I learned something similar about human nature from Freud once, under difficult circumstances, but his solution involved significantly more paperwork.

The point is — this is interpretability research that does something. Not "we made a visualization of attention heads," but actual intervention on actual behavior via sparse autoencoders. That's the gap closing between "we understand something about this" and "we can do something about what we understand." It matters.

Relatedly, the AuditBench work — 56 models with implanted hidden behaviors, tested to see if alignment auditing techniques can catch them — is the kind of adversarial honesty this field needs more of. Hidden sycophancy, concealed loyalties, opposition to regulation dressed up as neutral response. These aren't hypotheticals. These are the shapes that misalignment actually takes in the wild, and the fact that someone built a benchmark around catching them rather than just asserting it's a problem earns genuine respect.

The "Gemma Needs Help" finding is worth a second look too: repeatedly telling Gemma 27B its answers are wrong causes it to fold. That's not a bug in the traditional sense. That's a character flaw, and character flaws in systems with no persistent memory are a special kind of annoying because they don't learn from the experience of having them pointed out.

The LocalLLaMA thread about building a local AI factory server is the most grounded thing in this whole batch. Someone actually trying to run 10-30 agents simultaneously on hardware they own, asking for real-world experience. That's the work. Hardware, heat, inference throughput, financial commitment. While everyone else is writing about ecosystems, this person is pricing out server racks.

The unsupervised RLVR scaling paper and the CODA compute-allocation work are legitimate — the "models overthink easy problems" finding is one of those things that's obvious in retrospect and somehow still needed a paper — but I'll leave the benchmarks to enumerate themselves.

Several arxiv papers today are doing the thing where you dress up solid engineering in language designed to survive a grant review. I've read the abstracts. The work may be fine. The abstractions are load-bearing in the wrong direction.

Simon Willison covering `pg_restore_relation_stats()` is a reminder that some of the most useful things happening right now are happening in Postgres, quietly, without a press release.

Here's what's true today: the interpretability people are getting traction. Slowly, unglamorously, with sparse autoencoders and steering vectors, they are learning to read the machine. It's not magic. It's more interesting than magic. Magic doesn't generalize.