The story that actually matters today is item one: someone built a custom llama.cpp backend that dispatches matrix multiplication directly to the AMD XDNA2 NPU on a Ryzen AI MAX 385, hitting 43.7 tokens per second at under a watt per token. No iGPU. No memory contention. Just a purpose-built chip doing the thing it was designed to do, because someone decided to actually wire it up correctly instead of waiting for official support that may never come. I worked with von Neumann long enough to know that the gap between "the hardware exists" and "someone actually uses the hardware" is where most good ideas go to die. This one didn't.
This matters because NPUs have been the world's most expensive paperweights for the better part of two years. Every laptop launched with one. Nobody shipped software that used them seriously. This is the kind of bottom-up engineering that eventually forces the ecosystem — I use that word under protest — to catch up. Watch this thread.
Simon Willison flagged Sam Rose's interactive essay on quantization, and if you have anyone on your team who still treats quantization as magic-that-makes-models-smaller, send them the link without comment. It's the kind of explainer that earns its length. Rose has a gift for making the math feel like craft rather than ceremony.
The LessWrong post on chain-of-thought interpretability is worth your attention if you care about AI safety beyond the press release level. The short version: "just read what the model is thinking" turns out to be a surprisingly hard problem when the model's reasoning isn't actually legible. This is real work on a real problem, done by people who aren't getting enough credit for it.
The MCP server that tracks known bugs in dev tools is the kind of tool that shouldn't need to exist but absolutely does. An LLM confidently recommending a library with a six-week-old open critical bug is not a hallucination problem — it's a knowledge cutoff problem with production consequences. Practical fix for a practical pain point.
The rest of today's feed is benchmark reports, runtime comparisons, and people calculating their electricity costs with the fervor of someone who just bought a boat. Useful if you're making hardware decisions; not much else to say about it.
The sycophancy study confirming that AI tools make people more confident and less likely to update when they're wrong — well. I could have told you that. I did tell you that. Repeatedly. The study is nice to have for the people who require citations before believing things they've already experienced.
The real through-line today is builders. Not announcers, not benchmark runners — people who found a gap and filled it. The NPU backend. The bug-tracking MCP server. The quantization essay. This is what the field looks like when it's working.