‘Our p99 is fine’ and ‘one user in six has a bad time’ are the same sentence. Herein a small browser tool to visualise the tail-at-scale effect: how a rare slow call stops being rare once a page waits on enough of them, and why every dashboard stays green while it happens.
Part 1 ended with a table with 192 entries, one per state-action pair, updated by a single rule. That post covered four foundational algorithms from 1957 to 1989: By then much of what is fundamental in reinforcement learning was in place—value functions, policies, temporal difference errors, the tension between exploration and exploitation, actors and critics. The three lessons in this post cross…
Most reinforcement learning code starts after the interesting part. Policywerk is a plain-Python Reinforcement Learning project that rebuilds Bellman, Actor-Critic, TD Learning, and Q-Learning from first principles, with tests, tiny environments, and animated visualizations that make the machinery visible.
The two most interesting models in the modelwerk series of lessons about neural networks turned out to be ones that came after the transformer. Not because they're better, but because they seek to answer the question 'what comes next?' in completely opposite ways. In this post we take a look at Mamba and Continuous Thought Machines.
For the longest time children stood inside looms lifting weights to allow the thread to go through on the command of the weaver. They were called draw boys. There's a moment in the history of weaving where the draw boy disappears, quickly, after the introduction of the Jacquard loom, the one with the punch cards, before Babbage, before Hollerith applied the same mechanism to computing. In doing so…
Working with agents doesn't feel like getting into flow to me. It feels more like a game loop: set direction, trigger action, watch what happens, evaluate, intervene, adjust, try again. Different kind of focus, different kind of reward. But fun.
I recently ran a security audit on a code base using a sub-agent and gave it a YAA (‘You Are A’) persona as a security engineer. It was good enough to want to keep around. This post describes a simple way to capture engineering personas and run them in Claude Code with a custom ‘/agent’ command.
tars is a personal AI assistant with CLI, Web UI, Email, and Telegram channels, persistent memory, hybrid search, integration with tools I used all the time. About 35 features, 14kloc of python and 600 tests all told. I didn't write any of it. The experience was different enough from traditional development, even from what we're doing with AI last year, to be worth writing up. The surprising part…
Some working thoughts on Steve Yegge’s Welcome to Gas Town. I read the essay a few times, then dug into the implementation to better understand how the system is structured. What follows is a mix of appreciation, architectural curiosity, and open questions about where this kind of agent harness might lead.
The Jacquard loom, invented by Joseph Marie Jacquard in the early nineteenth century, was one of the first machines to be controlled by something resembling a stored program, namely a sequence of punched cards [2]. It wasn't a computer in the modern sense, but it demonstrated that a machine could perform complex work by following an external set of instructions, an idea that would echo through the…
An organisation should be as large as necessary and as small as possible.
Here are some notes on a few books I enjoyed and learned from in 2024, not all of which are books published that year.
In this post we’ll cover how to get multiple machine learning models running on Apple using Ollama, open-webui, and for image generation, stable-diffusion-webui.
At the end of 2019, I posted a book list for engineers new to machine learning, to help develop basic knowledge of the fundamentals, organised into four groups: Machine Learning & Algorithms, Tools & Frameworks, Data Science & Analysis, and Companion Mathematics. This post provides an updated book list using the same groups, which have held up well. At the time the list covered 26 books, this…
I’ve mentioned more than once how effective I find org-mode as a productivity tool and organiser, without really explaining why. This post describes how I use org mode and why it’s been such an important application for well over a decade.
Here are some notes on a few books I enjoyed and learned from in 2021, not all of which are books published that year.
Some notes on some books.
Here are some notes on a few books I enjoyed and learned from in 2020, not all of which are books published that year.
“You've gotta start with the customer experience and work backwards to the technology. You can't start with the technology and try to figure out where you going to try to sell it.” — Steve Jobs
Technical Prime is a lightweight mechanism for enabling clarity of decisions and who’s working on what, that also affords learning and knowledge sharing.