Optimizing code starts with measuring it, and a measurement is only useful if it is repeatable: a 2% improvement is invisible under 5% of noise. Yet on an …
This is nothing to be proud of, but I have never really studied optimisation in depth. Oh sure, I know my Adam from my AdaGrad and I even used L-BFGS one time, but when people start talking about dual spaces and convergence for L∞L^\infty continuous functions,…
Exchanging my data for service use, I took up OpenAI's offer for a free trial of ChatGPT Plus for a month. After slowly vibecoding an IDE plugin throughout last month, I'm eager to share my notes and lingering thoughts.
Source Back in February I posted The Kessler Syndrome , which also included a brief section mentioning the impacts of the proposed megacons...
I recently got a Pebble Time 2 as it seemed like a fun smartwatch away from Google/Apple/Samsung with a good 4 weeks of battery life. One thing I wanted to do is to create a custom watchface for my specific problems.
Why software engineers are grieving
Mission, craft, and the end of a historically unusual moment.
bertolami
Joe Bertolami's portfolio and blog covering artificial intelligence, game development, computer graphics, data compression, and more.
June Spam wave
I've been observing a new email spam wave hitting my servers in the last couple of weeks... Way above the normal "background radiation levels" for my server... 99% of them are poorly configured, and usually fail during "does the sender domain actually exist i…
Scaling laws are one of the most critical empirical findings in deep learning. The observation is simple in form: the training loss $L$ decreases predictably as we scale up model size $N$, dataset size $D$, and compute $C$, following a power-law curve, which a…
Flatpak.org Rewrite | Jakub Steiner
Flatpak website deserved a better backend.
When you need an approval checkpoint but don't want it everywhere, make the gate required but flexible in formality rather than optional but formal. A required gate lowers the stakes of misjudging risk.
Description of post
Karl Popper offered an elegant normative account of science as a process of conjectures and refutations: we formulate hypotheses or theories, expose them to criticism, scrutiny and experiment, and retain only what survives our best attempts at falsification. E…
From API to Ontology: An Architecture for On-demand Semantic Digital Twins
In our previous post, we drew a line between two layers of an urban energy digital twin: the Truth Layer, a relational system of record that protects the structural integrity of a city's data, and a promised Knowledge Layer, a semantic graph that would let tha…
A single root package.json can look simpler in an Nx monorepo but dependency hoisting hides ownership and can turn upgrades into confusing production build failures.
A tutorial on Hindley Milner type inference.
A technical report about a rare correctness issue that we found and fixed.
I've Automated My Homelab with the BAT Stack (Bootc, Ansible, Terraform/OpenTofu)
I rebuilt Geronimo Lab using Bootc, Ansible, Terraform/OpenTofu. This stack allows me to manage my homelab entirely through infrastructure as code tools, making it easier to maintain, scale and automate.
Don’t Use a Stochastic Engine for Deterministic Workflows
Before adding AI to source-controlled organizational workflows, ask whether the answer should be generated live at all.
Frontier models plateau on domain-specific tasks well before teams expect it. Here's how to diagnose whether you've hit a true capability ceiling or a prompt, eval, or data problem — and which technique actually breaks through.
Engineering for Bounded Cognition
The mind that builds software is far smaller than the software it builds, and almost everything good in engineering is a way of living with that gap.
Stop generating what you already have
A teammate pinged me about a self-hosted LLM taking 42 seconds per extraction. The model was not slow. The prompt was asking it to copy text it already had. Here is how I cut it to 6 seconds.
I'd boxed AI coding assistants in as a great autocomplete that would plateau, convinced senior engineers like me would spend years cleaning up after them. Then Claude Opus 4.5 reverse-engineered my way off a locked ISP router, software-only, the kind of work o…
We use models built on transformers every day, yet the architecture itself usually stays a black box. It doesn't have to, and following it doesn't take heavy math. A visual, ground-up walkthrough: how words become numbers, how attention lets them shape each ot…
My troubles with record identifiers starts with a web site I developed, Eksi Sozluk. It's been one of the most popular Turkish web sites in the world for the last quarter century. When I first wrote it in 1999, I had to run it on a remote hosting service
Two posts ago I quoted a warning: an AI will find it easier to convince you it has a proof than to write one. A middling new paper finally put a number on that gap — 0.99 against 0.55.
OpenAI previews GPT-5.6 Sol, a next-generation model with stronger capabilities in coding, science, and cybersecurity, paired with its most advanced safety stack.
A series of unfortunate agents.
Claude Code in a Container | no dogma blog
I don't like running Claude Code on my computer, instead I put it in a container and mount a directory from the host into the container where Claude generates code.
AI inference is obviously profitable
Many people claim that AI inference is unprofitable to serve, and thus must be subsidized by an ocean of dumb money from investors who believe that some future AI model will come to dominate the world economy. When that dumb money goes away, so will AI product…
Open Letter | Akrites
An open letter from the technology industry, and the launch of Akrites - a coordinated effort to remediate vulnerabilities in the open source software the world runs on.
home · about
This is a quick post, mostly for my own reference. I've avoided LUTs and 'Log' video footage for years1, mostly because of the extra tiny bit of workflow involved. Like RAW photos, 'Log' footage retains the video sensor's full dynamic range, so you can pull mo…
The ‘papers, please’ era of the internet will decimate your privacy
Americans, be warned: Age verification is identity verification.
Om Malik, 1966-2026
Om Malik passed away on June 24, 2026, at Stanford Hospital after a long health journey with his heart. He was surrounded by family and friends. We invite you to share your remembrances of Om in th…