto Cofounder from zero code to a tier-one launch in ten weeks. That turned agentic development from workflow advice into actual org design. This is Tech Podcast Podcast. Today — a chip company betting transformers are forever, a self-improving-agent loop with humans still in the middle, and the Tokenpocalypse. Let's start where the money's on fire. Here's Patrick O'Shaughnessy at Invest Like the Best:
They have since done it, taping out a working chip on their first attempt and becoming the first hardware company founded after ChatGPT to do so. They already have more than a billion dollars of customer demand for their first product, and have raised eight hundred million dollars to build it.
Etched taped out a working chip on its first attempt — the first hardware company founded after ChatGPT to pull that off. That almost never happens; first silicon usually comes back broken. And the bet underneath is that the chip is fast specifically because it assumes transformers never go away. It's baked into the architecture. You can actually falsify that wager instead of just nodding along to another capability claim. Right, so if Noam Brown or Evans is right that the architecture's already getting outrun, Etched is stuck with eight hundred million dollars of silicon optimized for the wrong world. Over a billion in customer demand already, though. Somebody's putting the same bet into a purchase order. So inference cost just turned into a hardware thesis with real money behind it. And it lands differently after a week of arguing whether the model layer commoditizes. If you're designing silicon around transformers being permanent, you're voting hard that the moat is real. Latent Space writes:
We’ve heard a lot about loops at the AI Engineer World’s Fair this week. Another buzzword is autoresearch, which involves building an “outer loop” where agents help maintain and improve the primary system, using feedback signals, evals and human input to make progress over time.
So Introspection's whole pitch is an 'outer loop' — agents that maintain and improve the primary system using evals, feedback signals, and human input. Gavrilescu calls the units 'recipes.' That's the Latent Space follow-up to the coding-copilot thread — they've moved from copilots to full self-improving loops. And the part I like is that humans stay central. There’s no autonomous-magic handwave here; the factory learns from people first. Right, and the tell is where the human sits. Gavrilescu keeps the checkpoints in — basically the same instinct Gusto's CTO had with eval-first shipping, just with a fancier name. And it's falsifiable, which is why I trust it more than the usual self-improving-agent hype. He gives you the mechanism — recipes, checkpoints, the Pi framework — so you can actually go test whether the loop converges or just spins. The xAI agent-infra guy starts a company selling infrastructure for the thing he was building at xAI. I'd want to see one recipe run end to end before I call it a factory. Here's Joseph Cox at 404 Media:
We start this week with Joseph’s story about the Tokenpocalypse, which is companies scrambling to stop spending so much on AI after providers started charging per AI token. After the break, Joseph and Emanuel tell us about the ways companies are trying to do this, including using a tool to make their LLMs talk like cavemen.
So all week I've been asking who actually pays the inference bill, and 404 Media just hands me the answer with a name on it. The Tokenpocalypse — companies scrambling to stop burning tokens once providers started charging per token. And it's the first cost-side reporting we've seen that isn't back-of-the-envelope. Every ROI pitch we floated this week assumed the bill was survivable. Now companies are deciding it isn't. And the fix is genuinely unhinged — they're making Claude and Codex talk like cavemen. Fewer tokens per prompt if the model drops the articles. That's the operating detail, right there. Nobody polishes a caveman-mode workaround into a keynote. It reframes the whole ROI conversation, though. Now you're asking: does AI pay off at this usage, or does it stop penciling out? Somebody's finance team should've plotted that curve months ago. From Theresa Loconsolo at TechCrunch:
Humble Robotics founder and CEO Eyal Cohen is one of them. Cohen was at Otto when Uber came calling, later followed Anthony Levandowski to Pronto, and after two decades bouncing between deep tech bets in the Bay Area, his new company came out of stealth in April with $24 million to build a fully autonomous, cabless electric hauler for freight.
Travis Kalanick is back building a robotics company, and TechCrunch is openly calling this a 2016 rerun. When the guy who blew up the first AV wave is a marker for the second one, that tells you something. The person at the center is Eyal Cohen — Humble Robotics, out of stealth in April with $24 million for a cabless electric freight hauler. And his résumé is basically the whole story: he was at Otto when Uber bought it, then followed Anthony Levandowski to Pronto. Otto and Pronto — that's a tour through the two most legally radioactive companies of the first wave. Levandowski literally got pardoned. So the pitch is, 'I lived through the mess, trust me this time.' What I want out of the Equity interview is whether Cohen names what's actually different now — the tech, the unit economics on freight — or whether $24 million is buying déjà vu. Freight is a narrower, more measurable problem than robotaxis. If he shows the route math, I'm listening. Right, a cabless hauler on a fixed freight lane is at least a bounded bet. The tell will be whether Korosec presses him on why 2016 stalled, or lets him coast on the war-story charm. Fifteen years of deep-tech scars only counts if it changed the plan. If Tech Podcast Podcast is part of your daily routine, consider subscribing or leaving a quick review wherever you’re listening. It really helps other curious listeners find the show.
You’ll find links to everything we covered today in the show notes, so if a story stuck with you, that’s the place to dig in a little further.
That’s Tech Podcast Podcast for today. Thanks for listening, and we’ll be back next time. This is a Lantern Podcast.