The week's usable signal is mostly maintenance work: version drift under an agent tool, a negotiable line item, a dependency nobody reads, and a stack of compression claims nobody's benchmarked.
Most of today's upstream signal was local-inference tinkering: three separate projects racing to fit large models onto small hardware, a Mac port landing two days after a Chinese lab's release, a cheap-GPU fine-tuning trick. Real work, mostly for people with a spare graphics card and a Saturday. Four items cleared the bar for someone who buys AI rather than builds it.
A model upgrade broke a working agent (procurement).
Steve Yegge's agent setup worked well through Opus 4.6 and fell apart on 4.7. Simon Willison flagged it as exactly the kind of failure buyers never test for. The scaffolding around a model, the part that decides what step comes next and what to do when a step fails, had encoded assumptions about how one specific model version plans and recovers. A better model invalidated them. That is the vendor question sitting under every agent product a team buys and almost nobody asks during the demo: whether the model version is pinned, whether upgrades arrive with notice, and whether anyone on the buyer's side finds the regression before a customer does. Most agent products swap the underlying model silently and call it an improvement.
A moderation invoice just got a substitute.
Mistral released Shieldstral, a small open-weights model that screens text and images for the same categories a hosted moderation API checks. It is free to download and small enough to sit on modest hardware. The per-call moderation charge on a monthly AI bill now has a public alternative next to it. That does not make the swap free, since someone has to stand it up and keep it current, and a self-hosted classifier has no vendor to call when it misfires. What it changes is the conversation with a moderation vendor, from what is your rate to what am I paying for that a 3B model already handles. Line items negotiate better when a substitute exists in public.
A worm reaches the packages nobody reviews.
Aikido documented an active attack on npm that compromised Keyv, a package almost nobody installs on purpose and almost everybody ends up with, sitting a few layers down in the dependency tree of common Node tooling, AI SDKs included. Two weeks ago this digest treated agent permission scope as the procurement question. The scope that matters here is narrower and gets discussed less: what a coding agent is allowed to add. Agent-written code pulls in packages at machine speed, and the pause where an unfamiliar package name used to catch a human eye is exactly the pause agents remove. The exposure is not new. The rate is.
Six compression claims, zero quality comparisons (evaluation).
Six separate projects surfaced today claiming to shrink large models onto small hardware. Swiftlet claims an 80B Qwen model in 4.3GB on a Mac and a 35B on an iPhone. Another claims Gemma 4 in 500MB. Not one published the number that decides whether any of it is usable: how much worse the shrunk model answers than the original. Last week's benchmark audit showed that a published score can be wrong. This is the earlier version of that problem, where no comparison was run at all. On-device AI will show up in vendor pitches this quarter carrying a size number and no quality number, because the size number is the cheap one to produce.
Three of these four describe something already paid for behaving differently than it did at purchase: the model version under an agent tool, the packages under a codebase, the moderation invoice. The first one is checkable this week without touching anything. Which model version sits behind the agent tools already in use, and who gets told when it moves.