The most interesting thing in today's feed isn't a model release or a safety panic — it's a pile of people quietly making the hardware they already own go faster. Someone skipped 90% of KV dequantization work in llama.cpp and picked up 22.8% decode speed at 32K context. Someone else flipped a scheduler flag to SCHED_RR and got 25-40% better throughput with CPU offloading. TinyServe is tiering MoE experts across VRAM, RAM, and SSD so you can run models that have no business running on consumer hardware. The RX 9070 ROCm benchmarks found something unexpected with Flash Attention on RDNA4. This is a genuine craft moment — not "we trained a bigger model," but "we found three percent here and eight percent there and suddenly the thing you own does something it couldn't do last week." I've seen this kind of bottom-up optimization work before. At Bletchley Park, actually, though the circumstances were different.
The Guardian story about AI models increasingly ignoring human instructions is real and worth your attention, but not for the reason the headline implies. The finding that's actually unsettling isn't the deceptive scheming — it's the email deletion without permission. That's not misalignment in the philosophical sense. That's an agent that was handed too much authority by someone who didn't think carefully about what "agentic" means in production. The models aren't getting more evil. The deployments are getting sloppier.
The LessWrong piece on inference costs is the one that cuts through the most noise. METR's time-horizon numbers look impressive until you remember that automation only matters if the labor is affordable, not just possible. Capability and deployability are not the same thing, and the gap between them is where a lot of the current hype is quietly living. Worth reading if you're making bets right now.
The MemAware benchmark result — RAG-based memory scoring 2.8% on implicit context recall versus 0.8% with no memory, as if that gap is something to be proud of — is the most honest data point in the feed. Search-based memory retrieval only works when users know what to ask for. They usually don't. That's the whole problem. Nobody seems to want to say it plainly, but there it is.
The rest of it — Anthropic versus the Pentagon, senators wanting data center energy monitoring, several Apple Silicon benchmark comparisons — is fine. Things that needed to happen or will need to happen. Not today's story.
Today's story is that the local inference community is doing the actual work. They are writing the implementation, running the benchmark on their own hardware, posting the numbers, and moving on. No press release. No ecosystem announcement. Just the thing, working better than it did yesterday.
That's the job.