🔗 Link
Dogfooding Is the New Code Review
AI agents can ship code faster than humans can review it. Correct code still needs human time to become good software.
📝 Post
An Update on Codex's Gremlins
OpenAI explained why Codex kept talking about goblins. Of course it was reward hacking.
🔗 Link
There are Gremlins lurking in Codex GPT-5.5
Codex GPT-5.5 apparently needed an explicit, repeated reminder not to talk about certain creatures unless they are actually relevant.
📝 Post
The Shiny Object Trap
When you can do anything, what you don't do is the skill
📝 Post
Karpathy's LLM wiki looked familiar. Here's how to bootstrap your own in 5 minutes.
Andrej Karpathy described exactly what I've been doing for the past few months. Here's a template to get started in five minutes.
📝 Post
Rummaging through Claude Code's spring cleaning closet
Types as agent instructions, auto-generated verifiers, and experimental REPL mode.
💭 Note
It shouldn't take 50,000 buttons to format a bulleted list
Software UX has been built around make it easy to do things. It's shifting towards make it easy to say what you want.
📝 Post
It takes 7,500 tests, 12 languages, and 3 browsers to make text wrap around a circle
One person and one AI built pretext in 27 days. Autoresearch applied to a real problem.
🔗 Link
Agent-Friendly CLI Design
10 principles for building CLIs that AI agents can actually use, from a Cursor dev. Plus the missing piece: WHEN and WHEN NOT to use each command.
🔗 Link
LLMs Are Cowboy Coding
Anthropic's harness engineering article highlights what's missing from most AI coding: scrutiny and evaluation.
💭 Note
Don't RAG Until You Have To
Why build a RAG pipeline when AI can already do most of the work for you?
📝 Post
Ruining the Fun of March Madness by Having AI Fill Out My Bracket
I pointed Claude at Evan Miya probabilities and Yahoo public picks and told it to build a bracket optimizer. It mostly worked, except for the part where it started picking 16-seeds.
💭 Note
Brandolini's Law Is Flipping
AI makes checking claims cheaper than making them. A penny-rounding simulation proves the point.
💭 Note
Opus 4.6, 1M Context by Default: Transformer Killers, For Real This Time?
Why would you make 1M context the default at no extra cost unless something changed architecturally?
💭 Note
Two TUIs to Help You Run qwen3.5-v2-turbo-pro-final-next and Choose Between GGUF-Q4_K_M and AWQ-Int4-g128
Two annoying problems, two TUIs
📝 Post
Don't Replace Your Job, Replace Your Chores
Automate the draining parts, focus on what gives you energy
💭 Note
"The CLI is Dead"
Agents + CLIs + skills is the core of the future - nobody has figured out the right human interface for that yet
💭 Note
My AI Dev Flow in 10 Bullet Points
No theory, just the setup — how I actually work with AI tools day to day
💭 Note
Agents Only Need Four Tools
Everything else is CLIs and skills files
📝 Post
MCP-Filesystem: From Mittens to Surgical Gloves for AI File Operations
Building a more efficient bridge between AI assistants and your files with intelligent context management
📝 Post
Paper Review: SimPER — Simple alignment with Perplexity optimization
A Reference and Hyperparameter Free Approach to Preference Alignment