Decision models are rapidly standardizing across the open inference stack: llama.cpp shipped a /v1/systemone endpoint for decision models, Ollama 0.35 followed with the same API, and SGLang v0.5.21 added a native Decisions API that turns LLMs/VLMs into low-latency classifiers and scorers. A common request/response shape emerging across three major runtimes is a real de facto standard for typed, probabilistic decisions — a new layer agents can call for routing, verification, and scoring without fine-tuning.
Meta pushed its Muse personal agent into physical hardware, open-sourcing ESP32 firmware and a Linux SDK so anyone can build gadgets that plug into Muse, alongside the Muse Home Link USB-C dongle (the 5,000-unit figure is a media report, not Meta-confirmed). Anthropic's new "Claude-shaped science" essay argues frontier models are still badly matched to how scientists actually work, and the model-consciousness debate spilled into the open: Sam Altman called ascribing "religious force" to models a real safety issue, while François Chollet warned that treating model sentience and quantified suffering as settled could lead somewhere dark.
On the model front, GPT-6.1 Sol (Max) debuted at #5 in Agent Arena and reshaped the cost-performance Pareto frontier at a $0.56 median cost per task. NVIDIA announced a 64GB DGX Spark desktop configuration shipping October 23, widening the local-AI memory ceiling.