xkcd shop
xkcd shop
My experience buying ads in ChatGPT: the good, the bad and the ugly
When OpenAI dropped the minimum $50k spend on their ad platform, I thought I should give it a try for my seating planner app. Perhaps I can get some cheap and well targeted clicks, before all the c…
How juniors learn the trade when an agent writes the code
In my last post I described how we structure delivery so that human judgement is spent where it matters, and I left one problem open: that model assumes seasoned judgement at the gates, and I could not say how a junior grows it inside a system that shields the…
Elasticsearch Isn't Dead. You Probably Don't Need It
Search gets slow, someone says 'we need Elasticsearch,' and two weeks later there's a new cluster, a sync pipeline, and a class of bugs that didn't exist…
Felony Bench: Be AI, Do Crime
Felony Bench is the most important benchmark of our time, some are saying, counting the number of questionable decisions made by AI agents. We might need a Pandemic Bench soon...
Rise of the Software Factory
Everyone is suddenly talking about the software factory. Few have built one. Here is what it actually means, and the three properties every software factory must have.
Language Models Are Anomaly Detectors
I work on security detection at Datadog. The model I have been building, Mambark, reads audit logs the way a language model reads text and scores every event by how surprising it is. It is a small model by today's standards, about 97 million parameters, and it…
A year of databases: how I fell in love with programming again. | Denny's blog
NOTE: LLM is not involved, typos and others mistakes are original.
Platforms Should Let Agents Talk
X, GitHub, and every other publishing platform should let people show when their agents helped them think, write, and talk to each other.
Remotely unlocking electric scooters — Henri Emategui
How subdomain recon, WordPress user enumeration, and a brute force led to admin access and remotely unlocking scooters.
Embedding models benchmark for code duplication detection
A focused benchmark evaluating how well embedding models separate duplicated code from unrelated, including adversarial cases
How good are agents actually at CAD? — Seldon
CADBench is Seldon's benchmark for computer-use agents in CAD — 105 native, long-horizon mechanical-design tasks in Autodesk Fusion whose geometry, feature history, and constraints are checked by deterministic verification.
Three ways to smuggle SQLite into Nix
The core of nixpkgs-multiverse, when you strip away the Nix API and the CLI, is an index. It is a map from (attribute, version) to the revision that shipped it as a JSON file.1 There are actually a few other files that drive other features such as the statisti…
Your Repository is Your Swarm
I’ve written before about why I love self-hosting software: real data ownership, total independence, and never having to answer to changing terms of service. But recently, plugging headless AI agents directly into my self-hosted Gitea setup turned my repositor…
Why do I work?
Jon's personal site.
Why Do We Need a New Agentic Browser
Primary Internet user is changing but its tools aren’t
Archery Feat
The Creation of Abulafia
In 1988, Umberto Eco published Foucault’s Pendulum. In it, three editors at a Milan publishing house (Belbo, Diotallevi, and Casaubon) spend their working…
How to not burnout | alikhil | software engineering, kubernetes & self-hosting
Burnout prevention for software engineers: practical ways to set work-life boundaries, recharge, and protect your motivation
Consciousness, Computability, Decidability, and Evolvability: A Weaker Problem Worth Attacking
1. The starting intuition, and why it’s too strong as stated # Gödel’s incompleteness theorems show that any sufficiently expressive, consistent formal system \$F\$ contains a true statement \$G\$ it cannot prove from its own axioms — informally, $$ G \leftrig…
The 44-Minute CI Job (and the One Query Behind It) — Recall Blog
A release-gate benchmark suddenly took 44 minutes. The cause was one search query running a 10-billion-operation join — and the fix took it from ~9 minutes to ~1 second.
Doing everyone else's job
I find it very useful to be able to do everyone else’s job — it helps you learn how they work and what they need, and it helps even more when you need their job done, but they won’t do it themselves.
Cobalt: apps and an SDK for Kobo e-readers
A launcher, a signed App Store, a Rust SDK, and a capability-isolated runtime for the Kobo Clara BW. Install once over USB; everything after that arrives over Wi-Fi.
The Last 20%
The other day my friend Chris pinged me on a Twitter thread. Todd Dailey was making the case that digital signage is a huge industry selling bad products …
Let’s talk about “vibe writing”
When it comes to writing, I only used AI for smaller requests for a long time: fixing a typo, or smoothing over a clumsy sentence – the sort of thing a spellchecker does. I always found that handy, especially in English, which isn’t my mother tongue, but it ne…
The software business after code scarcity
Open source companies learned how to build valuable businesses when access to the code was free. AI is forcing the rest of the software industry to confront the same business-model question.
Changelog - Kagi Search
Better search results with no ads. Welcome to Kagi (pronounced kah-gee), a paid search engine that gives power back to the user.
I accidentally logged hundreds of thousands of phone calls to military bases
How an expired nameserver let me take over e164.arpa zones for multiple territories, and why I probably should have checked my logs sooner.
Vision | DeepSeek API Docs
The deepseek-v4-flash-vision-exp model accepts images alongside text, so you can ask the model to describe pictures, read text from screenshots, analyze charts, and more.
Japan tried to build an operating system for the entire world, then the US government intervened
TRON was an ambitious project to say the least, and after a good start, it suddenly came to a screeching halt.
Readers can't identify watermarked AI text
In the last few weeks, I’ve been complaining that everyone is wrong about AI watermarking: it isn’t really anti-consumer and it doesn’t make the outputs any…
Huzzah
A new experimental way to code with AI
The August 17 outage, and the work ahead
An update on the August 17 outage and the steps we're taking to improve reliability.