Interactive notes explain RLHF, attention, LoRA and LLM evals

Databricks MLflow now ingests native OpenTelemetry traces from the Claude Platform, targeting agent failure modes like returning a plausible answer while missing an important finding or re-reading the same source repeatedly.

A community-made LoRA for Krea 2, posted to Civitai, generates images in the Wall Street Journal's signature stipple/hedcut illustration style. The Reddit poster frames it as a way to imagine yourself rendered inside a WSJ article.

A Reddit user describes an AI assistant called Pickle phoning a hotel in Japan to book New Year's stays, identifying itself as AI and giving the user's full name before the call dropped and it called again. The post is a secondhand anecdote with no linked source or corroboration.

A LocalLLaMA user reports decode speed roughly doubling from 70-80 tk/s with llamaserver to 150 tk/s after switching to Strata for Qwen 3.8 Flash Next. The poster says they don't know what technique Strata uses.

The iOS 27.3 Developer Preview lets developers integrate Siri AI across "all apps," with personal context, natural language prompts, and on-screen awareness. A new "app skills" feature bundles several tasks into one action; Apple's own apps like Mail, Maps, and Photos get the expanded capabilities too.

Model Vault and Model Vault Encrypted now run leading open models from Cohere and other providers on single-tenant infrastructure, with zero data retention of prompts and outputs by default. Model Vault Encrypted keeps data encrypted while the model processes it.

Hazelbaker joins from Uber, where she was president, to lead OpenAI's global communications and corporate positioning. She is the latest veteran executive to leave Uber this year.