Paul's Lab: Theme Songs, Knowledge Bundles and Radios

Top-Line Summary

October means a guest, and this month it’s Paul, the person behind the show’s intro and outro songs. He walks through how those songs come out of several models critiquing each other, the OKF knowledge bundles he says save him about half his tokens, and a home lab where Claude has free rein. The back half goes to hardware: a $4 ESP32 display, software-defined radio, and mesh networks for farm sensors.

Show Video

The “Pre-Show” Context

Everyone is still cleaning up after the “patchpocalypse,” the end-of-October pile of patches Travers calls the end of the world from a PHP perspective. Brett has found a few more things this week that made him stop and say “wait, what is this?”

The Engineering Rundown

• Paul and the theme songs (01:47)

  Paul has about 30 years in IT, starting in aerospace and broadcast engineering before cloud and AI, and today runs enterprise solution architecture. Someone sent him articles on MCP and ACP, so he had AI stitch audio generation platforms together, fed it meeting notes because he can’t write songs, and asked for something catchy. The intro song’s lyrics came from ChatGPT and were then run through Suno several times, because he found some platforms weaker on lyrics than others. Brett and his wife walk around the house singing it, and the new ones tend to land in Slack around 12:30 or 1 a.m., which Paul says is because Claude Code sometimes takes 10 to 20 minutes on a request and he opens another tab to make a song while he waits. Brett will take the upbeat electronic ones over the country ones.

• OKF knowledge bundles (07:31)

  The songs stay consistent because Paul keeps a music repository in OKF (Open Knowledge Format), using a tool at okfgem.com that he says came out of Google research. Point it at a source folder, Word documents, Azure DevOps or GitHub and it builds a “knowledge bundle”: summaries by concept, tag or category, a graph of the knowledge, and hints that tell an agent where to look instead of reading the whole folder. You load one with /okf consume, and at the end of a heavy session it asks whether to update the bundle because it has drifted. Paul thinks it’s saving him about 50% of his token consumption and says he can’t live without it. The hosts haven’t tried it.

• Second brains and smart wikis (11:29)

  Brett would be toast without his own brain, mostly because Claude still occasionally skips the “check the brain first” instructions, assumes something, and admits it didn’t check when he asks. Travers is still on his smart wiki: markdown with cross links, a graph DB on the side, and lately some experiments with PageRank. His point is that an agent spawning into a project starts with nothing every time, so a cross-linked wiki is the difference.

• Copilot for search, Claude for reasoning (13:37)

  At work Paul has to balance enterprise data protection against tool choice, so he jumps between Microsoft Copilot and enterprise Claude, with Gemini showing up in every Google search whether he wants it or not. His split: Copilot is good at SharePoint and enterprise content search, Claude is good at orchestration and deep reasoning, so Claude takes the hard problem and calls Copilot as a tool to find real artifacts that back up its reasoning. The models talk to each other over MCP or plain API calls.

• Local models and a lab Claude runs itself (15:26)

  At home Paul has GPUs, LM Studio, models off Hugging Face and Ollama, and a growing interest in ComfyUI and diffusion models for video, which he warns is a hobby where you run out of money fast. He mostly lets the harness pick a local model (GLM comes up a lot) and runs out of disk space quickly. The stack is Coolify for rented servers on Hetzner or AWS, Proxmox for Docker hosts, and Portainer to supervise them. Then he tells Claude to do what it wants in the lab, which is contained, and it builds servers to support his local models, full websites and VPN bridges to reach them from outside. As Paul put it, “when I say it out loud, it sounds horrifying.”

• Brett’s YOLO Home Assistant (18:53)

  Brett has gone partway down the same road: Claude can SSH into his Home Assistant box, keeps the config in its own git repo, and runs deploys and restarts. He likes home automation but got to the point of wanting to throw the whole thing into the backyard, so now he tells Claude what he wants and lets it maintain it. So far, knock on wood, everything still works, and Claude regularly suggests things he hadn’t thought of. Travers’ comparison: at the start of the year you had to name the exact files and the exact change in your IDE, and now you tell it to consume a PR and babysit it.

• AI-generated podcasts and narrated courses (21:23)

  Brett hasn’t forgotten Sam Altman’s suggestion to make a podcast of your calendar to listen to with your kids (talk to your kids), but he does generate 30-minute study podcasts from his notes before exams and is routinely shocked by the quality. Paul took a “front-end slides” skill from GitHub and added narration, so each slide has a narrator button that speaks as an instructor guiding students rather than reading the slide, plus background music on the intro.

• What happens to training? (24:03)

  Brett started his business as a trainer, so this one is personal. If you spend half an hour feeding a model good sources (vendor documentation, references you’ve vetted), it produces content about as good as most paid courses, and he doesn’t know what happens to that industry. His view is that it’s in for a tough haul as models keep improving.

• Gemini Argon and paying Google (25:33)

  A new Gemini snapshot came out the day before recording and is topping the benchmarks. Brett’s running complaint is that he still can’t figure out how to give Google his money: a business account doesn’t seem to include it, and the API console told him he already had two keys and couldn’t have a third. Travers has used it a little in the Gemini CLI and says it still has “Gemini-isms” in code work, going off on tangents and losing the plot when editing files or landing a PR. Paul navigated 16 menus to turn on one feature, so he mostly uses NotebookLM and Gemini’s image generation, and goes to Claude for anything deeper.

• Architecture diagrams with real AWS icons (28:54)

  Brett has been getting Claude to read a code base and produce data flow and system design diagrams alongside the markdown, and they’ve gotten very good. He told it once to use the official AWS icons and gave it the source, and now it does that every time. You still have to read everything and understand the code base yourself. Paul generates HTML documentation with SVG graphics and interactive graph views using D3-style JavaScript libraries. Travers has every research task produce an HTML review page with a diagram, using SVGs, AWS icons, and Mermaid for flowcharts.

• Personalized software (31:18)

  Brett and Travers joke every week that they’re building the same tools, each built around how they think about the problem. Paul has the same pattern: a template he reuses, so “make an architecture diagram from the meeting we had 10 minutes ago” comes back in the same theme every time with almost no prompt work.

• ArtCraft instead of Adobe (32:24)

  Brett asked himself why he pays US$30 a month, about $360 a year, for an Adobe PDF reader, and went looking for ArtCraft’s clean-room alternative. The GitHub page wanted him to compile it, which nearly put him off, until he found the loader, which works like Creative Cloud. The PDF tool is tiny, opens instantly, and doesn’t make him log in 14 times. Their presentation app opens PPT files fine too, though Brett rarely makes PowerPoint anymore. His presentations are HTML now, and Claude builds them the same way every time, probably from a mix of the model and his brain’s record of past decks. Paul has tried the tools and calls them close to pixel-perfect on features and fast. Travers has no Adobe use case.

• A $4 ESP32 that fakes hacking your computer (36:27)

  Paul plugged a tiny ESP32 display board into USB and told Claude to make it show a fake “hacking your computer” screen. It now scrolls through probing a drone fleet and talking to satellites. Claude found the display’s hardware spec, pulled the right libraries and got it working in one pass. The board came from AliExpress for four dollars, with instructions in Chinese he can’t read, and you can’t buy just one.

• Software-defined radio and mesh networks (38:16)

  Paul is a licensed amateur radio operator, and his main hardware interest is software-defined radio: tuning radio frequencies through Python libraries instead of expensive hardware, from (as he said it) around 10 Hz up to 6 GHz. Flipper devices work on the same idea, and some can transmit too, which needs a license. He showed a LilyGo T-Deck, which looks like an old BlackBerry, and uses license-free frequencies to talk over long distances without a cell carrier. Friends in agriculture want to mesh 16 or so of these to report soil potassium, nitrate and moisture from greenhouses and fields covering thousands of acres, back to a server with logs and dashboards. Some units run two or three years on battery, or off a small solar panel on a post. Paul says this would have taken him weeks or months to build before.

• Building what’s in your head (43:07)

  Brett’s friend, who is not a developer, has used Claude to build a phone app to remote start his TR6, a thumbprint reader to start it, and remote thermostats for his barbecues, going from breadboard to PCBs printed in China. Travers’ line: it’s a great time to be a maker, and 3D modelling and printing push it further. Brett is now wondering whether Claude could write better software for his HOTAS controller, which he regrets buying because he crashes constantly and the bundled software looks like it’s from the 80s.

• Reflashing firmware (46:23)

  Paul says reflashing firmware on hardware you find (Brett: don’t pick up hardware off the street) is within what Claude can do. He mentioned JTAG debugging, where people “accidentally” copy the ROM out, re-engineer it and send it back. Nobody meant for the model to do that.

Off-the-Clock Recommendations

• Buddy: Travers’ absurdist horror pick, a homicidal take on Barney where an entity traps kids in a TV-show world. Thumbs up.
• Mandy (2018): Brett’s Nicolas Cage pick. He compared it to John Wick with strange undertones, and loved the “that’s my favorite t-shirt” line. Travers says it has one of the best fake commercials ever, the Cheddar Goblin, which Brett doesn’t remember and now has to rewatch for.
• Victorian Psycho: Paul’s pick based on the trailer, which he says is a capital B on B movie.
• Rubber (2010): the killer tire movie, one of the late-night picks Brett’s Prime algorithm keeps feeding him.
• Letterboxd: still Brett’s watch list.