Crazy how much uReview has changed since our talk at AI Engineer Worlds Fair last month. Our harness, models, tool calls and evals have all seen significant revamps. Sol running on Pydantic replaced Claude Code + Opus.
Security Researcher 1: there is a vulnerability in some package you ported to Python 3 over 10 years ago and hasn’t been touched in years
Me: sorry, I don’t support it any more, your company can sign a maintenance contract if you want
SR1: You should jump through these hoops
Caring about low level details such as these is a big part of the latency and performance differences users experience with tools like Sublime Text vs VS Code.
Minimizing allocations, using memory arenas, being smart about memory lifetime, struct packing, (C++) move semantics,
If you're interested in "how" or "why": the major culprit is that each row in Alacritty has 32 bytes of metadata and each cell is 24 bytes. In Ghostty, every row and cell is represented by exactly 8 byte each. How do we do this?
The first major culprit is styles. Alacritty
You should be using LLMs to write more deterministic development scripts and helpers, and then have the LLMs make them faster.
Skills are the crappy, slow, expensive version of this.
