Writing
Notes from the workbench
AI, local LLMs, cloud architecture, and developer tools — benchmarked, built, and broken so you don't have to.
Gemini 4 Argon Built the Eiffel Tower and Paris in 3D From 8 Images
Gemini 4 Argon built a walkable 3D Paris around the Eiffel Tower in one night: $48.97 billed, 3 h 24 min of agent work, 159 prompts. The making-of.
Gemini 4 Argon vs Opus 5.5 vs GPT-6 Astra: Same Prompt, Real Bills
Gemini 4 Argon built a 3D car in Blender from one picture ($22.13), wrote 2,540 lines for one Lamborghini, and won a one-shot kart game 2.3x cheaper than Astra.
EmbeddingGemma 2 Review: 30/30 vs Mac Spotlight 5/30, Offline Photo Search
EmbeddingGemma 2 found the right file 30/30 times on a Mac, Spotlight 5/30. Plus offline face search by name: 114/121 found, 0 wrong. Free app and code.
Microsoft Decision-1 vs Every Decision Model on OpenRouter: 6th of 14
I ran 82 hand-labelled questions through every decision model on OpenRouter and 4 on a Mac. Microsoft Decision-1 got 75/82, 6th of 14, at $0.0000036 a question.
Nano Banana 2.1 vs GPT Image 2.5: 8 Prompts, Real Cost per Image
Nano Banana 2.1 cost me $0.042 a picture, GPT Image 2.5 $0.033. Same 8 prompts, 5 models, first picture only: OpenAI won 3 to 2. Prompts, costs, code.
Claude Sonnet 5.5 Built a Kart Racer for My TV: 4 Players, Gamepads
Turbo Kartz: a split-screen kart racer built in Claude Code. 4 tracks, jumps, gamepads on a Sony TV and in the browser. Every prompt inside.
I Made a Need for Speed Style Game With Claude Opus 5.5: Every Prompt
Kartit Nitro, a browser racer built with Claude Opus 5.5: the 3D car cost $0 on a free Kaggle GPU, and $0.64 of AI music ships in the game.
Free AI Image-to-3D on a Kaggle T4: a TRELLIS Car in a three.js Racer
TRELLIS image to 3D on a free Kaggle T4: 39.4 s to a textured car with 12,444 faces, cut to a 300 KB GLB and raced in a three.js game you can play here.
Claude Opus 5.5: Pricing, Benchmarks vs GPT-6 Astra, Banked Reset and Breaking Changes
Claude Opus 5.5 costs $4/$20 per million tokens, 40% less than Opus 5, and leads Terminal-Bench 4.0 at 66.4%. Full model card, plan changes, the banked rate-limit reset and 4 breaking API changes.
Qwen3.8-27B on a 24GB Mac: 3 Kernel Panics and the 17.76 GiB Limit Apple Doesn't Print
Your Mac has 24GB of RAM. Metal reports 17.76 GiB as the amount the GPU can use without hurting performance. Comparing a GGUF file size against total RAM is the wrong arithmetic, and that gap is where the kernel panics live. Measured on one M4 Pro, including the runs that failed.
Gemma 4 26B Won't Fit on My 24GB MacBook — Until I Did This
Ollama gives 2 tok/s with broken tool calling. I got 49 tok/s with perfect tool calling using Unsloth Q3_K_XL + llama.cpp. Then I built a Claude Code clone on top of it.
I Tested Every Gemma 4 Model Locally on My MacBook — What Actually Works
Audio ASR in 3 languages, image understanding, full-stack app generation, coding, and agentic behavior -- all on a MacBook M4 Pro 24GB.
The Ralph Wiggum Technique: Autonomous AI Development with Claude Code
Learn how to use the Ralph Wiggum technique for autonomous AI-powered coding. Install the Ralph plugin for Claude Code and let your AI write code while you sleep.
7 Best AI Coding Tools Compared: WindSURF, Cursor, Bolt New, Cline, Roocline, GitHub Copilot, and Replit
Comprehensive comparison of WindSURF, Cursor, Bolt New, Cline, Roocline, GitHub Copilot, and Replit. Learn which AI coding tool best fits your needs in 2025.