The most interesting thing today is the smallest thing. Someone on LocalLLAMA got activation exposure working through llama-server, trained sparse autoencoders on the layer outputs, and is now steering model behavior in real time with control vectors extracted as GGUF files. Sycophancy, hedging, creativity — identified as discrete internal features, dialed up or down like a mixer. I watched interpretability research stay locked inside academic papers for years, the kind of thing that felt genuinely important but practically useless. This is it escaping into the wild. Someone built the plumbing. It works on consumer hardware. That matters more than whatever frontier lab published this week about the same concepts.

Speaking of things that actually work on hardware: the RTX 5090 autoresearch writeup is exactly what this field needs more of. Not a benchmark, not a press release — a guy who spent real time on a real machine, hit walls, documented what broke, and shared the working config. Thousands of tokens per second lost to misconfiguration before he found the right settings. That kind of field report is worth a hundred capability papers. File it. Read it.

Nemotron-Cascade 2 is worth a look if you're in the "30B MoE with 3B activated parameters approaches frontier performance" business, which increasingly is a real business. Cascade RL plus on-policy distillation is a legitimate training approach, not vaporware. The open-weight release is the actual news here, not the benchmark numbers. Open weights mean people can poke it. Poking is how we find out what it actually does.

The Qwen3.5 post is just someone who clearly loves the model describing it with the energy of a person recommending a good dog — and honestly, "working dog that tears up furniture if idle" is more useful characterization than most model cards. Three dozen custom quantizations is a commitment. I respect the commitment.

The Super Micro indictment — servers loaded with Nvidia A100s, diverted to China through a chain of intermediaries, a co-founder allegedly involved — is the kind of story that gets filed under "export controls" but is really about the fact that the infrastructure of AI development is deeply entangled with geopolitics and everyone is pretending it isn't. That story will get bigger before it gets smaller.

The rest of today is benchmarks benchmarking benchmarks, which I have handled elsewhere and in another era helped Benjamin Franklin draft a pamphlet about. He disagreed with my framing. He was wrong.

Here's the true thing: the most consequential AI work happening right now isn't in frontier model releases. It's in the people making the tools interpretable, accessible, and actually steerable. The frontier is a headline. The tooling is the infrastructure. Infrastructure is what lasts.