Search // Blog Archive
Filter by category, tag, or year — or run a text search across notes on LLM-assisted development, systems architecture, and coding agents.
10/23/2026 · Game Development
Acting is a multiplayer RPG-style dialogue experiment. Cars is a complete car fabricator and multiplayer racing simulator with Box3D physics. Both are up to play.
Acting is a multiplayer RPG test with dialogue. Cars lets you build a car and race it with other players. Both are in the arcade.
Read article10/23/2026 · LLM Research
Astra handles the wide coding jobs I used to spend more time correcting. Opus 5.5 promises cheaper work at that level. I want to see what that buys over a full project.
Astra earned its place on the hard jobs. Opus 5.5 needs the same test, at a lower price. With Acting and Cars up, I have plenty of work to try it on.
Read article09/23/2026 · LLM Research
Jev is TypeSafe's System One decision model. RLCD is the interesting part. The 25-line Python version works. I am exploring, not claiming.
Not a chat model. A decision interface. TypeSafe shipped it because nobody else was. RLCD might be the real story. The local Python demo already works.
Read article09/11/2026 · LLM Research
Fall update to my coding LLM rankings. OpenAI Astra is a different class on intelligence and wide work, and last on usage. Grok 4.6 is still the daily driver.
OpenAI Astra took intelligence and wide work, then made itself hard to live with. Grok 4.6 kept the hours. Same three dimensions, a split board.
Read article08/22/2026 · LLM Research
AI img2mesh usually means image-to-mesh generators. This is a different path: frontier LLMs author mesh source from a reference image inside a compile-and-compare harness.
AI img2mesh through code. Reference image in, DSL syntax out, repair up to six rounds, silhouette compare up to two. Fable 5, Opus 5, GPT Sol xhigh, and Grok 4.6 xhigh clear it.
Read article08/22/2026 · LLM Research
Late summer update to my coding LLM rankings. Same three dimensions as June, plus open source. Grok is now #1 on usage, intelligence, and wide work.
Two months after the summer edition, July's model wave reshuffled my board. Grok leads usage, intelligence, and wide work. Open source leadership moved to Qwen.
Read article06/27/2026 · LLM Research
Updated practical ranking of coding LLM tools three months later. Same three dimensions, some shifts in the answers based on more real work.
Spring had one set of tradeoffs. Summer is showing different ones in practice. The three questions have not changed much. The answers have.
Read article05/28/2026 · LLM Research
A proposal for Squire: a local, context-engineered awareness service that runs inside apps, uses CLI tools as capabilities, and prepares useful work without taking over user judgment.
Chatbots move the user out of the workflow. Squire moves the model into the workflow: a local awareness layer that understands the app, reads its structures, and prepares the next useful thing.
Read article05/10/2026 · LLM Research
A practical first-person guide to using LLMs in game development, covering code, assets, prototyping, debugging, multiplayer, tools, and workflow management.
LLM-assisted game development is no longer just autocomplete. This guide shows what LLMs can realistically add to a game dev workflow today, from design docs and gameplay code to assets, debugging, performance, and multi-agent workflows.
Read article05/06/2026 · LLM Research
A practical baseline workflow for using LLMs and coding agents without losing control of the work: specs, examples, clarification, diff review, runtime QA, and earned commits.
Prompt-and-pray feels productive because the model moves fast. A better workflow keeps the human in charge of taste, context, review, and runtime behavior.
Read article05/05/2026 · Software Culture
A practical argument for software that respects users, keeps data portable, prices itself honestly, and stops turning every workflow into a funnel.
Deshitification is not anti-business. It is anti-extraction: software should fit the user, respect their data, and make leaving as easy as arriving.
Read article04/27/2026 · LLM Research
A practical ranking of coding LLM tools in 2026 by usage allowance, state-of-the-art intelligence, and wide problem complexity handling.
The best LLM for coding depends on the job: daily usage volume, top-end intelligence, or wide multi-file problem solving. This ranks the current stack by those three dimensions.
Read article04/26/2026 · LLM Research
High-end tools still lead on polish and stability, but modern GPUs, open knowledge, and fast prototyping are making much of that advantage reusable by smaller teams in far less time.
A field note on how speed and accessibility are eroding software moats in 3D, tooling, and beyond.
Read article04/17/2026 · LLM Research
A visual map of how software value changes as LLM capability expands outward from simple CRUD into finance apps, vertical SaaS, games, infrastructure, embedded systems, and expert software.
An interactive star map for tracking which software categories are becoming easier to produce with LLMs, and which still hold value through judgment, context, and expertise.
Read article04/17/2026 · LLM Research
LLM operating harnesses and LLM agent orchestration are not autonomous companies yet. They are early coordination tools for founders, managers, and small teams.
A practical look at coding harnesses, company harnesses, agent orchestration, and why the useful near-term shape is a hybrid operating layer for humans.
Read article04/13/2026 · Projects
How I got a multiplayer Three.js game to load instantly in the browser. Asset pipelines, meshopt compression, texture atlases, and streaming GLBs without a loading screen.
The asset pipeline behind Mouse Trouble: meshopt compression, texture atlases, streaming GLBs, audio optimization, and browser-first Three.js performance.
Read article04/12/2026 · LLM Research
A timeline of experimenting with LLM tools from ChatGPT through agent maxxing. What worked, what was overhyped, and how my thinking changed along the way.
A field report from years of testing ChatGPT, Claude, Cursor, aider, local models, and coding agents against real software projects.
Read article04/12/2026 · LLM Research
Cursor's game jam requires 90% LLM written code. That constraint makes it a useful benchmark for what agentic development can and can't do right now.
Cursor's LLM game jam is a useful pressure test for agentic development, architecture decisions, and what happens when generated code has to become a playable product.
Read article01/01/2026 · Architecture Notes
Who I am, what 17 years of building enterprise and consumer software taught me, and why I started writing about LLM-assisted development.
A starting point for writing about systems architecture, full stack software, and where LLM-assisted development helps or creates risk.
Read articleNo posts match. Clear filters to see everything.
