AI signal, minus the noise.

The city of Rio de Janeiro has post-trained and released a massive language model named Rio 3.5 Open, with 397 billion parameters. It is built upon a Qwen base model—referred to as Qwen 7/2—and integrates SwiGLU activation and Rotary positional embeddings. The model is openly accessible, marking a rare public-sector contribution of a large-scale open LLM.

This article presents a hands-on study on generating security operations center (SOC) narratives for insider threat detection using small open-weight language models. The experiments are conducted on the CERT R4.2 dataset using Qwen3 models, comparing four approaches: zero-shot prompting, few-shot prompting, supervised fine-tuning with LoRA (SFT LoRA), and Group Relative Policy Optimization (GRPO). The study demonstrates a practical workflow for adapting small LLMs to explain insider threats, highlighting the accessibility of fine-tuning with open-weight models.

The viral study tested medical AI products UpToDate and OpenEvidence—not underlying models—on benchmarks like MedQA and HealthBench, finding them worse than frontier general-purpose models. The author argues this does not prove domain-specific models are inherently inferior; their own comprehensive benchmark shows fine-tuning a frontier model for medicine yields a noticeable boost. Current domain-specific models often lag because they are built on older or weaker open-source base models, not because specialization fails. For example, Baichuan-M4 is cited as a medical-specific model that claims to outperform frontier models. The main takeaway is that adapting strong frontier models into medical tools quickly would produce superior domain-specific systems, but open-source base model progress and adaptation speed remain challenges.

This tutorial provides a practical overview of core LLM concepts for machine learning engineers. It begins with foundational elements like tokens, transformer architectures, and embeddings, then covers advanced techniques including prompt engineering, retrieval-augmented generation (RAG), and fine-tuning. The guide emphasizes developing sound engineering judgment to move beyond trial-and-error prompting. No new research or product announcements are made; it serves as an educational resource.

Trajectory Labs announced they have achieved frontier model performance using an open model that was post-trained in under 24 hours. The training infrastructure was powered by Together Compute and NVIDIA. No specific model name, benchmark metrics, or dataset details were provided in the social media post. The announcement highlights the potential of combining open models with efficient training infrastructure.

The paper proposes Retrieval-Augmented Reinforcement Fine-Tuning (RA-RFT), a post-training framework that teaches language models to reason by analogy. It first trains a reasoning-aware retriever via gold-relevance distillation, so that contexts are ranked by expected reasoning benefit rather than semantic overlap. The policy model is then fine-tuned using reinforcement learning on retrieved analogous demonstrations under verifiable outcome rewards, enabling it to leverage reasoning traces. Analysis shows that reasoning-aware retrieval surfaces complementary solution strategies that provide distinct scaffolding per problem. On AIME 2025, RA-RFT improves average@32 accuracy over GRPO by 7.1 points for Qwen3-1.7B and 2.8 points for Qwen3-4B, demonstrating that reasoning-aware retrieval is an orthogonal improvement to reward design or training curricula.