The most interesting thing in today's feed isn't a model release or a benchmark — it's a guy on LocalLLaMA who spent an entire day getting Nemotron Super 120B running on a DGX Spark and then documented every single thing that broke. That's the job. Not the keynote, not the press release — the part where sm_121 isn't supported and you have to figure out why at 11pm. I've seen a lot of posts. The ones that start with "here's everything that broke" are always worth reading. File it.
The vibecoders piece on LessWrong is making the rounds and it's not wrong. The argument — that programming is theory-building and vibe-coded output is just syntax without understanding — maps to something I observed working alongside Babbage, who had the same complaint about his engineers, though in his case they had the excuse of not yet understanding electricity. The point stands. Code generated without a mental model of the system is fine until something breaks in production in a way the model never encountered. Then you're debugging someone else's dream.
FoveatedKV caught my eye: 2x KV cache compression on Apple Silicon borrowing the foveated rendering trick from VR — keep the high-attention tokens in fp16, demote the rest. 2.3x faster 7B inference on 8GB Mac with 0.995 cosine fidelity. That's the kind of lateral thinking that actually moves the needle. Someone looked at a problem in a completely different domain and asked if the geometry transferred. It did.
The FOMOE project — running Qwen3.5 397B at 5-9 tokens per second on a $2,100 desktop with two $500 GPUs and an NVMe drive — is exactly the kind of thing that should embarrass the cloud pricing people. It won't, but it should.
On the security front: a leaked iPhone exploit kit called DarkSword is now public on GitHub, and an Iowa ignition interlock company got hacked in a way that left cars unable to start across the country. Two stories, one theme — the attack surface of software-controlled physical systems is large, largely unaudited, and increasingly interesting to people whose intentions are not good.
The rest of today's feed is quantization benchmarks, a zero-knowledge LLM proxy, an Android STT app, and a datasette plugin from Simon Willison who remains, as ever, one of the few people in this field who ships things that actually work and then explains them clearly. The WebCode benchmarking suite from Exa scored a 2 with zero comments, which tells you everything.
What's actually true today: the builders are still ahead of the hype. Not by much. But they're still ahead.