Lost Bytes
Lost Bytes
Notes, experiments, and fragments I didn't want to lose.
A public notebook — topics evolve over time.
Start here
Entry points
Latest notes
-
Small Models, Serious Work
Why smaller language models are quietly reshaping AI in 2026. The most interesting trend is not scale. It is restraint.
Featured
-
Small Models, Serious Work
Why smaller language models are quietly reshaping AI in 2026. The most interesting trend is not scale. It is restraint.
-
Why Latin Isn't Called Roman
What the name of a language tells us about history and identity. Latin points to the people of Latium, not the empire that spread it.
-
Only One-Third of Universal Grammar Rules Survived Testing
Why language is both structured and unpredictable. A large empirical test found most proposed universal grammar rules did not hold across languages.
-
The Dark Factory Problem: Why AI Is Writing All the Code and Most Developers Are Still Slower
Why most developers are slower with AI, and why the future belongs to Dark Factories, not copilots.
-
Inside the Dark Factory: Where AI Ships Software and Humans Fall Behind
Most developers are slower with AI. The gap between vibe coding and real Dark Factories is the most important organizational gap in software.
-
The Web Is Forking in the Age of Agents
Agentic commerce, content for non-human readers, and agent search are being standardized in parallel. The existing web is growing a second interface — for agents.
-
Reverse-Engineering Your Software Architecture with Claude Code (So Claude Can Actually Help)
Use Claude Code to reverse-engineer your system into versioned flow docs, then use those flows as a source of truth for investigations and support.
-
Do JPEGs age?
JPEG files do not slowly degrade over time. What actually causes quality loss is recompression and workflow, not age.
-
Plan–Code–Execute: Building Agents That Create Their Own Tools
A tutorial on building PCE agents that create their own tools on demand — no massive tool registry, just plan, code, and execute.
-
Vercel's Agent Skills: The Missing Middle Layer in Agent Architectures
Agent Skills add a versioned, reusable execution layer between reasoning and chaos — and they matter more than they look.
-
I Put Claude Code Inside Obsidian — and It Changed How I Use My Notes
Embedding Claude Code inside my Obsidian vault gave the AI real filesystem access. No more copy-paste — an agent that lives inside my notes.