devotd.
Recent developments, checked against their sources.
Black Forest Labs has released weights for a 5.6 billion-parameter image generator that handles both generation and editing. Self-hosting requires checking the memory cost of a separate prompt expansion model and the scope of conditional commercial use.
Code can now read annotations supplied by external tools and format execution errors from eligible functions through a shared handler. Automatic history compaction is deferred while approval is pending. Existing error handlers need checks for the new error kind.
The update adds input-only decorations and standard attributes for logging generative AI work. The advisor requires opt-in. Shared logs now contain more information for reproducing requests and warrant review.
The terminal coding tool adds a settings search panel and support for follow-up instructions during responses. Saving settings differs from applying them, and interrupting work does not guarantee automatic rollback.
GitHub Copilot plans to introduce automatic device–cloud routing by late October 2026. Developers must assess execution location and command permissions separately. Device measurements and multi-model research evaluations use different conditions.
EmbeddingGemma 2 can load only the input encoders it needs and shorten retrieval vectors. The first choice affects model memory; the second affects storage and search quality. Google reports substantial quality loss for image, video and audio retrieval at 128 dimensions.
Ai2 replaced a priority-based scheduler with GPU-time budgets, hierarchical fair-share and a minimum-runtime contract. The reported 98% figures are cluster occupancy and delivery of owed time; whether useful computation survives preemption is a separate question.
LightOn released LightOnOCR-3 on October 8. The grounding prompt adds region labels, 0–1000 normalized coordinates, image descriptions and chart tables, and it changes output-token and rendering-resolution budgets as well.
On October 7 Microsoft made Microsoft Execution Containers (MXC) generally available to enforce access policy outside AI agents. Learning mode records while keeping denials, but permissive mode actually allows what it records, so the two modes produce different kinds of evidence.
Cohere’s Embed 5 Pro and Fast, released September 30, share an embedding space. With matching output dimensions, an index built with Pro can be queried by either model, so changing the query model does not require rebuilding the index.