A new SQL parser for PostHog just smashed performance records. Built with AI in a few weeks, it outpaces the old C++ ANTLR version by 70x. The team fed property-based tests into prompts, had the model generate and tweak code, then stitched it into a recursive-descent structure that speaks the same query language as before. Now SQL queries run in a fraction of the time without changing how users write them. Linear is ditching styled-components for Facebook’s StyleX to cut runtime overhead and lock down CSS contracts. Styled-components hasn’t kept pace with React’s latest features, so moving to StyleX means fewer style recalculations, faster page loads, and more predictable class merging. They’ve rewritten key UI modules, tracked bundle sizes, and already seen a drop in style-sheet overhead. Prompt injection gets a deep dive. It turns out LLMs blur the line between user instructions and system roles. Tags like “user:” or “assistant:” don’t stop a clever payload from slipping in commands. Models lean on tone, not tag hierarchy, so you can sneak in “ignore previous” or “execute this” inside a user block and bypass checks. The fix? Stronger role enforcement, token-level guards, and training that punishes style-based hijacks. On the tooling front, check out peerd and Nub. peerd is a browser extension for Chrome and Firefox that runs tasks locally through peer-to-peer sandboxes—no cloud needed. Nub brings Bun-style speed to Node.js, with a fast package manager and script runner that still obeys Node’s module rules. Gemini 3.5 Flash also added computer use hooks so agents can call APIs, run code or manage files. And AI labs are hiring philosophers to frame ethical logic with deontology and critical thinking, helping shape safer, more responsible models.