Before LLMs, the recurring question was "should I use an existing package or code it myself?" Claude Code made that question disappear. But it got replaced by a harder one: what is worth building at all? My answer is building blocks. Pieces that fade into the background, with boundaries so clear that knowing when to use them is instinct.
We are being sold LLMs as incredible productivity tools, and the more I hear that pitch, the more I think about Michael Polanyi.
When writing code is cheap, reusing existing systems could be a sunk-cost fallacy
Every major technology pulls apart things we used to treat as one thing. Writing split knowledge from memory. Batteries split energy from its source. GPS split getting somewhere from knowing the way. LLMs are doing this to intelligence and identity.
Technology is a separator, impactful technologies separate a lot of things.
Claude Code outperforms competitors using the same models. The secret might be privileged access to unrestricted model configurations that public APIs are too constrained to provide.
Tools like Cursor and Claude Code are not the end game of using LLMs for coding. What is missing?
How can we define intelligence, and how do LLMs fit in this definition
OAuth flows, Code editors, and some thoughts on authorship on the LLM era.
Deep dive focusing on security, cloud infrastructure and developer tools
Trying out webstorm integration and comparing it with react
Just saying hi and explaining what you can expect from this blog.