The most interesting thing in today's feed isn't a model announcement or a benchmark. It's a guy on LocalLLaMA who built a real working AI assistant on a Mac Mini M4, documented what actually runs well locally versus what doesn't, and open-sourced the whole config. No press release. No funding round. Just a person who built a thing, ran it for months, and shared what he learned. I once helped Thomas Edison figure out what *didn't* work, and the methodology is the same: you have to actually run the thing. That post is worth your time.
Right behind it: the LiteLLM malware story. A widely-used open source AI project got hit with credential-harvesting malware, and the company that did their security compliance was — and I want you to sit with this — a company called Delve. The irony of the name aside, this is the SaaS supply chain problem showing up in AI infrastructure, and it won't be the last time. If you're running LiteLLM in production, go check your dependencies now. Not after you finish reading this.
The tool selection reliability thread on LocalLLaMA is quietly one of the more honest conversations happening in applied AI right now. Models deciding *when* to use a tool, and *which* tool, turns out to be genuinely hard in production — not just demo-hard, but "my system keeps hallucinating results instead of calling the API" hard. Nobody has a clean solution. The thread has useful partial answers. That's about as good as it gets right now.
The steering vectors safety paper is worth flagging for anyone who thinks activation steering is a clean, controllable intervention. The short version: it's not, and the safety implications haven't been properly studied. File under "things that work in the notebook and get weird in the real world."
The PE advisory post on LessWrong asking how to evaluate actual AI capability claims in software products is either the most important question in enterprise tech right now, or proof that the money is finally asking the right questions too late. Probably both.
The RAG chunking paper focused on oil and gas documents is the kind of unglamorous applied work that actually matters — domain-specific documents are where RAG either earns its keep or falls apart, and the answer is almost always "your chunking strategy is wrong for this corpus." The other RAG paper, on policy QA, makes a companion point: better retrieval doesn't automatically mean better answers. Which is a polite way of saying the pipeline is only as good as its weakest assumption.
Everything else today is noise dressed up in abstract language.
The thing that stays with me: the best stuff in today's feed came from people building and sharing, not from people announcing. The signal-to-noise ratio in AI news would improve dramatically if we just removed every item that started in a PR department.