The guy who spent years dunking on AI coding just shipped 100 pull requests in 90 minutes. So what changed his mind? This is Tech Podcast Podcast — DHH runs a two-model setup, Jeff Dean does napkin math, and somewhere in there the hype finally coughs up receipts. Let's start with the power-user workflow: Rails and Claude Code. Then we'll get to self-improving agents and the arithmetic behind a TPU. Here's Danar at The AI & Tech Society:
David Heinemeier Hansson, creator of Ruby on Rails and CTO of 37signals, went from AI coding skeptic to agent-first developer. This episode breaks down DHH’s Claude Code workflow, how he used AI agents to process 100 pull requests in 90 minutes, his two-model setup with Gemini and Opus, why Rails is ideal for AI coding agents, and what “peak programmer” means for the future of software engineering.
DHH spent years dunking on AI coding, and now he's run 100 pull requests through Claude Code in 90 minutes. That's the skeptic-to-power-user turn I wanted receipts on this week, and here they are. The detail I care about is that he uses two models: Gemini 2.5 and Opus 4.5. That tells us something concrete about where the frontier is. Right, but why two? If Opus could carry it alone, he'd run Opus. The fact that a user at DHH's level splits the work tells you where the ceiling is for any one model. The line I want them to press him on is 'Rails is ideal for AI coding agents.' It could be a real property of the framework's conventions, or just the Rails guy loving Rails. Those are very different episodes. Convention over configuration was always half marketing. But if agents really do work better with opinionated defaults, that's the most useful thing DHH has said in years. I'll even take the Neovim-and-tmux setup, 'peak programmer' framing and all. From Peter Yang at Creator Economy:
Karan is the co-founder of Nous Research, the company behind Hermes, the #1 personal agent and AI app on OpenRouter. We had a great chat about why open-source AI must win and how Hermes gets better the more you use it. Karan also shared a live demo of how he used Hermes to modify his favorite childhood game.
So Hermes is the number-one personal agent on OpenRouter. Karan's pitch? The personality comes from the memory and skills you build with it, whatever model sits underneath. Swap the underlying model and he barely notices a difference. Then that's the bet I want tested. If the model's interchangeable, Hermes is betting its moat on portable context, with the underlying weights treated as interchangeable. And here's the part I actually care about — it can create and clean up its own skills without bloating the working context. Everybody says agents get better with use; almost nobody explains the garbage collection. This gives us a different answer to the problem DHH's two-model setup was solving. DHH points two frontier models at the work. Karan says accumulated context is what earns trust, regardless of the infrastructure around it. The chapter titled 'how to stop AI from saying you're absolutely right' — that's the most honest line on the whole rundown. When every model sounds the same, the sycophancy tells you there's no there there. Here's where I'd hold his feet to the fire: if the skills are self-improving and model-agnostic, does that really loosen your dependence on any single frontier lab — or just move the lock-in to the memory layer he owns? Yeah, and his proof point is modding Sonic Adventure 2 with it. Weird flex, but at least it's a real workflow. I'll take the childhood-game mod over one more slide about agentic futures. Y Combinator, with Diana Hu:
In 2001, Jeff Dean and Sanjay Ghemawat did the math and realized Google’s entire search index would fit in RAM — then shipped it in a few days, and search got fast. In 2013, another napkin calculation showed that three minutes of daily speech recognition per user would require doubling Google’s server fleet. That one became the TPU.
Okay, this is the one I've been waiting all week to get to. In 2001, Jeff Dean and Sanjay Ghemawat ran the numbers and realized the entire Google search index would fit in RAM — then shipped it in a few days, and search got fast. And the 2013 calculation is even better: three minutes of daily speech rec per user would've forced Google to double its entire server fleet. That napkin math became the TPU. This is the kind of mechanism story YC almost never gets: an actual number that forced a hardware decision. Honestly, it has the same shape as the DHH piece we just hit. Small crew, does the math, ships. That's why it lands as a quiet rebuttal to the whole 'straight line forever' sermon. Dean followed the arithmetic to a hard choice: double the fleet or build the chip. And he tells Diana Hu that inference hardware is the next big area for specialization, and explains where two or three people in a room can still win. It's fifty-seven minutes, and I'd queue it over anything with 'the future of AI' in the title. Give me the guy who shipped the index in a weekend over another keynote about disruption, every time. From The Twenty Minute VC:
Joon Sung Park is the Founder and CEO of Simile, the AI simulation company building foundation models of human behaviour; allowing companies to test how real people may think, decide and act before making a decision in the real world. Simile has now raised $300 million in total, including a $200 million Series B announced last week at a $2 billion valuation, led by Greenoaks and Index Ventures.
Look at this 20VC rundown. Jensen writes an open-weights letter. Kalanick's back with $1.7 billion for something called Atoms. Then Etched pulls in $300 million to take a swing at Nvidia. It's a whole cinematic universe in one episode. The Etched number's the one I'd actually chew on. Three hundred million to build inference silicon against Nvidia and Google's own TPUs, while Google Cloud posts 82 percent growth. Those are two fresh data points for the margins-and-moats fight. Google Cloud grows 82 percent and the market still tanks? Investors see cloud growth like that and still price in a leaky moat the second custom chips show up. The wild Simile line is Joon Sung Park saying there are models where they'd pay $100 million for a single query. That's a data-acquisition claim dressed as a flex. A hundred million a query, and then five minutes later he's arguing stock markets won't exist in five years. Pace yourself, my guy. Give me the reward-function detail before you retire the NASDAQ. Right — the useful bit's buried at minute nineteen: clear, fast reward functions. That's a real operating filter. The stock-markets-won't-exist prediction can make the clip; I want the reward-function detail. This one's from The a16z Show:
Inspired by time spent working inside dental practices, Pelle and Renken set out to automate the administrative work that keeps healthcare providers away from patients. They discuss how AI agents are changing billing, insurance claims, patient payments, and other operational workflows, allowing practices to spend less time on paperwork and more time delivering care.
A16z sits down with two guys who actually spent time inside dental practices, watching front desks drown in insurance claims. Okay — I'll take that origin story over another founder who read a McKinsey deck. And it's the second episode running with vertical AI aimed straight at healthcare admin — Abridge on Friday, now Lassie's billing and claims automation. We've got a real pattern here. Here's what I want them to press on: Abridge was pitching a universal interoperability layer. Does Lassie plug into it, or is every one of these agents rebuilding the same insurance plumbing from scratch? Right. The pitch is that these agents do the work instead of just storing the information. Fine — but small dental practices are where reliability really bites. One botched claim and you've lost the customer. I want to hear what breaks, not the demo. Exactly the DHH lesson from earlier — throughput is where you find the receipts. Show me the claims that clear clean, then we'll talk. If Tech Podcast Podcast keeps you informed, subscribe and leave a quick review wherever you're listening. It helps other people find the show, and we really appreciate it.
You'll find links to every story in today's show notes if you want to dig into anything. That's Tech Podcast Podcast for today. This is a Lantern Podcast.