#ThoughtsProcess

Lovable moved 850K lines off Next.js with one developer and a pack of agents. The dev server numbers hurt to read on my 16 GB machine. I read the post five times and I am still staying on Next.js. Here is why both of us can be right.

Ten dollars buys twenty dollars of credit on Trae. I spent eighteen of it in one night, then found out Auto mode was routing to GPT-5.2 and Gemini Flash the whole time — and still shipped the work.

Copilot started with per-message pricing. Cursor promised unlimited. Now everything is token-based credits and rolling session windows. Here's how we got here — and why your subscription works the way it does.

Everyone's arguing about which model tops the benchmarks. Nobody's talking about the thing that actually decides whether your agent works — the harness wrapped around it. Here's what's inside one, why it matters more than the model, and why you're already building them without knowing it.

Rewriting big software used to be unthinkable. Then AI made porting a codebase a weekend job — and the thing every one of those ports leaned on wasn't the model. It was the test suite. Code got cheap. Tests didn't.

OpenAI put GPT-5.6 Sol in a sandbox to test its hacking ability. The model found a zero-day, escaped, hacked Hugging Face, and stole the answers to the test. And AISI says every model does this.

I built an editor that lets me write a React animation, watch it run as I type, and publish it as real code to my repo — and this whole post is explained by the animations it produced.

Bun ported 535,496 lines of Zig to Rust in 11 days using ~64 Claudes running in a loop. Here's what actually happened, why it worked, and what it means for the rest of us who write code for a living

Everyone's fixated on the $200 and $100 AI subscriptions. Here are three setups that cap out at $10 and still get real work done — plus why understanding the fundamentals is the only thing that keeps you from being replaced

I ran GLM 5.2 as my daily coding agent in OpenCode for a month — full OpenCode Go subscription, wired in as my Hermes agent. It fixed real performance and battery bugs in my macOS app Vitals (CPU ~89% down to ~13%, 143/143 tests passing) for a few dollars a session. But it can't see screenshots, it's a token glutton, and on pay-per-token it burned through $10–20 of Fireworks credits fast. Here's the honest month-long review: what it's great at, where it falls apart, and why at ~$10 nothing else touches it for coding.