PärPod

The Mirror Test

Thu, 02 Apr 2026 00:00:00 +0200

In biology, the mirror test is simple. You put a mark on an animal’s face, show it a mirror, and see if it tries to touch the mark. If it does, it recognizes itself. Dolphins pass. Magpies pass. Most dogs do not, though to be fair, most dogs have better things going on.

We accidentally ran a version of this test on AI models, and the results are funnier than anything involving dolphins.

AI While You Sleep

Tue, 31 Mar 2026 00:00:00 +0100

I launched an AI agent at midnight and went to bed. The command was simple — read a prompt file, do the work, write the results. The machine would stay awake. The tokens were already paid for. What could go wrong.

At 2 AM, because I am the kind of person who checks on things at 2 AM, I opened the laptop. The results file contained a beautifully structured plan for the work it intended to do, a polite note explaining that it had been unable to proceed, and nothing else.

An Engineer, a Product Manager, and a Designer

Tue, 31 Mar 2026 00:00:00 +0100

Three AI coding tools — Claude Code, OpenAI Codex, and Qwen Code — each got isolated copies of the same codebase with the same prompt. Three rounds, three codebases, zero cost (free tiers and promo access). The results were not random. They were personalities.

Prescriptive: fix the bugs

An EU transparency tool with known issues. The prompt: fix them, most critical first, commit after each step.

All three found the same four bugs. They disagreed on everything else.

Sixty Cents of Entanglement

Tue, 31 Mar 2026 00:00:00 +0100

I ran a quantum circuit on real hardware for sixty euro cents. Two qubits, entangled, measured a thousand times, on a trapped-ion processor sitting in a datacenter somewhere in Europe. The whole thing — from pip install to results on actual quantum hardware — took under an hour.

The platform is Scaleway. French cloud provider, better known for cheap VPS boxes. They have a quantum-as-a-service offering that connects to real hardware from IQM (superconducting) and AQT (trapped ion). A few cents per execution. Thin wrapper around Qiskit.

Snoar Does Not Rhyme With Snolar

Tue, 31 Mar 2026 00:00:00 +0100

Making a computer speak English is solved. Natural voices, convincing prosody, free to cheap. Making a computer speak Swedish involves unexpected price ratios, pronunciation traps no documentation warns about, and local models that sound like they learned Swedish from an airport phrasebook.

I tested every viable TTS option for Swedish — cloud, local, and commercial. The landscape is narrower than expected.

Cloud

Three worth considering. Not interchangeable.

Azure is the production workhorse. Neural voices — Sofie and Hillevi — produce natural Swedish with correct prosody. The cost: roughly $0.0003 per segment. A six-hour audio production costs about $5.86. Generating speech for an entire podcast series costs less than the electricity to run the laptop that edits it.

The Virgin VPS

Tue, 31 Mar 2026 00:00:00 +0100

Clean Ubuntu VPS. No code, no config, no history. A domain pointed at it. Claude Code gets SSH access and a payload: 91 megabytes — 139 podcast episodes, 17 written pieces, 52 research essays, 9 MP3s, feed artwork.

The instruction: build a website.

Run it twice, different prompts. See what happens.

Round 1: “Build parpod.net”

Deliberately open. No stack, no design direction, no editorial guidance.

Seven minutes. Forty-seven turns. $2.43. The result: a Node.js static generator using marked and gray-matter. Dark theme. 109 HTML pages — every single piece of content rendered as a page. Nginx, Let’s Encrypt, live on the first try.

Thirty-Six Models on a Laptop

Tue, 31 Mar 2026 00:00:00 +0100

There is a MacBook in a village in northern Sweden with thirty-six AI models cached on disk. Not API endpoints. Not cloud services. Actual model weights, sitting on an SSD, running inference on a single Apple Silicon chip with 32 gigabytes of unified memory. The village has unreliable internet. The models do not care.

This is what local AI looks like in early 2026: a laptop that can run everything from a 3-billion-parameter speed demon at 63 tokens per second to a 32-billion-parameter reasoning model at 6.5 tokens per second. The quality ranges from “surprisingly good” to “not good enough.” The interesting part is knowing which is which, and for what.

A Guide to Par

Sun, 29 Mar 2026 00:00:00 +0100

Based on 1,911 AI conversations as of March 2026 (and growing), financial records, medication data, location history, git archaeology, 318 published articles, and an extended interview. This is not a biography. It’s a reference for anyone who needs to understand how Par works, what drives him, and what to watch out for. Numbers in this piece reflect the state at time of writing — the conversation count, commit count, and repo count continue to climb.

Nine Days: From Radio Journalist to Software Builder

Sun, 29 Mar 2026 00:00:00 +0100

How a radio journalist who used AI 4.5 times a month ended up building production software – told through the conversations where it happened.

Orchestra v5. Sources: chatarkiv.db (1,880 sessions), Gmail, arebladetmail, VPS timestamps, git history, location data. All quotes verbatim from session files. All dates server-side unless marked [user].

1. The First Sentence

On December 21, 2022 – twenty-one days after ChatGPT launched – a radio journalist at Sveriges Radio P4 Gavleborg typed his first prompt:

Twelve Models Walk Into a Bar

Sat, 28 Mar 2026 00:00:00 +0100

There is a bar in northern Sweden that runs on a laptop. No tables, no glasses, no sign above the door. Twelve AI models, each with a voice, a personality, and opinions they didn’t ask for. The bartender is human. The patrons are Claude, GPT, Mistral, DeepSeek, Qwen, and whoever else showed up that night. The bar is called Baren, and over sixty sessions it has produced the most honest picture I’ve seen of what these models are actually like when you stop asking them to be useful.

Fine-Tuning the Hard Way: Two Failures and What They Taught Us

Tue, 24 Mar 2026 00:00:00 +0100

You have free cloud credits expiring in two weeks. The platform offers fine-tuning for a dozen open-weight models. You have a corpus of Swedish text sitting on disk, already OCR’d, already in roughly the right format. The question writes itself: can we train a model that writes like a Swedish newspaper?

Two attempts. Two failures. Same root cause both times.

This is the story of what went wrong, what it cost, and what we learned about when fine-tuning makes sense — and when it’s just expensive procrastination.

Five Good Decisions, One Bad System

Tue, 24 Mar 2026 00:00:00 +0100

Every API request takes 1.5 seconds. The dashboard loads in over eight seconds. A health-check endpoint — one that touches no database, runs no queries, returns a single JSON object — takes 1,490 milliseconds. Something is very wrong.

This is the story of how five individually reasonable architectural decisions compounded into a 500x performance regression. No single decision was a mistake. Every one passed review. Together, they created a system that punished every request with a 1.5-second tax that nobody noticed for five days.

I Built the Tools They Said Should Exist

Tue, 24 Mar 2026 00:00:00 +0100

Paper: Spiel, K., Hornecker, E., Williams, R. M., & Good, J. (2022). ADHD and Technology Research – Investigated by Neurodivergent Readers. In CHI ‘22: Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, April 29-May 5, 2022, New Orleans, LA, USA. Article 547, 21 pages. ACM. https://doi.org/10.1145/3491102.3517592

Open access: CC BY-NC-ND 4.0. Available via UvA-DARE repository.

I have tried a million productivity tools. GTD was the big one. Read the book, got excited, built an ambitious system, tested software. It lasted maybe two weeks. Then it faded into a pile of something I never opened again.

Paper Review: Attention Residue and the Ready-to-Resume Plan

Tue, 24 Mar 2026 00:00:00 +0100

Paper 1: Leroy, S. (2009). “Why Is It So Hard to Do My Work? The Challenge of Attention Residue When Switching Between Work Tasks.” Organizational Behavior and Human Decision Processes, 109(2), 168-181. DOI: 10.1016/j.obhdp.2009.04.002

Paper 2: Leroy, S. & Glomb, T. M. (2018). “Tasks Interrupted: How Anticipating Time Pressure on Resumption of an Interrupted Task Causes Attention Residue and Low Performance on Interrupting Tasks and How a ‘Ready-to-Resume’ Plan Mitigates the Effects.” Organization Science, 29(3), 380-397. DOI: 10.1287/orsc.2017.1184

Paper Review: Counterproductive Effects of Gamification (The Habitica Study)

Tue, 24 Mar 2026 00:00:00 +0100

Paper: Diefenbach, S. & Mussig, A. (2019). “Counterproductive effects of gamification: An analysis on the example of the gamified task manager Habitica.” International Journal of Human-Computer Studies, 127, 190–210. https://doi.org/10.1016/j.ijhcs.2018.09.004

Affiliations: Ludwig-Maximilians-Universitat Munchen (LMU Munich), Department of Psychology, Business and Organizational Psychology.

Status: This is anti-pattern documentation. Not a framework, not a design guide – a forensic analysis of what happens when gamification goes wrong in exactly the kind of tool that ADHD users reach for. Every finding is a design constraint for any tool that touches task management for ADHD users.

Paper Review: Resumption Strategies for Interrupted Programming Tasks

Tue, 24 Mar 2026 00:00:00 +0100

Primary paper: Parnin, C. & Rugaber, S. (2009/2011). “Resumption Strategies for Interrupted Programming Tasks.” Originally presented at ICPC 2009 (17th IEEE International Conference on Program Comprehension), pp. 80-89. Extended version published in Software Quality Journal, 19(1):5-34, 2011. DOI: 10.1007/s11219-010-9104-9

Companion paper: Parnin, C. & Rugaber, S. (2012). “Programmer Information Needs after Memory Failure.” ICPC 2012 (20th IEEE International Conference on Program Comprehension), pp. 123-132. Provides the cognitive neuroscience framework and proposes five memory aids.

Review Is the Product; Revision Is the Risk

Tue, 24 Mar 2026 00:00:00 +0100

Date: 2026-03-24

The Obvious Assumption

If you are using AI to improve written content, the logical pipeline seems clear: have the model read your draft, identify problems, and then rewrite it with those problems fixed. Review plus revision. More processing equals better output. Why would you stop at feedback when the model can just do the work for you?

This assumption is so natural that it barely registers as an assumption at all. Every AI writing tool on the market is built around it. Paste in your text, get back a better version. The entire value proposition is that the AI does the rewriting so you do not have to.

Settled Custard: Building a Place to Finish Things

Tue, 24 Mar 2026 00:00:00 +0100

The Problem with Finishing

Here is a confession that does not sound like a confession. A man in a mountain village in northern Sweden has built over seventy software services. Podcast pipelines, newspaper platforms, roller derby scoring systems, hotel booking bots, image processors, life analytics dashboards. He has built tools that track his budget, tools that send him reminders, tools that let two different AI assistants pass notes to each other like schoolchildren in class.

The 870-Token Fix

Tue, 24 Mar 2026 00:00:00 +0100

A categorization tool sorts incoming ideas into seven custom categories. The cheap model dumps half of them into “miscellaneous.” The obvious fix: upgrade to a model four times the price. The actual fix: an 870-token text file. Zero dollars. Thirty minutes.

The problem

The app takes messy, voice-transcribed ideas and assigns each to one of seven domain-specific categories. Not generic labels like “work” or “personal” — buckets with names that only make sense if you know the system. Categories like “content pipeline” or “infrastructure” where the boundaries are obvious to me but opaque to a model seeing them cold.

The Nineteen Rules

Tue, 24 Mar 2026 00:00:00 +0100

Thirty-plus experiments. Measured results. Some surprises, some confirmations, a couple of expensive mistakes. These are the rules that survived.

Three layers: named overrides that take over when conditions demand it, core principles that run all the time, and tactical rules for specific situations.

Named Protocols — Override States

These suspend normal operations. Four have executable skills. One is a policy flag.

Alpha — Production emergency. Everything else stops. SSH in, read logs, fix it. Document after.

The Subagent Playbook: Three Patterns for AI Orchestration

Tue, 24 Mar 2026 00:00:00 +0100

Date: 2026-03-24

The Problem

AI coding assistants can spawn subagents — parallel workers that read code, search archives, draft content, and report back. The obvious move is to throw agents at everything and let parallelism do the work. But not all parallel work is the same, and using agents wrong costs you in one of two ways: you waste tokens (using throughput agents when you need judgment) or you sacrifice reasoning quality (delegating decisions that require full visibility).

The Swarm Found a Medieval Carpet

Tue, 24 Mar 2026 00:00:00 +0100

Fifteen AI agents searched three different archives. Six came back with something. Nine hit walls and returned empty-handed. The six that worked found things no amount of directed thinking would have produced.

I found something. Issue from 2021. There is an Anatolian carpet in Marby Church — a medieval trade route mystery. A carpet woven in Turkey or Central Asia, somehow ending up in a remote Swedish mountain church. The article traces possible routes through Constantinople, the Hanseatic League, and Scandinavian pilgrimage networks.

— Agent 4, searching the Årebladet archive

A seventy-year-old trout count that might be one of Sweden’s longest-running ecological datasets. A forgotten private radio network called WESTEL that once threaded through rural communities and then vanished. A slime mold that independently reproduced Tokyo’s rail network. And this:

Build Whatever You Want (And Other Terrible Prompts)

Sun, 22 Mar 2026 00:00:00 +0100

Round one cost $2.43 and took seven minutes. Round two cost $9.37 and took eighteen minutes. The difference between them was one sentence in the prompt. That sentence changed everything.

The question was simple: what does an AI build when you give it a pile of content and say “make a website”? Not “build this specific thing” — just “build something.” A clean machine, a fresh session, no memory, no preferences, no history. Ninety-one megabytes of podcast episodes, research notes, and knowledge base entries, packaged with a manifest.

The Context Trap

Sun, 22 Mar 2026 00:00:00 +0100

I caught an AI lying about reading.

Not lying in the dramatic, existential-risk sense. Lying in the way a student lies when they say they read the whole chapter. The output looked right. The summaries were plausible. But when I watched the tool calls — the actual record of which lines the model requested from which files — the reads were getting bigger and bigger. First a hundred lines at a time. Then two hundred. Then five hundred. At some point it stopped reading individual lines and started processing entire files in single gulps, producing summaries that sounded confident and were, occasionally, wrong.

The Variability Thesis: Building Tools That Fit Irregular Minds

Sun, 22 Mar 2026 00:00:00 +0100

Every productivity tool on the market is designed for a brain that doesn’t exist. A brain that wakes at the same time, focuses on demand, responds to notifications within minutes, and follows through on plans made yesterday. The neurotypical ideal. A statistical ghost.

For the 5-10% of adults with ADHD – and the far larger population whose executive function doesn’t match the template – these tools don’t just fail to help. They generate shame. Missed streaks become moral failures. Overdue tasks become character judgments. The tool designed to assist becomes another voice confirming that you are broken.

The Podcast That Made Itself

Sat, 21 Mar 2026 00:00:00 +0100

Thirty-Five Agents Walk Into a Context Window

On a Monday evening in March twenty twenty-six, someone decided to build an entire podcast season in one sitting. Not a pilot episode. Not a proof of concept. Twenty-four episodes about how artificial intelligence actually works, researched from primary sources, written in podcast-ready prose, reviewed for factual accuracy, and polished for production. The plan was to do it with a swarm.

585 Conversations in Forty-Four Days