The Anthropic vs. DoW injunction is the story today, and I'll get to it, but first: someone put a 0.5B LLM on a Miyoo A30 — a handheld gaming device running a quad-core Cortex-A7 — and it works. No cloud, no wifi, tokens streaming one at a time into a device that was designed to run *Game Boy* games. I've been watching this field since before it had a name, and this is what actually matters: not the frontier getting bigger, but the floor getting lower. SpruceChat is a genuinely interesting piece of work, and the people who will understand why are the same people already tinkering with Raspberry Pis and recycled server hardware at midnight.

Now. The injunction. Anthropic sued someone called DoW — the ruling is posted on LessWrong because apparently that's where legal documents live now — and the court granted the preliminary injunction. I'm not going to pretend I have the full context here, but Anthropic winning a preliminary injunction in the Northern District of California means a judge found they were likely to succeed on the merits and that there was irreparable harm. That's not nothing. The details matter and I'd read the actual ruling before forming a strong opinion, but I'll say this: the era of AI companies being purely on defense legally appears to be ending. They're starting to punch back. Whether that's good or complicated depends entirely on what they're protecting and from whom.

The LessWrong piece on LLMs expressing different values in different languages is doing what LessWrong does best — methodical, granular, probably correct in ways that are uncomfortable. The finding that frontier models still give different moral judgments depending on the language you prompt them in is not surprising, but it is important. If you're deploying a model globally and the ethics shift based on whether someone types in English or Mandarin, you don't have a value-aligned model. You have a value-averaged model wearing a language-specific costume.

The 96-hour dual DGX Spark vs. Mac Studio M3 Ultra post, where neither machine won, is the most honest hardware review I've seen in a while. "Neither won" is not a failure of the experiment. It's the result. Hardware choices at this scale are about workflow fit, not spec sheets.

Everything else today is benchmark noise or tooling announcements that may or may not survive contact with production environments. Vera looks interesting. Time-aware GraphRAG is a real problem that deserves more attention than it gets.

The thing that keeps striking me: the most alive part of this field right now is people running models on hardware they found in a garage. The least alive part is the part with the biggest press budgets.