sethrylan.org on Typing, Fast and Slow

Calibrates

Mon, 01 Jan 0001 00:00:00 +0000

SLO Burn

Mon, 01 Jan 0001 00:00:00 +0000

readis

Mon, 01 Jan 0001 00:00:00 +0000

boop

Mon, 01 Jan 0001 00:00:00 +0000

Counting with Less

Mon, 01 Jan 0001 00:00:00 +0000

Something Old, Something New

Mon, 01 Jan 0001 00:00:00 +0000

Oxford

Fri, 12 Jun 2026 00:00:00 +0000

Efficiency and Knowledge Transfer

Tue, 26 May 2026 00:00:00 +0000

I’ve been thinking more on the economics and efficiency of LLMs, and have been meaning to put together a John Henry-esque comparison of human and machine speeds.

Mike Veerman’s lovely tokenspeed (via) beat me to it, so I took the weekend to work on my interactive essay skills with Claude and build something new.

→ Efficient Knowledge — WPM & TPM, for humans and models

This should work work on most platforms that don’t have reduced-motion enabled.

Counting with Less

Wed, 18 Mar 2026 00:00:00 +0000

Count-Min Sketch is a beautiful data structure. It’s relatively young, first described by Cormode and Muthukrishnan in 2003. I came across this algorithm while looking at rate limiting strategies, and the core idea is so clever that I wanted to build something interactive around it.

I promise there is very little math needed to follow along.

We’re in our everything-is-a-bottleneck era. Traffic volumes have never been higher, and compute/memory has never been harder to come by. AI agents are using more and more of those resources, both directly via inference/training and through API calls on behalf of their handlers. I think probabilistic data structures like the count-min sketch will be increasingly important for building efficient systems that can handle higher volumes of data. And they also lean into the non-exact, non-deterministic nature of these systems, which is fun to learn about. It’s simple to expect exact answers from our data structures, but in many cases an estimate is good enough, and can be obtained much cheaper.

Simple Agent Patterns

Mon, 09 Mar 2026 00:00:00 +0000

Note

This was a short presentation on the current state of software engineering with agents.

Calibrate Your Confidence

Wed, 04 Mar 2026 00:00:00 +0000

Intuition is not about being right. It’s about knowing when you are.

The world is changing at a quietly astonishing rate. Mental models update slowly - we carry assumptions from the last year, or the last decade. The data changes every day, and our intuition doesn’t keep up. We’re miscalibrated, and we don’t even know it.

Calibrated thinking isn’t about having the right answer. It’s about holding beliefs with appropriate uncertainty, and updating them when evidence arrives. It’s a skill, and it’s trainable, and I wrote to game to practice.

Tools for Writing Tools for Agents

Wed, 25 Feb 2026 00:00:00 +0000

2026 is turning into a good year for CLIs. And how do we help agents write CLIs for agents and humans to use? With more CLIs, of course!

Here are two tools I’ve been building so my agents can write more tools.

Tool #1: Let the Agents Demo Their Work
#

I was impressed by Simon Willison’s Showboat and Rodney. Showboat creates executable demo documents - markdown with code blocks that an agent runs, captures output, and a verifier can re-execute to confirm the results match. Rodney drives a persistent headless Chrome instance from the command line. Together, they let an agent prove its web-facing work.

A Good Year for CLIs

Mon, 23 Feb 2026 00:00:00 +0000

My Christmas wish was for a more complete Datadog CLI:

a better version of dogshell that supports read/write, OAuth, and a more intuitive interface.

And the folks at Datadog shipped it! Pup is a Go Rust CLI for the Datadog API: 283+ subcommands, OAuth2, and output formatting designed for both humans and AI agents. It’s everything I wanted, and more.

What Pup Does
#

Designed for progressive disclosure:

2026: The Year of Skills

Wed, 24 Dec 2025 00:00:00 +0000

This was MCP’s year. We re-arranged our schedules to accomodate the highly anticipated server ships from our vendors. The promise of connecting agents to tools was irresistible. We wanted to see it, we wanted to use it, we wanted to build on it.

It wasn’t all good. MCP, as a protocol and a suite of products, is half-vibes. It mostly works, but I couldn’t tell you how. And that hasn’t stopped anyone from unlocking an incredible amount of value by letting their agents churn through tokens and make API calls on their behalf.

Photostream

Fri, 17 Oct 2025 00:00:00 +0000

Photostream

Thu, 12 Sep 2024 00:00:00 +0000

Photostream

Sun, 16 Jun 2024 00:00:00 +0000

Things I learned in 2023

Sun, 31 Dec 2023 00:00:00 +0000

Face muscles are the only muscles directly attached to the skin.
Flint forms in nodules, which take the shape of marine animal burrows from 80MA. These burrows would fill with chalk, and flint would form in a sedimentary process.
The universe is beige. Cosmic latte, “the average color of the universe as perceived from the Earth.” In hex, #FFF8E7.

Photostream

Sat, 02 Dec 2023 00:00:00 +0000

Chimayo

Wed, 13 Sep 2023 00:00:00 +0000

Photostream

Sat, 15 Jul 2023 00:00:00 +0000

Things I learned in 2022

Sat, 31 Dec 2022 00:00:00 +0000

Cotton was independently domesticated in the Old and New Worlds.
Just more than half of total carbon emissions came after 1991.
The dwarf sperm whale is unique among aquatic mammals for being able to release “ink”.
Hummingbirds and bees generate a static field during flight that allows them to electrodynamically sense flowers and when that flower was last harvested.
One influenza variant may have gone extinct between 2020 and 2022.
“Perverse polymorphism” is a phenomenon where the crystalline form of a molecule is replaced by a more energetically favorable one, like ice-nine in Vonnegut’s novel Ice Nine. In real life, this can make some crystals impossible to manufacture, because the molecules will revert to the “infectious” form in the presence of microscopic seed crystals. In 1998, the HIV antiretroviral Ritonavir disappeared from the market due to Form II seed crystals permeating the air of the manufacturing environment. Abbott, the manufacturer, spent $250 million trying to remove the Form II contamination. The problem was eventually solved… by using a gelatin capsule to deliver the new form of the molecule (source).

Chicago

Sat, 02 Jul 2022 00:00:00 +0000

Cassis

Fri, 25 Mar 2022 00:00:00 +0000

Photostream

Sun, 19 Dec 2021 00:00:00 +0000

Things I learned in 2021

Fri, 03 Dec 2021 00:00:00 +0000

There are only 15 bobsled tracks in the world.
Company Demonyms: Googler (also Noogler, Xoogler, and Doogler and Greygler), Microsoftie, Tweep, Amazonian, Zoomie, Hubber, Oktanauts, Shopifolk, Pinployee, Twilion, Sentaur, Tabloid.
All animals combined weigh 4 gigatonnes. All plastics, 8GT.
Sperm whale oil was used as an automatic transmission oil additive until the 1970s.
BP runs their own fiber backbone through the Gulf of Mexico.
Staghorn ferns are eusocial.
Marsquakes are seasonal and occur most frequently in the winter.
“Geofoam” is a marketing term for styrofoam blocks used in large scale engineering.
Brandolini’s Law states that it takes a lot more energy to refute bullshit than to produce it. Hence, the world is full of unrefuted bullshit. Also know as the Bullshit Asymmetry Principle.

Photostream

Mon, 01 Mar 2021 00:00:00 +0000

Photostream

Thu, 03 Dec 2020 00:00:00 +0000

Photostream

Wed, 13 Jun 2018 00:00:00 +0000

Baltimore

Photostream

Sat, 17 Mar 2018 00:00:00 +0000

Irazú

Playón, CR

Photostream

Tue, 21 Nov 2017 00:00:00 +0000

Valencia

Barcelona

Grad School (the good parts)

Wed, 09 Aug 2017 00:00:00 +0000

On Abstractions
#

Systems designers are abstraction merchants.¹

Nothing is so difficult it cannot be solved by another level of indirection.²

On Files
#

The life of an average file is tedious³ and brief⁴. Sequential access is rarely sequential with multiple threads.

On Queueing
#

Little’s Law has bizarre and counter-intuitive conclusions. Suppose bank customers take an average (exponential distribution) of 10 minutes to serve and they arrive at the rate of 5.8 per hour. With 1 teller, the average wait time is over 5 hours. With 2 tellers, it is 3 minutes.

Photostream

Sat, 15 Apr 2017 00:00:00 +0000

Verona

Photostream

Sun, 15 Jan 2017 00:00:00 +0000

Reykholt, Borgarfjörður

Akureyri

Dettifoss, Skútustaðahreppur

Photostream

Sun, 26 Jun 2016 00:00:00 +0000

Montreal, Quebec

Indian Pipe

Photostream

Mon, 04 Jan 2016 00:00:00 +0000

Fox Theatre, Detroit

Photostream

Sun, 01 Nov 2015 00:00:00 +0000

Earl's Ford, Georgia

Photostream

Wed, 10 Jun 2015 00:00:00 +0000

{{ page.title }}
#

Edinburgh

Giardini Pubblici

Workarea

Thu, 15 May 2014 00:00:00 +0000

Photostream

Mon, 10 Mar 2014 00:00:00 +0000

Photostream

Mon, 24 Feb 2014 00:00:00 +0000

Something Old, Something New

Sun, 15 Dec 2013 00:00:00 +0000

Choropleth map of the age of structures in the City of Charleston, Old and Historic Districts.

Weekend Project - Bookcase Door

Sun, 20 Oct 2013 00:00:00 +0000

Materials
#

Dimensional lumber in nominal sizes.

1x6 (36 feet)
1x5 (16 feet)
1x2 (2 feet)
1/4" luan
paint (shelf and wall)
2 1/2" trim nails
sandpaper
1 1/2" rollers

Tools
#

Compound Miter Saw
Drill/Driver
Pneumatic Finish Nailer
Carpenter’s Square
Level
.5" Chisel

Total Cost
#

~$170
20 hours

Obviously, before you begin the project you should check your local building codes. Usually, there are no issues as long as your are not modifying the rough opening dimensions or changing an egress door.

FindBugs for Android

Sun, 14 Jul 2013 00:00:00 +0000

The very useful code quality plugins for Gradle (Checksytle, FindBugs, PMD, etc) expect to have the Java plugin applied, which is incompatible with the new Android gradle build system. Some developers, like Netflix, resolve this problem by omitting the plugins from their otherwise standardized build process. Others have tried to create a build system inside another build system, at which point you can have whatever plugins you want, and most use a wrapper around the Ant tasks which predate Gradle. Behind the scenes, this is what the Gradle plugins do anyway.

IntelliJ Scala Plugin

Fri, 05 Jul 2013 00:00:00 +0000

The IntelliJ Scala plugin v0.13.286 does a few things which were previously manual steps. This means that following the old documentation can lead to a few unhelpful error messages, when all you want is syntax highlighting, autocomplete and the occassional REPL support.

Assuming you start with a Gradle project using the Scala plugin and clean IntelliJ IDEA 12.1.4 install, the steps to add Scala support are:

File -> Import Project. Select folder with parent build.gradle.
Import project from external model
Select whether to use the gradle wrapper or a local installation, then finish import with default settings.
Right-click parent folder in project structure, select Add Framework Support....
Depending on your Scala installation, some expected folders may be missing.

And finished. If you had existing scala-compiler/scala-library global libraries, or tried following some of the older plugin documentation, then you will likely find errors such as

ScalaConsole with Gradle

Tue, 02 Jul 2013 00:00:00 +0000

The usual build tool for Scala is SBT, but I’ve been trying to stay consistent with Gradle builds. Of course, the most lacking feature is a decent REPL for manipulating the main/test classes, and there appear to be two possible solutions to this problem.

ScalaConsole in the Gradle scala plugin
#

The ScalaBasePlugin for Gradle defines the scalaConsole and scalaTestConsole tasks, which aren’t widely used (or documented) due to known issues. In particular, you have to make a few changes before the feature is usable. In order the fix the

Khan Academy

Thu, 13 Sep 2012 00:00:00 +0000

After finishing the Khan Academy math exercises in ~80 hours, I made a few notes regarding deliberate practice and its place in online learning media. Khan Academy (KA) focuses far more on rout work that Udacity or Coursera, with most of modules requiring one of 1-3 techniques for solving. The adaptive training plan mostly makes up for this, but one could easily accomplish every KA exercise without any understanding of the background.

Arduino Barometric Altimeter

Tue, 31 Jul 2012 00:00:00 +0000

Using Sparkfun’s BMP085 sensor board (SEN-09694), I put together some quick Arduino code to measure temperature-adjusted barometric pressure, store the data on an SD card using Adafruit’s data logger shield (v1.0), and then calculate altitude. This is a pretty simple project. In fact, the biggest challenge was getting a suitable formula for calculating altitude from atmospheric pressure.

Testing the sensor was very simple, following the tutorial and schematic on Sparkfun’s site with breadboard wiring.

Multiplex Easy Star Build

Tue, 17 Jul 2012 00:00:00 +0000

Materials:
#

Where possible, I’ve listed an online retailer, but some of the generic items may require searching.

Easy Star Electric Kit
Hitec HS-65HB Micro Karbo Gear Servo (1.8kg/11g) x2
TURNIGY Plush 30amp Speed Controller (25g)
S2838-2800 Brushless Inrunner (2800kv)
Industrial Strength Velcro
Boyd Red Eamel Spray Paint
Testors Gloss Gray (1238) Enamel Spray Paint
Spinner/Collet
12awg wire
3mm basswood
1.3mm x 4.5mm carbon fiber strip
5.3mm hollow carbon fiber tube
2.5mm solid carbon fiber tube
6x4e APC propeller
Nylon hinges
5/16" zinc washer

Prior to the build, I tested three types of glue:

ST101 Homework in R

Wed, 04 Jul 2012 00:00:00 +0000

As part of Udacity’s ST101, I’ve started completing the homeworks in R using RStudio. Below is the solution for homework 2.6:

print.matrix <- function(m){
 write.table(format(m, justify="right"),
 row.names=T, col.names=T, quote=F)
}
p_loaded = 0.1
p_hfair = 0.5
p_hloaded = 0.9
cells <- c(4, 0,10, 0, 20, 0, 0, 5, 2, 3)
rnames <- c("# heads", "# tails")
# creates a list of in form ("#1", "#2", "#n"...) to n=number of trials
cnames <- c(paste("#", 1:(length(cells)/2), sep = "")) 
trials <- matrix(cells, nrow=2, ncol=5, byrow=FALSE, dimnames=list(rnames, cnames))
#print(trials)
 
# P(fair|flips) = (P(flips|fair) * P(fair))/ P(flips)
 
for (trial in 1:ncol(trials)) {
 # joint probablities for fair and loaded coins given observed results
 jp_fair = trials[1, trial] * p_hfair + trials[2, trial] * (1-p_hfair)
 jp_loaded = trials[1, trial] * p_hloaded + trials[2, trial] * (1-p_hloaded)
 if (jp_fair < jp_loaded) {
 print(paste("trial ", trial, " is < 0.5 probable for a fair coin!", sep=""))
 } 
}

Bio

Mon, 01 Jan 0001 00:00:00 +0000

I make things and live in Charleston, South Carolina

sethrylan.org on Typing, Fast and Slow

Calibrates

SLO Burn

readis

boop

Counting with Less

Something Old, Something New

Oxford

Efficiency and Knowledge Transfer

Counting with Less

Simple Agent Patterns

Calibrate Your Confidence

Tools for Writing Tools for Agents

Tool #1: Let the Agents Demo Their Work #

A Good Year for CLIs

What Pup Does #

2026: The Year of Skills

Photostream

Photostream

Photostream

Things I learned in 2023

Photostream

Chimayo

Photostream

Things I learned in 2022

Chicago

Cassis

Photostream

Things I learned in 2021

Photostream

Photostream

Photostream

Baltimore

Photostream

Irazú

Playón, CR

Photostream

Valencia

Barcelona

Grad School (the good parts)

On Abstractions #

On Files #

On Queueing #

Photostream

Verona

Photostream

Reykholt, Borgarfjörður

Akureyri

Dettifoss, Skútustaðahreppur

Photostream

Montreal, Quebec

Indian Pipe

Photostream

Fox Theatre, Detroit

Photostream

Earl's Ford, Georgia

Photostream

{{ page.title }} #

Edinburgh

Giardini Pubblici

Workarea

Photostream

Photostream

Something Old, Something New

Weekend Project - Bookcase Door

Materials #

Tools #

Total Cost #

FindBugs for Android

IntelliJ Scala Plugin

ScalaConsole with Gradle

ScalaConsole in the Gradle scala plugin #

Khan Academy

Arduino Barometric Altimeter

Multiplex Easy Star Build

Materials: #

ST101 Homework in R

Bio

Tool #1: Let the Agents Demo Their Work
#

What Pup Does
#

On Abstractions
#

On Files
#

On Queueing
#

{{ page.title }}
#

Materials
#

Tools
#

Total Cost
#

ScalaConsole in the Gradle scala plugin
#

Materials:
#