<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:dc="http://purl.org/dc/elements/1.1/"><channel><title>T KOLSTO · Blog</title><link>https://t.kolsto.no/blog/</link><description>Posts on technology, infrastructure, and startups by Thomas Kolstø.</description><generator>Hugo</generator><language>en</language><copyright>© 2010 - 2026 Thomas Kolstø</copyright><lastBuildDate>Sun, 04 Oct 2026 05:15:00 +0000</lastBuildDate><atom:link href="https://t.kolsto.no/blog/" rel="self" type="application/rss+xml"/><image><url>https://t.kolsto.no/img/avatar.jpg</url><title>T KOLSTO · Blog</title><link>https://t.kolsto.no/blog/</link></image><item><title>Eight Arms, All Billable</title><link>https://t.kolsto.no/blog/clocktopus/</link><pubDate>Sun, 04 Oct 2026 05:15:00 +0000</pubDate><guid>https://t.kolsto.no/blog/clocktopus/</guid><dc:creator>Thomas Kolstø</dc:creator><category>clocktopus</category><category>macos</category><category>swift</category><category>sidequest</category><category>time-tracking</category><description>Clocktopus is a local macOS time tracker that suggests work sessions and asks before recording them. No screenshots, no cloud. Suggest, never surveil.</description><enclosure url="https://t.kolsto.no/img/clocktopus-share-v1.png" length="51601" type="image/png"/><content:encoded>&lt;p&gt;I used to perform a ritual every Friday that I suspect you&amp;rsquo;ll recognize. Timesheets were due in xledger—hours split across projects and activities because my time belonged to multiple budgets. I would fill them out with the confidence of someone who had been present all week, only to realize I couldn&amp;rsquo;t recall what I&amp;rsquo;d actually done.&lt;/p&gt;&#10;&lt;p&gt;So, I reconstructed it: running &lt;code&gt;git log --since=monday&lt;/code&gt; across repos, checking calendars for meetings, and scrolling Slack history to account for missing time. Then I translated this archaeology into tidy quarter-hour blocks and submitted it as fact.&lt;/p&gt;&#10;&lt;p&gt;Memory is the worst database. It lacks a write-ahead log, runs compaction nightly whether you like it or not, and every query returns a plausible answer with no indication of its truthfulness.&lt;/p&gt;&#10;&lt;h2 id="the-tools-are-all-wrong"&gt;The tools are all wrong&lt;/h2&gt;&#10;&lt;p&gt;The obvious fix is a timer: start when working, stop when done, making Friday an export button. But manual trackers fail because they require &lt;em&gt;remembering&lt;/em&gt; to use them. I don&amp;rsquo;t forget how to work; I forget to say that I am working. Any tool relying on me forming a new habit at the precise moment my brain is busy with something else is designed to fail. My collection of abandoned timewarrior databases proves this.&lt;/p&gt;&#10;&lt;p&gt;At the other end are &amp;ldquo;automatic&amp;rdquo; trackers. They solve forgetting by removing you from the loop entirely. They screenshot your screen every thirty seconds, run it through AI, and upload &amp;ldquo;productivity insights&amp;rdquo; to a cloud dashboard—often read by your manager. I work with patient data for a telemedicine company. A tool that photographs my display and posts it to someone else&amp;rsquo;s cloud isn&amp;rsquo;t an aid; it&amp;rsquo;s an incident report with a subscription fee.&lt;/p&gt;&#10;&lt;p&gt;This leaves a gap: one end trusts you to remember, the other doesn&amp;rsquo;t trust you at all.&lt;/p&gt;&#10;&lt;h2 id="the-sidequest"&gt;The sidequest&lt;/h2&gt;&#10;&lt;p&gt;So, obviously, I built one. This is a proper sidequest: no cloud, no accounts, no roadmap committee, and no ambitions of becoming a company. Just the tracker I wanted, built how any tool touching my work machine should be.&lt;/p&gt;&#10;&lt;p&gt;It&amp;rsquo;s called &lt;a href="https://github.com/tkolsto/clocktopus"&gt;Clocktopus&lt;/a&gt;. Eight arms, all billable.&lt;/p&gt;&#10;&lt;img src="https://t.kolsto.no/img/clocktopus-popover.webp" alt="Clocktopus menubar popover with a running timer, project list, and weekly activity chart" width="600" height="764" loading="lazy"&gt;&#10;&lt;p&gt;The insight is embarrassingly simple: your machine already knows what you&amp;rsquo;re working on. Not via AI—but via &lt;code&gt;pwd&lt;/code&gt;. My shells are cd&amp;rsquo;d into a client&amp;rsquo;s repo. My tmux panes sit in project directories. A running process has a working directory that names the project outright. The frontmost window title shows the open codebase. None of this needs a vision model; it needs a scorer over deterministic signals and a config file where &lt;em&gt;you&lt;/em&gt; write the rules:&lt;/p&gt;&#10;&lt;div class="tk-code"&gt;&lt;div class="tk-code__tools"&gt;&lt;span class="tk-code__lang"&gt;toml&lt;/span&gt;&lt;/div&gt;&lt;div class="highlight"&gt;&lt;pre tabindex="0" class="chroma"&gt;&lt;code class="language-toml" data-lang="toml"&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt;&lt;span class="p"&gt;[[&lt;/span&gt;&lt;span class="nx"&gt;project&lt;/span&gt;&lt;span class="p"&gt;]]&lt;/span&gt;&#10;&lt;/span&gt;&lt;/span&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt;&lt;span class="nx"&gt;name&lt;/span&gt; &lt;span class="p"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;&amp;#34;Initech&amp;#34;&lt;/span&gt;&#10;&lt;/span&gt;&lt;/span&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt;&lt;span class="nx"&gt;xledger_project&lt;/span&gt; &lt;span class="p"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;&amp;#34;T200001&amp;#34;&lt;/span&gt;&#10;&lt;/span&gt;&lt;/span&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt;&lt;span class="nx"&gt;xledger_activity&lt;/span&gt; &lt;span class="p"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;&amp;#34;DEV&amp;#34;&lt;/span&gt;&#10;&lt;/span&gt;&lt;/span&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt;&lt;span class="nx"&gt;dirs&lt;/span&gt; &lt;span class="p"&gt;=&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;&amp;#34;~/src/initech*&amp;#34;&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="c"&gt;# shell / tmux / AI-agent cwds&lt;/span&gt;&#10;&lt;/span&gt;&lt;/span&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt;&lt;span class="nx"&gt;keywords&lt;/span&gt; &lt;span class="p"&gt;=&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;&amp;#34;initech&amp;#34;&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="c"&gt;# frontmost window title&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&#10;&lt;/div&gt;&#10;&lt;p&gt;Longest prefix wins, so &lt;code&gt;~/src/initech-api&lt;/code&gt; beats your &lt;code&gt;~/src&lt;/code&gt; catch-all. Every rule is yours, in a version-controllable TOML file, and every guess shows exactly which signals fired. No vibes.&lt;/p&gt;&#10;&lt;h2 id="suggest-never-surveil"&gt;Suggest, never surveil&lt;/h2&gt;&#10;&lt;p&gt;This is the core feature. When a project holds my focus for minutes, Clocktopus doesn&amp;rsquo;t clock me in. It &lt;em&gt;asks&lt;/em&gt;. A dashed ghost block appears on the day timeline—&amp;ldquo;is this Initech?&amp;quot;—with evidence attached. One click confirms it, backfilling to when I started (the whole point: the app remembers the start time I was too busy to announce). One click dismisses it. Nothing lands in the database without my say-so.&lt;/p&gt;&#10;&lt;img src="https://t.kolsto.no/img/clocktopus-day-timeline.webp" alt="Clocktopus day timeline with a dashed ghost block suggesting a detected work session" width="1800" height="1800" srcset="https://t.kolsto.no/img/clocktopus-day-timeline_hu_6a68ec9bc18832fe.webp 720w, https://t.kolsto.no/img/clocktopus-day-timeline_hu_228a98bf6b529b21.webp 1440w, https://t.kolsto.no/img/clocktopus-day-timeline.webp 1800w" sizes="(min-width: 1120px) 1040px, (min-width: 601px) calc(100vw - 80px), calc(100vw - 40px)" loading="eager" fetchpriority="high"&gt;&#10;&lt;p&gt;The surveillance question is settled by architecture, not policy: the app makes zero network requests. Not for sync, telemetry, or updates—grep the source. Your hours are one SQLite file in &lt;code&gt;~/Library/Application Support&lt;/code&gt;, and the only thing leaving the machine is the CSV you export yourself.&lt;/p&gt;&#10;&lt;h2 id="friday-after"&gt;Friday, after&lt;/h2&gt;&#10;&lt;p&gt;The ritual now: open the week grid (ISO weeks, Monday start), with per-project totals already rounded to the billing increment. I skim it, nudge a block edge if the octopus and I disagree about lunch, and export. The CSV speaks xledger&amp;rsquo;s upload dialect—semicolon-separated, &lt;code&gt;yyyymmdd&lt;/code&gt; dates, decimal hours—because that&amp;rsquo;s what my Fridays demanded, but it&amp;rsquo;s just CSV; point it at whatever your accounting department worships.&lt;/p&gt;&#10;&lt;img src="https://t.kolsto.no/img/clocktopus-week-grid.webp" alt="Clocktopus week review with per-project hours, separate billable and personal totals, and CSV export" width="1800" height="1380" srcset="https://t.kolsto.no/img/clocktopus-week-grid_hu_e107ec88c1b571e0.webp 720w, https://t.kolsto.no/img/clocktopus-week-grid_hu_2e807c2023dde021.webp 1440w, https://t.kolsto.no/img/clocktopus-week-grid.webp 1800w" sizes="(min-width: 1120px) 1040px, (min-width: 601px) calc(100vw - 80px), calc(100vw - 40px)" loading="lazy"&gt;&#10;&lt;p&gt;The archaeology is gone. That&amp;rsquo;s the entire return on investment, and it pays out weekly.&lt;/p&gt;&#10;&lt;h2 id="questions-youre-already-typing"&gt;Questions you&amp;rsquo;re already typing&lt;/h2&gt;&#10;&lt;p&gt;&lt;strong&gt;&amp;ldquo;Why not ActivityWatch?&amp;rdquo;&lt;/strong&gt; A genuinely good project for passive analytics—it lets you study yourself afterwards. Clocktopus is suggest-and-confirm for &lt;em&gt;billing&lt;/em&gt;: nothing is recorded until you approve it, and the output is a timesheet, not a chart. And timewarrior is fine, but it still trusts me to remember. We&amp;rsquo;ve established how that goes.&lt;/p&gt;&#10;&lt;p&gt;&lt;strong&gt;&amp;ldquo;Linux port?&amp;rdquo;&lt;/strong&gt; The core—domain models, scoring, rounding, export, ~90 tests—is a pure SwiftPM package with no AppKit, so it&amp;rsquo;s portable-ish. Signal collectors are the macOS-specific part. If cwd-sniffing on Linux sounds interesting, the repo is MIT.&lt;/p&gt;&#10;&lt;p&gt;&lt;strong&gt;&amp;ldquo;Why isn&amp;rsquo;t it notarized?&amp;rdquo;&lt;/strong&gt; An Apple Developer ID is $99/year, and this octopus is young. The release build is ad-hoc signed: right-click to open the first time, or build from source (Gatekeeper has no opinions).&lt;/p&gt;&#10;&lt;p&gt;&lt;strong&gt;&amp;ldquo;Can a tmux pane I forgot about bill a client from three days ago?&amp;rdquo;&lt;/strong&gt; No. This was the heuristic I sweated most. Background signals—parked panes, long-running agent sessions—can only &lt;em&gt;reinforce&lt;/em&gt; a project that also has foreground evidence. A forgotten pane can&amp;rsquo;t clock you in on its own; it can only corroborate what you are demonstrably looking at.&lt;/p&gt;&#10;&lt;h2 id="coda"&gt;Coda&lt;/h2&gt;&#10;&lt;p&gt;Clocktopus is &lt;a href="https://github.com/tkolsto/clocktopus"&gt;on GitHub&lt;/a&gt;, MIT licensed, macOS 13+, Swift/SwiftUI, native menubar app. It has exactly one user in production, whose Fridays are much improved. If your accounting department also worships a CSV, maybe it&amp;rsquo;ll improve yours.&lt;/p&gt;&#10;</content:encoded></item><item><title>The Ash AI Tax</title><link>https://t.kolsto.no/blog/ash-ai-tax/</link><pubDate>Sat, 30 May 2026 00:00:00 +0700</pubDate><guid>https://t.kolsto.no/blog/ash-ai-tax/</guid><dc:creator>Thomas Kolstø</dc:creator><category>ash</category><category>elixir</category><category>AI</category><category>LLM</category><description>I kept noticing my AI agent burned more tokens on our Ash code than on plain Phoenix. So I measured it: ~2.6x more on Claude, ~1.9x on Codex — building the exact same feature.</description><enclosure url="https://t.kolsto.no/img/ash-ai-tax-chart.jpg" length="53026" type="image/jpeg"/><content:encoded>&lt;p&gt;For months I harbored a nagging suspicion I couldn&amp;rsquo;t prove. Every time I pointed a coding agent at our Ash codebase, it consumed tokens noticeably faster than when working on plain Phoenix. I initially dismissed this as mere confirmation bias and filed it away.&lt;/p&gt;&#10;&lt;p&gt;I finally tested this by building the same feature against the passing test suite in two stacks—idiomatic Ash and plain Phoenix + Ecto—and having an agent build it five times in each. The result was bigger than my hunch:&lt;/p&gt;&#10;&lt;blockquote&gt;&#10;&lt;p&gt;Building an identical feature, Claude (Opus 4.7) used &lt;strong&gt;+162% more tokens&lt;/strong&gt; in Ash than in plain Ecto. Codex (GPT-5.5) used &lt;strong&gt;+90% more&lt;/strong&gt;. This effect showed up on both agents.&lt;/p&gt;&#10;&lt;/blockquote&gt;&#10;&lt;img src="https://t.kolsto.no/img/ash-ai-tax-chart.jpg" alt="Bar chart: Claude used 1.76M tokens on Ash vs 672k on Ecto (&amp;#43;162%); Codex used 1.02M on Ash vs 533k on Ecto (&amp;#43;90%). 20 runs, all green." width="1200" height="630" srcset="https://t.kolsto.no/img/ash-ai-tax-chart_hu_6de352feda187a39.webp 720w, https://t.kolsto.no/img/ash-ai-tax-chart_hu_66e814372288085d.webp 1200w" sizes="(min-width: 1120px) 1040px, (min-width: 601px) calc(100vw - 80px), calc(100vw - 40px)" loading="eager" fetchpriority="high"&gt;&#10;&lt;p&gt;I&amp;rsquo;d have bet on 25%. It was closer to 2.6x.&lt;/p&gt;&#10;&lt;h2 id="why-this-is-worth-measuring"&gt;Why this is worth measuring&lt;/h2&gt;&#10;&lt;p&gt;We have crossed a line where most code is now written with an agent in the loop. Tokens are not free; they represent money, latency, and pressure on a finite context window. This means framework choice must account for a dimension nobody includes in comparison tables: &lt;strong&gt;how expensive is it for an AI to work in?&lt;/strong&gt;&lt;/p&gt;&#10;&lt;p&gt;This is an uncomfortable question because the answer may contradict everything that makes a framework pleasant for &lt;em&gt;humans&lt;/em&gt;. Ash is genuinely delightful for certain problems, but the core issue isn&amp;rsquo;t whether it&amp;rsquo;s good. It&amp;rsquo;s whether the feature that helps me costs the robot extra—and by how much.&lt;/p&gt;&#10;&lt;h2 id="the-experiment"&gt;The experiment&lt;/h2&gt;&#10;&lt;p&gt;I wanted a number I could defend, not a vibe. So I pinned everything I could.&lt;/p&gt;&#10;&lt;p&gt;The goal was building a small Support-ticket domain: a &lt;code&gt;Ticket&lt;/code&gt; with validations, a &lt;code&gt;Comment&lt;/code&gt; relationship, an assignee, custom &lt;code&gt;close&lt;/code&gt;/&lt;code&gt;reopen&lt;/code&gt; actions with state-transition guards, and authorization (assignee or admin can close; admin only can reopen). This provided enough surface area to exercise the parts of Ash people actually use, moving beyond simple CRUD.&lt;/p&gt;&#10;&lt;p&gt;The key was defining the contract in a plain Elixir module boundary (&lt;code&gt;Support.create_ticket/2&lt;/code&gt;, &lt;code&gt;Support.close_ticket/2&lt;/code&gt;, etc.) and writing one framework-agnostic ExUnit suite against it. This boundary is idiomatic for both sides—an Ash &lt;code&gt;code_interface&lt;/code&gt; and an Ecto context function expose the exact same six functions. Success meant the suite went green, the project compiled, and the linter was happy. No partial credit allowed.&lt;/p&gt;&#10;&lt;p&gt;The controls were strict:&lt;/p&gt;&#10;&lt;ul&gt;&#10;&lt;li&gt;Both stacks used the same agent, model, and prompt; only the framework differed.&lt;/li&gt;&#10;&lt;li&gt;Reasoning effort was pinned to &lt;code&gt;high&lt;/code&gt; for every run—identical across agents and frameworks. (Measuring whether the agent reasons more in Ash was a key goal, so this setting could not drift.)&lt;/li&gt;&#10;&lt;li&gt;The Ash side received its idiomatic context (&lt;code&gt;usage_rules&lt;/code&gt; AGENTS.md, all 1,556 lines), mimicking what a real Ash developer&amp;rsquo;s agent would get. The Ecto side required no special input as it is inherent to the model&amp;rsquo;s structure.&lt;/li&gt;&#10;&lt;li&gt;We ran five fresh checks per cell, and crucially, I personally verified green status after every run rather than relying on the agent&amp;rsquo;s confirmation.&lt;/li&gt;&#10;&lt;/ul&gt;&#10;&lt;p&gt;I repeated this process using two different agents—Claude Code and Codex—because &amp;ldquo;it replicates across models&amp;rdquo; is what separates a finding from an anecdote.&lt;/p&gt;&#10;&lt;p&gt;The entire harness—including the specification, shared suite, both reference implementations, the runner, and raw per-run token logs—is open source for reproduction or inspection:&#10;&lt;a href="https://github.com/tkolsto/ash-ai-tax-bench"&gt;github.com/tkolsto/ash-ai-tax-bench&lt;/a&gt;.&lt;/p&gt;&#10;&lt;h2 id="how-much-more-does-ash-cost-in-tokens"&gt;How much more does Ash cost in tokens?&lt;/h2&gt;&#10;&lt;p&gt;About 2× more — verified across 20 runs, every one reached green:&lt;/p&gt;&#10;&lt;section class="tk-table-scroll" aria-label="The Ash AI Tax — table 1" tabindex="0"&gt;&#10; &lt;table&gt;&#10; &lt;thead&gt;&#10; &lt;tr&gt;&#10; &lt;th scope="col"&gt;Agent&lt;/th&gt;&#10; &lt;th scope="col" style="text-align: right"&gt;Ecto (avg tokens)&lt;/th&gt;&#10; &lt;th scope="col" style="text-align: right"&gt;Ash (avg tokens)&lt;/th&gt;&#10; &lt;th scope="col" style="text-align: right"&gt;Ash tax&lt;/th&gt;&#10; &lt;th scope="col" style="text-align: center"&gt;Turns: Ecto → Ash&lt;/th&gt;&#10; &lt;/tr&gt;&#10; &lt;/thead&gt;&#10; &lt;tbody&gt;&#10; &lt;tr&gt;&#10; &lt;td&gt;Claude (Opus 4.7)&lt;/td&gt;&#10; &lt;td style="text-align: right"&gt;671,886&lt;/td&gt;&#10; &lt;td style="text-align: right"&gt;1,758,599&lt;/td&gt;&#10; &lt;td style="text-align: right"&gt;&lt;strong&gt;+162%&lt;/strong&gt;&lt;/td&gt;&#10; &lt;td style="text-align: center"&gt;22.4 → 31.0&lt;/td&gt;&#10; &lt;/tr&gt;&#10; &lt;tr&gt;&#10; &lt;td&gt;Codex (GPT-5.5)&lt;/td&gt;&#10; &lt;td style="text-align: right"&gt;533,001&lt;/td&gt;&#10; &lt;td style="text-align: right"&gt;1,015,218&lt;/td&gt;&#10; &lt;td style="text-align: right"&gt;&lt;strong&gt;+90%&lt;/strong&gt;&lt;/td&gt;&#10; &lt;td style="text-align: center"&gt;23.2 → 32.8&lt;/td&gt;&#10; &lt;/tr&gt;&#10; &lt;/tbody&gt;&#10; &lt;/table&gt;&#10;&lt;/section&gt;&#10;&lt;p&gt;A quick disclaimer: Per-run variance is high—Codex&amp;rsquo;s Ash runs, for example, ranged from 526k to 1.44M tokens—so treat these as rough estimates (&amp;ldquo;roughly 2x&amp;rdquo;) rather than precise percentages. However, the overall direction is rock solid: on neither agent did Ash come out cheaper. Furthermore, the deltas barely budged between my initial 3-run pilot and the full 5-run matrix (Claude went from 179% to 162%; Codex went from 85% to 90%), which is exactly what you want to see before trusting a number.&lt;/p&gt;&#10;&lt;p&gt;Both agents needed about 40% more &lt;em&gt;turns&lt;/em&gt; to get Ash green. This difference matters greatly, as it unravels the whole thing.&lt;/p&gt;&#10;&lt;h2 id="where-do-the-extra-tokens-actually-go"&gt;Where do the extra tokens actually go?&lt;/h2&gt;&#10;&lt;p&gt;The most interesting part is that it wasn&amp;rsquo;t what I expected. The issue isn&amp;rsquo;t that Ash code is bigger; the agents&amp;rsquo; diffs were comparable in size—sometimes even smaller. The cost isn&amp;rsquo;t in what the agent &lt;em&gt;writes&lt;/em&gt;, but in what it has to &lt;em&gt;understand&lt;/em&gt; before it can write anything correct.&lt;/p&gt;&#10;&lt;p&gt;I saw this concentration of effort in three places:&lt;/p&gt;&#10;&lt;p&gt;Context loading is a major hurdle. Frameworks like Ash require agents to pull in usage rules, resource definitions, and extensions for correct resources. Claude handles this poorly because it re-reads its cached context on every single turn. This means that a large starting context multiplied by many turns results in an exponential cost increase. That explains why Claude&amp;rsquo;s tax (+162%) significantly exceeds Codex&amp;rsquo;s (+90%): it is the same underlying phenomenon, just different token accounting.&lt;/p&gt;&#10;&lt;p&gt;&lt;strong&gt;Reasoning, not typing.&lt;/strong&gt; The agent spent significant effort thinking about Ash, determining which declarative knob produces the desired imperative behavior for a test. Ecto simply requires you to &amp;ldquo;write a query.&amp;rdquo; Ash, however, involves considering the action, change, validation, policy, and whether it will survive atomic execution.&lt;/p&gt;&#10;&lt;p&gt;&lt;strong&gt;The &amp;ldquo;that&amp;rsquo;s not how Ash works&amp;rdquo; tax.&lt;/strong&gt; This is the retry loop, and I have receipts. Watching the reference build, the friction was rarely domain logic; it was bending the framework to specific contracts:&lt;/p&gt;&#10;&lt;ul&gt;&#10;&lt;li&gt;A &lt;code&gt;code_interface&lt;/code&gt; named &lt;code&gt;create_ticket&lt;/code&gt; &lt;em&gt;collides&lt;/em&gt; with a hand-written &lt;code&gt;create_ticket/2&lt;/code&gt;. The solution involves defining internal interfaces and wrapping them.&lt;/li&gt;&#10;&lt;li&gt;The suite expects &lt;code&gt;get_ticket&lt;/code&gt; to return &lt;code&gt;{:error, :not_found}&lt;/code&gt;, but Ash returns a NotFound wrapped inside an &lt;code&gt;Ash.Error.Invalid&lt;/code&gt;, necessitating an error-normalization shim.&lt;/li&gt;&#10;&lt;li&gt;Reading the current &lt;code&gt;status&lt;/code&gt; to guard a transition triggers Ash 3&amp;rsquo;s &lt;code&gt;default_actions_require_atomic? true&lt;/code&gt;. The fix (&lt;code&gt;require_atomic? false&lt;/code&gt; plus a guard validation) is detailed in usage rules, but ignoring this results in baffling atomic-validation errors and recovery round trips.&lt;/li&gt;&#10;&lt;li&gt;Policies: Consecutive &lt;code&gt;authorize_if&lt;/code&gt; statements are &lt;strong&gt;OR&lt;/strong&gt;, not AND. Getting this mental model wrong causes authorization tests to fail in ways that look like logic bugs.&lt;/li&gt;&#10;&lt;/ul&gt;&#10;&lt;p&gt;None of this is hard &lt;em&gt;once you know it&lt;/em&gt;. That&amp;rsquo;s the point: it&amp;rsquo;s a large surface of structured knowledge, and every gap in the agent&amp;rsquo;s understanding costs tokens.&lt;/p&gt;&#10;&lt;h2 id="so-is-it-worth-it"&gt;So is it worth it?&lt;/h2&gt;&#10;&lt;p&gt;I need to be fair here. The tension is this: for a human, Ash often means writing less code; for an AI, it means understanding more before writing anything. These pull in opposite directions, and which one dominates depends on who—or what—is at the keyboard.&lt;/p&gt;&#10;&lt;p&gt;The tax buys real value: declarative actions, policies living next to their data, generated APIs, and a mountain of correctness you don&amp;rsquo;t have to hand-roll or maintain. On a long-lived system with a team, that leverage is worth far more than a pile of tokens. Furthermore, this tax is &lt;em&gt;front-loaded comprehension&lt;/em&gt;, not ongoing bloat—it’s the cost of the agent learning the framework&amp;rsquo;s shape, not a per-line penalty forever.&lt;/p&gt;&#10;&lt;p&gt;However, if you are spinning up something small, leaning hard on agents for speed, or watching your token bill, &amp;ldquo;2x the tokens, every feature&amp;rdquo; is a number that belongs in the decision, right next to all the reasons Ash is lovely.&lt;/p&gt;&#10;&lt;h2 id="how-to-shrink-the-ash-ai-tax"&gt;How to shrink the Ash AI tax&lt;/h2&gt;&#10;&lt;p&gt;The most expensive element is comprehension, but it&amp;rsquo;s also the easiest to fix: provide the agent with understanding rather than making it pay to rediscover it.&lt;/p&gt;&#10;&lt;ul&gt;&#10;&lt;li&gt;&lt;strong&gt;Invest in &lt;code&gt;usage_rules&lt;/code&gt;.&lt;/strong&gt; This is your highest-leverage move. The entire cost stems from the agent learning Ash&amp;rsquo;s structure; a current &lt;code&gt;usage_rules&lt;/code&gt; file preloads that shape. Keep this synced (Warning: the &lt;code&gt;usage_rules.sync&lt;/code&gt; tooling itself has sharp edges, costing me several iterations during testing).&lt;/li&gt;&#10;&lt;li&gt;&lt;strong&gt;Use &lt;code&gt;ash_ai&lt;/code&gt; / tidewave&lt;/strong&gt; to feed scoped, relevant context instead of forcing the agent to crawl your entire domain every turn.&lt;/li&gt;&#10;&lt;li&gt;&lt;strong&gt;Scope what you put in the window.&lt;/strong&gt; Because Claude re-reads the cache per turn, a bloated context is a recurring bill. Smaller, sharper inputs save money on every single interaction.&lt;/li&gt;&#10;&lt;li&gt;&lt;strong&gt;Prefer generators.&lt;/strong&gt; Tools like &lt;code&gt;mix ash.codegen&lt;/code&gt; and igniter installers absorb much of the boilerplate that the agent would otherwise have to reason through token by token.&lt;/li&gt;&#10;&lt;/ul&gt;&#10;&lt;h2 id="what-id-still-want-to-know"&gt;What I&amp;rsquo;d still want to know&lt;/h2&gt;&#10;&lt;p&gt;Remember this is just a snapshot: one task, two models. Model representation will only improve, so today&amp;rsquo;s results aren&amp;rsquo;t permanent. It remains unclear if larger, multi-feature tasks will amortize or compound comprehension cost. Also, note that I measured &lt;em&gt;processed tokens&lt;/em&gt;; because Claude bills cache reads cheaply, the dollar gap is smaller than the token gap—&amp;ldquo;2.6x the tokens&amp;rdquo; doesn&amp;rsquo;t mean &amp;ldquo;2.6x the bill.&amp;rdquo;&lt;/p&gt;&#10;&lt;p&gt;If you reproduce this and get a different number, please let me know. That’s what the repo is for.&lt;/p&gt;&#10;&lt;h2 id="the-thing-i-actually-took-away"&gt;The thing I actually took away&lt;/h2&gt;&#10;&lt;p&gt;As agents become the default way code gets written, my focus shifted from Ash&amp;rsquo;s specific token cost to something larger: &lt;strong&gt;how comprehensible is this framework to an AI?&lt;/strong&gt; This dimension—the ease with which a system can learn its shape in a fresh context window—is rapidly becoming a critical axis of design that we have barely started measuring.&lt;/p&gt;&#10;&lt;p&gt;The frameworks that will win the agent era might not be the most elegant for humans; they will be those whose structure is cheapest to &lt;em&gt;learn&lt;/em&gt;, over and over, by an AI. This is precisely the problem &lt;code&gt;usage_rules&lt;/code&gt; is quietly inventing solutions for.&lt;/p&gt;&#10;&lt;p&gt;Ash is paying a tax today. But with its commitment to usage-rules, it stands as one of the few frameworks actively building the mechanism that pays that comprehension debt back down.&lt;/p&gt;&#10;&lt;p&gt;&lt;em&gt;The benchmark, warts and all, is on &lt;a href="https://github.com/tkolsto/ash-ai-tax-bench"&gt;GitHub&lt;/a&gt;. Tell me where I&amp;rsquo;m wrong.&lt;/em&gt;&lt;/p&gt;&#10;</content:encoded></item><item><title>Hello again, again!</title><link>https://t.kolsto.no/blog/hello-again-again/</link><pubDate>Thu, 05 Mar 2026 18:45:00 +0700</pubDate><guid>https://t.kolsto.no/blog/hello-again-again/</guid><dc:creator>Thomas Kolstø</dc:creator><category>Hugo</category><category>AI</category><category>Meta</category><description>Third time I have lost the server this site runs on. This time an AI rebuilt it while I made demands about font sizes. New custom theme, same Hugo.</description><enclosure url="https://t.kolsto.no/sharing/card-base_hu_66ede5c272c03942.png" length="65974" type="image/png"/><content:encoded>&lt;p&gt;If you&amp;rsquo;re experiencing deja vu reading this, imagine how I feel writing it.&lt;/p&gt;&#10;&lt;p&gt;For the third time in the history of this site, I have managed to lose access to the server it was running on. At this point it&amp;rsquo;s less of a bug and more of a feature. Some people have a sourdough starter they keep alive. I have a personal website I keep accidentally killing.&lt;/p&gt;&#10;&lt;p&gt;This time around, the site was still live but I&amp;rsquo;d lost SSH access and the credentials to the hosting provider had apparently been sacrificed to the password gods sometime during a laptop migration. The content was right there, taunting me through a browser window like a dog on the other side of a glass door.&lt;/p&gt;&#10;&lt;p&gt;So I did what any reasonable person would do in 2026: I asked an AI to help me rebuild it. And by &amp;ldquo;help&amp;rdquo; I mean I sat here making increasingly specific demands about font sizes and button colors while the bots did the actual work. Two decades of managing teams was apparently just training for this.&lt;/p&gt;&#10;&lt;p&gt;The old &lt;a href="https://github.com/luizdepra/hugo-coder"&gt;Coder&lt;/a&gt; theme got the boot after one too many &lt;code&gt;!important&lt;/code&gt; overrides. We now run a custom theme built from scratch, with dark mode, a timeline that actually has content in it, and a toolbox section that no longer pretends I still write PHP.&lt;/p&gt;&#10;&lt;p&gt;Some things never change though. I still can&amp;rsquo;t design. The site still runs on Hugo. And I&amp;rsquo;m still not entirely sure anyone reads this.&lt;/p&gt;&#10;&lt;img src="https://t.kolsto.no/img/office-view.webp" alt="The office today" width="1200" height="1600" srcset="https://t.kolsto.no/img/office-view_hu_72bc87e87d8f39b9.webp 720w, https://t.kolsto.no/img/office-view.webp 1200w" sizes="(min-width: 1120px) 1040px, (min-width: 601px) calc(100vw - 80px), calc(100vw - 40px)" loading="lazy"&gt;&lt;p&gt;If you&amp;rsquo;ve made it this far, you are, as always, the curious one. Thanks for that.&lt;/p&gt;&#10;</content:encoded></item><item><title>Hello again!</title><link>https://t.kolsto.no/blog/hello-again/</link><pubDate>Fri, 03 Nov 2023 21:45:00 +0200</pubDate><guid>https://t.kolsto.no/blog/hello-again/</guid><dc:creator>Thomas Kolstø</dc:creator><category>Hugo</category><category>Meta</category><description>Resurrected from backup after the server aged like a fine glass of milk. New Hugo theme, new web server, new CI pipeline, and some oldies of Bob.</description><enclosure url="https://t.kolsto.no/sharing/card-base_hu_874f549cd1fbd9d.png" length="63878" type="image/png"/><content:encoded>&lt;p&gt;I had to resurect this page from backup after I forgot about it for a long period of time and the server it ran on aged like a fine glass of milk.&lt;/p&gt;&#10;&lt;p&gt;Doing so meant brushing up on hugo to adapt a new theme, setting up a new web server and a gitlab ci pipeline. Right now I&amp;rsquo;m having amnesia and deja vu at the same time. I think I&amp;rsquo;ve forgotten this before.&lt;/p&gt;&#10;&lt;p&gt;Anyhow, please enjoy some oldies but goodies of Bob.&lt;/p&gt;&#10;&lt;p&gt;&lt;img src="https://t.kolsto.no/img/bob3.jpg" alt="Bob standing on a tiled floor wrapped in a brown scarf, ears up, showing his front teeth" width="750" height="743" srcset="https://t.kolsto.no/img/bob3_hu_8e89ae8785141b06.webp 720w, https://t.kolsto.no/img/bob3_hu_bccc6bed1060b2ad.webp 750w" sizes="(min-width: 1120px) 1040px, (min-width: 601px) calc(100vw - 80px), calc(100vw - 40px)" loading="lazy"&gt; &lt;img src="https://t.kolsto.no/img/bob2.jpg" alt="Bob asleep on a terracotta floor, half on a leopard-print cushion, next to a bag of barbecue charcoal" width="750" height="744" srcset="https://t.kolsto.no/img/bob2_hu_93c22a7a38be9938.webp 720w, https://t.kolsto.no/img/bob2_hu_ca425a303ab2a15f.webp 750w" sizes="(min-width: 1120px) 1040px, (min-width: 601px) calc(100vw - 80px), calc(100vw - 40px)" loading="lazy"&gt; &lt;img src="https://t.kolsto.no/img/bob4.jpg" alt="Bob wrapped in a blanket, carried by a man in a grey hoodie in front of a brick wall" width="750" height="747" srcset="https://t.kolsto.no/img/bob4_hu_de730f3d78b3129a.webp 720w, https://t.kolsto.no/img/bob4_hu_11824519fa9e1cf.webp 750w" sizes="(min-width: 1120px) 1040px, (min-width: 601px) calc(100vw - 80px), calc(100vw - 40px)" loading="lazy"&gt;&lt;/p&gt;&#10;&lt;p&gt;RIP, Bob ❤&lt;/p&gt;&#10;</content:encoded></item><item><title>In other news</title><link>https://t.kolsto.no/blog/in-other-news/</link><pubDate>Tue, 11 Oct 2016 15:20:26 +0200</pubDate><guid>https://t.kolsto.no/blog/in-other-news/</guid><dc:creator>Thomas Kolstø</dc:creator><category>video</category><category>ffmpeg</category><category>sidequest</category><description>Serduno: a news channel aggregator, recorder and remixer I built as an ffmpeg playground. 15+ live channels, a week of backlog, 3x3 matrix view.</description><enclosure url="https://t.kolsto.no/img/serduno.png" length="54898" type="image/png"/><content:encoded>&lt;p&gt;I&amp;rsquo;ve been investing some of my spare time creating this news channel aggregator, recorder and (re-)mixer -thing lately, something that had been bubbling in the back of my mind since at least 2008.&lt;/p&gt;&#10;&lt;img src="https://t.kolsto.no/img/serduno.png" alt="Serduno home page: a retro television showing a three-by-three grid of live news channels, with a Jerry Seinfeld quote about the news always fitting the newspaper" width="731" height="515" srcset="https://t.kolsto.no/img/serduno_hu_e7d6c656dfc8ecd1.webp 720w, https://t.kolsto.no/img/serduno_hu_8add6925def3857e.webp 731w" sizes="(min-width: 1120px) 1040px, (min-width: 601px) calc(100vw - 80px), calc(100vw - 40px)" loading="eager" fetchpriority="high"&gt;&lt;p&gt;I also had to brush up on my ffmpeg-fu for some actual work related stuff, so this project ended up serving as a playground for doing both.&lt;/p&gt;&#10;&lt;p&gt;The start was an idea that it might be interesting to see how the different news outlets prioritises and covers (or not) the events that unfolds around the world. Then I figured it would also be nice to archive this for some period of time, so that if world war 3 were to break out while I&amp;rsquo;m walking the dogs, I can just check the instant replay (… or not.)&lt;/p&gt;&#10;&lt;p&gt;Way too many hours was spent looking into how to interpret the output of running encoders, and restarting them when needed; I&amp;rsquo;m relying on some rather unreliable/poorly connected sources for some of the channels. If anyone has any suggestions to how to improve this or have some private feeds to share I&amp;rsquo;m all ears.&lt;/p&gt;&#10;&lt;p&gt;There is a (limited) mechanism for &amp;ldquo;private&amp;rdquo; channels in which I&amp;rsquo;ve added a couple of commercial ones that I pay for just myself. I should probably do a writeup some day on how I ended up with OBS running in x11vnc doing screen-grabbing of firefox. It is nothing else than a hack, but as hacks go this one seems to work rather well.&lt;/p&gt;&#10;&lt;p&gt;&lt;del&gt;Try it out over at ser.du.no and please let me know what you think.&lt;/del&gt;&lt;/p&gt;&#10;&lt;p&gt;Features:&lt;/p&gt;&#10;&lt;ul&gt;&#10;&lt;li&gt;15+ live news channels from around the world, most in English&lt;/li&gt;&#10;&lt;li&gt;7 days backlog segmented by channel/day for all channels&lt;/li&gt;&#10;&lt;li&gt;Rotate/view multiple channels simultaneous in 1x1 2x2 and 3x3 matrixes&lt;/li&gt;&#10;&lt;li&gt;Embedded chat with live-breaking-news-alert-bot v0.1&lt;/li&gt;&#10;&lt;li&gt;Loop video to make it easy to extract soundbites&lt;/li&gt;&#10;&lt;li&gt;Chromecast &amp;amp; Airplay support (on some of the streams)&lt;/li&gt;&#10;&lt;/ul&gt;&#10;&lt;p&gt;Todo:&lt;/p&gt;&#10;&lt;ul&gt;&#10;&lt;li&gt;Chat/live-breaking-news-alert-bot needs serious work to provide less comedic relief. OCR is hard.&lt;/li&gt;&#10;&lt;li&gt;Re-enable DASH tech for video.js. Something broke with last version.&lt;/li&gt;&#10;&lt;li&gt;Make youtube-tech optional via selection, it&amp;rsquo;s not that great for matrixes (slow)&lt;/li&gt;&#10;&lt;li&gt;Create Recording/Export module for video.js&lt;/li&gt;&#10;&lt;li&gt;Set up nginx hls/dash secure links&lt;/li&gt;&#10;&lt;li&gt;Add moar channels. Linux/PC based sat recievers might be the way to go.&lt;/li&gt;&#10;&lt;li&gt;Make channel-matrixes refresh one by one instead of all of them at the same time somewhat fixed for 3x3&lt;/li&gt;&#10;&lt;li&gt;#hindenburg { float: none; }. I am not a designer. The lack of design + related issues will forever be banished to the end of this list, unless someone wants to help.&lt;/li&gt;&#10;&lt;/ul&gt;&#10;</content:encoded></item><item><title>WebGL Traffic Analyzer</title><link>https://t.kolsto.no/blog/webgl-traffic-analyzer/</link><pubDate>Sun, 09 Nov 2014 14:45:00 +0200</pubDate><guid>https://t.kolsto.no/blog/webgl-traffic-analyzer/</guid><dc:creator>Thomas Kolstø</dc:creator><category>webgl</category><category>html5</category><category>dataviz</category><category>sidequest</category><description>A weekend WebGL globe for the ops centre video wall, mixing IDS and firewall data into eye candy that only I could interpret.</description><enclosure url="https://t.kolsto.no/img/trafficanalyzer.jpg" length="27329" type="image/jpeg"/><content:encoded>&lt;p&gt;Last weekend I created a WebGL Traffic Globe Analyzer thing for our operations centre video wall.&lt;/p&gt;&#10;&lt;p&gt;Here it is.&lt;/p&gt;&#10;&lt;img src="https://t.kolsto.no/img/trafficanalyzer.jpg" alt="WebGL globe showing the Atlantic and Europe on a black background, with red and green spikes marking traffic sources and a clock in the corner" width="916" height="662" srcset="https://t.kolsto.no/img/trafficanalyzer_hu_29e91e3ec5c8f543.webp 720w, https://t.kolsto.no/img/trafficanalyzer_hu_cb944b1a5b04e08d.webp 916w" sizes="(min-width: 1120px) 1040px, (min-width: 601px) calc(100vw - 80px), calc(100vw - 40px)" loading="eager" fetchpriority="high"&gt;&lt;p&gt;Project is deeply inspired by a similar globe I once saw in a lobby somewhere in Silicon Valley, but that one was &amp;ldquo;just&amp;rdquo; displaying aggregate search traffic. I&amp;rsquo;ve tried to mix in some sensor data from our IDS system and firewalls to make it a bit more interesting to look at with our comparatively humble dataset.&lt;/p&gt;&#10;&lt;p&gt;It&amp;rsquo;s really more eye candy than useful, especially since only I know what the different colors represent (..but maybe you can guess some of them?).&lt;/p&gt;&#10;&lt;p&gt;You know what they say, there are only two kinds of people in this world — those who can extrapolate from incomplete data sets…&lt;/p&gt;&#10;</content:encoded></item><item><title>Hello World!</title><link>https://t.kolsto.no/blog/hello-world/</link><pubDate>Fri, 14 May 2010 04:20:00 +0200</pubDate><guid>https://t.kolsto.no/blog/hello-world/</guid><dc:creator>Thomas Kolstø</dc:creator><category>Meta</category><description>The obligatory first post, from 2010, written for the curious one who wonders where this all started. Includes a picture of my dog.</description><enclosure url="https://t.kolsto.no/sharing/card-base_hu_5c60c085684c7253.png" length="59030" type="image/png"/><content:encoded>&lt;p&gt;This is the obligatory first post on my new home on the intertubes. Being the first post, on a new site, I&amp;rsquo;m being realistic and don&amp;rsquo;t think it will be seen by too many people.&lt;/p&gt;&#10;&lt;p&gt;At best, someone who is interested in my earlier content will think hey I wonder what his first post was. This post is for you, the curious one.&lt;/p&gt;&#10;&lt;img src="https://t.kolsto.no/img/bob.jpg" alt="Bob, a black and tan miniature pinscher, lying on a sofa cushion under a pile of blankets with a toothy grin" width="990" height="743" srcset="https://t.kolsto.no/img/bob_hu_b29049ba5bf30d10.webp 720w, https://t.kolsto.no/img/bob_hu_8debb83e457a7275.webp 990w" sizes="(min-width: 1120px) 1040px, (min-width: 601px) calc(100vw - 80px), calc(100vw - 40px)" loading="lazy"&gt;&lt;p&gt;I hope you enjoy this picture of my dog.&lt;/p&gt;&#10;</content:encoded></item></channel></rss>