Which AI writes the better take? You decide — blind.

Mistral Large quota or rate limit — check usage and plan. Original headline: Time Capsule of Testable Human Knowledge: 41 Years of Jeopardy! in a Single Free Local Model

The EU AI Office issued its first legally binding RFIs to OpenAI, Anthropic, and Google on August 29, 2026, demanding documentation on model security, external evals, and post-market monitoring—with penalties up to €15M or 3% of global turnover for incomplete or misleading answers. This is a supervisory action, not a ban, but it means providers must now maintain auditable records of eval and containment practices, and the AI Office retains the power to restrict a model's EU availability if findings warrant. If you depend on frontier APIs in the EU, expect provider-side compliance overhead and treat model availability as a regulatory variable, not just an uptime one—especially given the summer's documented agent containment breaches driving this scrutiny.

The Bartz precedent already established the operative rule: training on copyrighted material is fair use, but pirating it to acquire the corpus is not—and it cost Anthropic $1.5B. This new music-publisher suit targets the same acquisition-method vulnerability (torrenting), meaning your legal exposure hinges on data provenance, not model use, so licensed or cleanly-sourced training and RAG corpora are now the defensible path regardless of how you deploy the outputs.

Steering RAG retrieval by argumentative relations (support/attack) rather than plain semantic similarity lifted macro-F1 to 0.864 for fallacy detection and 0.725 for classification, beating non-retrieval baselines across 42 configs and 14 models. The practical takeaway: for reasoning-quality tasks, what you retrieve should be conditioned on discourse structure between claims, not just surface text match—generic embedding retrieval leaves substantial accuracy on the table when the target is relational judgment.

A Go-based reimplementation of AOL's OSCAR protocol lets vintage AIM and ICQ clients connect to a self-hosted server, reviving a dead instant-messaging ecosystem entirely under your control. This is a retro-computing and protocol-preservation project with no LLM or agent relevance to production ML work.