Hard Stops
Simon Willison wants budget caps that actually stop, Wikimedia counted what the agent swarm did to its wikis, and a GNOME maintainer argues for AI-found bug reports. Anthropic cut Haiku to a tenth of the price, OpenAI shipped a decisions API that can say no, and Mistral previewed a trillion-parameter model. The star is Strands Decider, a 2B model that vets tool calls before they run.
The tools
Strands Decider 2B models open source
A small open model that makes typed decisions — pick one, yes or no, rate it — with a confidence score, in one forward pass.
281 points on HN. Its best use: vetting an agent's tool call before it runs, for one cheap forward pass instead of another LLM call.
tool page →Whistle audio open source
Speech-to-text in 16.9 MB: seven languages, word timestamps, runs on a plain CPU with no dependencies.
270 points on HN. Small enough to ship inside an app, so the audio never leaves the device.
tool page →SCM (Screen Memories) image-video open source
Search every photo and every video frame on your Mac in plain language — by scene, on-screen text, or spoken line. Fully local.
179 points as a Show HN. A neat local-first stack built from parts you already know.
tool page →openTPU research open source
An open AI accelerator — hardware design, simulator, and compiler — running real small models on an FPGA card.
340 points on HN. Asks how far AI agents can go at chip design, and teaches you how an accelerator works along the way.
tool page →Worth reading
Simon Willison: agents make runaway bills easy, so caps should stop spending and return errors — on by default, not as a warning email.
The swarm from the operator's side: unauthorised edits, failed exploit attempts, and millions of requests that may have contributed to a Wikidata outage.
GNOME's CVEs went from 21 in 2021 to 141 this year. Its security maintainer argues that banning AI-generated bug reports now means banning most vulnerability reports.
From the big labs
A tenth of Haiku 4.5's price ($0.10 / $0.50 per M tokens) and smarter than GPT-6 Luna on Artificial Analysis — but very wordy, so watch per-task cost.
A new endpoint that returns a probability, a choice, or a score instead of prose — with 'refuse' as an explicit answer. Public beta, $0.10 per M input tokens.
A trillion-parameter model pitched on refusing less for security work. API preview now; weights promised by the end of the month, licence not yet announced.