Jul 12, 2026

This week’s engineering news was dominated by AI inference breakthroughs and the quiet bugs they exposed. EdgeSync-LLM used llama.cpp’s KV-cache reuse to cut time-to-first-token by 9.9× on a real Android phone, while Colibri booted GLM-5.2’s 744B MoE on 25 GB of RAM with expert streaming and a 57× KV-cache shrink. Under the hood, a long‑overlooked P100 CUDA precision bug in llama.cpp was finally squashed, tightening KL‑divergence 2,300‑fold.

Beyond raw speed, testing whether AI genuinely understands a codebase got a practical answer: a change‑impact teardown audit that catches silent regressions. In materials science, a new CLI tool runs four falsification tests to detect models that cheat by learning bibliographic metadata instead of chemical structure.

Developer infrastructure revealed its own cracks. The widely reused LiteLLM pricing table relied on just `jq empty` for validation, leaving thousands of model rates unchecked. Claude Code’s shared prompt cache vanished when launched from a standalone terminal, while VS Code‑based launches hit it consistently. On the productivity front, a PHP DTO compiler hit 4.5 million hydrations per second using lazy ghosts, React 19 formalized which values cross the server‑component boundary (Date safe, class instances rejected), and a self‑hosted RAG stack gave an end‑to‑end blueprint from parsing to knowledge‑graph extraction.