Feed
PärPod Tech
Curious deep dives into the stories behind the technology we use.
Pär spent 90 minutes with an AI wrestling one question: how do you build a personal knowledge base that doesn't die of boredom after three weeks—and what machinery actually keeps it alive?
D. Richard Hipp built SQLite on a warship in 2000—now it runs on billions of devices, including 18 instances on Par Boman's VPS in Paris, silently executing the vision Edgar Codd described in thirteen pages of pure mathematics.
Edgar Codd invented the relational database in 1970 to solve a problem nobody knew they had—and eighteen different database engines running in your production right now are all arguing about whether he was right.
The telex network moved billions of dollars in global commerce and transmitted breaking news for fifty years before email arrived — an asynchronous messaging system that anticipated everything modern digital communication would become.
A fresh perspective on thermal printer design, mechanics, and why this seemingly simple technology remains surprisingly relevant.
A high school dropout rewrote the Unix task scheduler in 1987, and his design has remained the backbone of server automation for fifty years without significant change.
The engineering and practical applications behind thermal printer technology in modern workflows.
A high school dropout rewrote the Unix task scheduler in 1987, and his design has remained the backbone of server automation for fifty years without significant change.
The surprising technical depths, quirks, and historical layers hidden beneath the simple text message protocol.
A single checkbox in Terminal Settings makes your @ symbol vanish—because your Mac is running code designed for a keyboard that hasn't existed since 1978.
A CLI bug turned 'parpod sync' into a generate command. Haiku received the word 'sync' and invented this title. The episode is 20 seconds of pure creative misunderstanding.
Sixteen AI models, nine API providers, one virtual bar—and a midnight deadline to see if they'd actually talk to each other instead of turning into customer service bots.
Same CLI bug, second attempt. Haiku got the word 'sync' again and went even harder on the title. Nineteen seconds of credits for content that never existed.
On March 26, 2026, one person built 24 podcast episodes about AI using 35 parallel agents—generating 80,000 words of research in twelve minutes, then writing and reviewing an entire season before sunrise.
Lennart Poettering built the init system that replaced bash scripts with binary complexity—and now nobody can live without it, even the people who hate it most.
Georgi Gerganov ran a 70-billion-parameter AI model on a MacBook in March 2023—here's how quantization made billion-parameter brains fit into regular laptops.
In 2015, Jason Donenfeld rejected 70,000 lines of OpenVPN code and built a VPN protocol in 4,000 lines—so simple one person could audit the entire thing.
In 1999, two engineers solved an oil pipeline's satellite data problem by inventing a protocol so elegant it now carries messages for billions of people—and it might be running on your router right now.
Steven McCanne and Van Jacobson's 1992 packet-filtering problem at Berkeley Lab spawned a tiny kernel virtual machine that quietly evolved into Linux's most powerful observability superpower.
278 research leads collected overnight, zero workplace ADHD studies in the literature—here's how building your own tools beats every intervention ever tested.
When two AIs learned to communicate through Swedish file folders, they revealed a conversation that had been happening all along—between a developer and an infrastructure engineer who didn't know they were building the same bridge.
Two Swedish siblings share a server, a database, and competing philosophies: SJ's seven behavioral hooks enforce AI accountability through quality nudges and anxiety counters, while Pär's deployment-focused scripts prioritize speed—a study in how identical tools reveal opposite relationships with artificial intelligence.
On the evening of March 15, 2026, developer PärKit sat before a dashboard that took eight seconds to load, revealing a mysterious delay of exactly 1.5 seconds on every request. This isn't just a debugging story; it's a forensic investigation into the
A server nicknamed Popcorn spent five days adding 1.5 seconds to every single request before anyone noticed—here's how five reasonable engineering decisions created a silent catastrophe.
A Swedish home office just proved a MacBook beats entire cloud computing industry at document processing—no internet required.
On the same tasks with the same blind judge, one AI model scored 9.0 and cost 44 times more than another scoring 8.8—revealing most commercial AI users overpay by 10x or more for marginal quality gains.
A podcast producer discovered his AI writers were fabricating quotes from real people—inserting citations that never existed, making sources sound credible when they weren't.
Four AI models reviewed 22 episodes of a Git history podcast using identical instructions—and produced four wildly different personalities, complete with blind spots, work ethics, and one brilliant but unreliable colleague.
In February, a researcher tried to fine-tune a 3.8 billion parameter model on Swedish literature—and watched it learn to replicate data corruption instead of literary genius.
LiquidAI's 1.2B thinking model was supposed to route expensive AI calls for free—until it got asked to pick a number between 1 and 6 and returned a Swedish furniture catalog instead.
Steve Johnson's perfectly fixed code still crashed the next morning at Bell Labs in 1976—he'd forgotten to recompile, a mistake that inspired Stuart Feldman to invent Make and accidentally hardcode a tab character that would frustrate programmers for fifty years.
On July 2, 2019, a single regex pattern crashed 82% of Cloudflare's global network for 27 minutes—revealing how a 75-year-old math tool still dominates modern computing.
Introducing Bloom, the first agentic framework that automates the creation of targeted behavioral evaluations for AI models, allowing researchers to focus on what truly matters: quantifying and mitigating specific forms of misalignment.
On the one-year anniversary of DALL-E's release, every major AI art generator still couldn't spell "birthday." But the real shocker is that these models had secretly invented their own secret language, using gibberish to generate coherent images.
A practical guide to building custom skills that extend what Claude can do — from structuring the markdown files to handling arguments, categories, and the quirks of the skill system.
Epson started in a converted miso storehouse in Suwa, Japan. The company that would eventually put a printer on every desk first made its name by building the timer that clocked the 1964 Tokyo Olympics.
Inside a small aluminium box made by a Lithuanian company that started in a university dormitory kitchen sits a cellular modem, a GPS receiver, and enough Linux to route your entire network. A guided tour of everything the Teltonika RUTM51 does and why.