AI Daily Digest · 2026-11-14
🔬 New AI Craft
1. Safety evals need conversation-level criteria and calibrated graders
Anthropic describes risk-specific rubrics that assess an entire conversation, alongside paired harmful and benign prompts. Its practical recommendations include checking graders against human labels, preregistering analyses, and reporting uncertainty; refusal frequency alone does not establish safer behavior.
https://www.anthropic.com/research/building-strong-safety-evaluations
2. Train multi-vector encoders for the pooling budget used at retrieval time
A Sentence Transformers walkthrough combines token merging with pooling-aware encoder training. In the author's NanoBEIR experiment, a fourfold reduction in document vectors retained 97.8% of unpooled nDCG@10 after targeted training, versus 94.6% without it; the storage-quality tradeoff still needs validation on the deployment corpus.
https://huggingface.co/blog/token-pooling
🛠️ Tools & Tips
(no items this day)
⭐ Open Source Highlights
1. llama.cpp and ggml v1.0.0 make interface compatibility visible in version numbers
The projects now signal public C API breaks with major releases, compatible features and model support with minor releases, and API-preserving fixes with patches. This gives downstream inference integrations clearer upgrade boundaries while explicitly allowing future breaking changes.
https://huggingface.co/blog/ggml-org/llama-cpp-v1
📰 Industry News
(no items this day)
🚀 Major Releases
1. Salesforce releases Apache-2.0 open-weight SOV-Transcribe-1 for non-streaming ASR
