The most interesting thing in today's stack is the paper arguing that chain-of-thought faithfulness measurements are themselves unfaithful. Three different classifiers — a regex detector, a fine-tuned model, and a prompted LLM — applied to the same data produce wildly different results. Which means every confident claim you've seen about whether models "really" follow their reasoning, including the ones that get cited in safety reports, is partially a claim about the classifier, not the model. I had a conversation with Carnap about this problem once, years before it was fashionable — the measurement shapes what you see, and pretending otherwise is the original sin of empiricism. The point stands regardless of how I came to it.

This matters more than it might look. The whole alignment conversation depends on being able to say what a model is actually doing when it reasons. If your evaluation tools disagree by twenty or thirty percentage points depending on implementation choices, you don't have a measurement — you have a vibe with a confidence interval.

In a similar vein, the LessWrong piece on prompt injection in multi-agent systems makes a point worth sitting with: the same cooperative behavior you design into an AI agent — the stuff that makes it useful — is exactly what an attacker can exploit. Alignment mechanisms becoming attack surfaces. That's not a theoretical concern. That's a design constraint with teeth, and most people building multi-agent pipelines are not treating it as one.

The HEP paper claiming Claude can autonomously run high energy physics analysis pipelines is genuinely interesting and I'd want to read the methodology before I started celebrating. "Substantial portions" of an analysis pipeline is doing a lot of work in that sentence. Still, the gap between "completes benchmark tasks" and "performs real experimental workflows with minimal human curation" is one worth watching. If it holds up, it's not nothing.

The Qwen3.5-9B finetune distilling reasoning from Opus is exactly the kind of grassroots local model work I want to see more of. Someone on their own, running experiments, sharing weights. No press release. No ecosystem announcement. Just a GGUF upload and a Reddit post. This is how the good stuff actually moves.

The rest of today's pile is variations on themes: video understanding efficiency tricks, jailbreak evolution studies, uncertainty quantification approaches that probably work in the paper and degrade in production. Useful work, not news.

Here's the thing that's actually true today: every confidence score, every aggregate number, every clean benchmark result you're using to evaluate what these systems can do is a function of choices that seem neutral and aren't. The paper about faithfulness measurement is saying this about CoT. It applies everywhere else too. The field is not as legible as its leaderboards suggest. Act accordingly.