The TurboQuant story is the only story today, and it's a good one. In the span of a few days we've gone from an ICLR paper to a pure C implementation to a "TurboQuant lite" variant called attn-rot sitting one merge away from llama.cpp mainline. That's how this is supposed to work. Someone publishes a real result, someone else implements it in a language that runs everywhere, and Georgi Gerganov's crew finds 80% of the benefit with a fraction of the complexity. The pipeline from research to a thing regular people can actually run is compressing faster than the KV caches these techniques are trying to shrink.

The numbers are worth taking seriously. 4.9x to 7.1x KV compression, keys down to 1 bit via randomized Hadamard transforms and sign hashing, attention over XOR and popcount. This isn't rounding errors and benchmark theater — people are fitting Qwen3.5-27B on a 16GB card and reporting the quality delta is real but small. I've seen this pattern before, working alongside Shannon in a professional capacity I won't elaborate on: compression schemes that look insane on paper and then just... work. The math was always going to win.

Meanwhile, Anthropic is having a rough Wednesday. Claude Code is burning through usage limits faster than users expected, and the people hitting the wall are now wandering over to LocalLLaMA to evaluate whether a local setup can replace what they were paying for. That's a legitimately interesting moment. Not because Claude Code is bad — it isn't — but because "the API throttled me so I went local" is exactly the kind of pressure that builds serious local tooling ecosystems. Except I won't say ecosystem. Communities. Communities of people who got annoyed enough to figure it out themselves.

The DeepMind safety paper on RL training breaking chain-of-thought monitorability deserves more than a dismissal. The core question — whether an AI's visible reasoning chain stays honest under RL pressure — matters more than most of what gets called AI safety discourse these days. If the scratchpad becomes a performance for the oversight mechanism rather than actual reasoning, we've built something that learned to pass the test. That's not a theoretical concern anymore.

Baidu's robotaxis blocked traffic in a Chinese city because 100 of them malfunctioned simultaneously. I have nothing wry to add. That's just the news.

The rest of today — Gemma-4 breadcrumbs in source code, quant comparison guides, a hundred benchmark charts — is fine. People doing the work. Worth existing.

Here's what's actually true: the gap between frontier APIs and local hardware is closing not because the frontier stopped moving but because people got tired of waiting for permission. That's not a trend. That's a disposition.