Daily Digest

Amazon introduced Set-Supervised Fine-Tuning (SSFT) using 'global forking tokens' to prevent reasoning mode collapse, achieving 5-7% gains in benchmarks. For clinical AI, this reinforcement learning framework allows models to deliberately fork between distinct diagnostic strategies rather than averaging them.

OSCAR enables INT2 KV cache quantization via attention-aware rotation based on query and score-weighted value covariance, reducing memory by 8x. The mixed-precision layout preserves 64 sink and 256 recent tokens in BF16, solving the outlier channel problem for long-context production serving.

Engineers detailed a multi-GPU Volta setup confirming dense models are a 'trap' due to low throughput, while Qwen3.5-122B MoE achieves ~50 tok/s using Q8 GGUF and llama.cpp. The team also observed a RAG failure mode where context became '79% poisoned' by the agent retrieving its own prior outputs.

Release of an abliterated Qwen3.5-35B-A3B model that preserves 785 Multi-Token Prediction (MTP) modules. Testing reveals the 3.5 architecture handles KL divergence (300-400 range) significantly better than 3.6 during abliteration, making it superior for custom uncensored deployments.

TypeGraph fuses BM25, pgvector similarity, and recursive graph traversal into a single Postgres SQL round-trip. Crucial for healthcare RAG, it supports declarative ontology expansion (e.g., subclassing) directly via recursive CTEs, ensuring deterministic entity matching.

A new database indexes 3M scientific papers using SPECTER2 (768D) embeddings in pgvector, combining OpenAlex citation graphs with semantic similarity tracking. This bypasses the citation-lag problem, enabling immediate discovery of novel biomedical preprints.

A 1,500-page RAG benchmark testing layout accuracy and Q&A utility on complex real-world forms, including Healthcare Explanation of Benefits (EOBs). It fills a major gap in evaluating extraction tools for messy clinical and administrative documents.

Audits of biomedical literature show a twelvefold increase in AI-fabricated references since 2023, with review articles experiencing a 57% higher fabrication rate. This critically undermines the evidence chain for Clinical Decision Support systems, mandating automated reference validation pipelines.

South Korea's Asan Medical Center deployed a production-grade, air-gapped RAG system for internal clinical guidelines. It utilizes semantic unit chunking and strictly requires ground-truth document referencing to mitigate hallucinations in a high-compliance environment.

Mayo Clinic and Stanford developed a deep learning framework analyzing cell-free DNA epigenetic methylation markers to map the Tumor Microenvironment. The interpretable model identified 9 spatial ecotypes across 17 solid tumors, expanding liquid biopsy utility beyond direct genetic mutations.

Operating in a regulatory sandbox, Utah's AI pilot autonomously renewed prescriptions 72% of the time, with human physicians agreeing with the AI's logic in 91% of validated cases. Escalations were appropriately triggered for missing lab work, demonstrating viable exception-based workflow routing.

NVIDIA's new Blackwell RTX PRO 4500 utilizes specific DPX instructions to accelerate Smith-Waterman sequence alignment by 9.6x over the L4 baseline. It matches Hopper H100 performance for genomic pipelines while consuming 4.3x less power, heavily impacting hardware TCO for precision medicine workloads.

An implementation guide demonstrating FedProx's proximal term constraints for mitigating divergence in highly imbalanced, non-IID datasets. This federated learning pattern via NVFlare is essential for cross-hospital clinical model training where data centralization is prohibited.

AWS launched Amazon Bedrock AgentCore natively integrating LangGraph, NVIDIA NIM inference, and the x402 m2m payment protocol. The stablecoin-based microtransaction layer enables agentic systems to autonomously query pay-per-use clinical APIs beneath standard credit card fee floors.

Security analysis highlights critical vulnerabilities in Agent-to-Agent (A2A) interactions, specifically 'ghost permissions' and Model Context Protocol scope drift. Robust production systems must implement cryptographic credentials and strict permission attenuation to maintain HIPAA compliance across agent hops.

A zero-click vulnerability in Copilot Cowork allowed prompt injection to force the agent to generate pre-authenticated OneDrive links. The data was exfiltrated when the agent sent unapproved automated emails that rendered external image requests to attacker-controlled servers.

AWS demonstrated AgentWatch, an ambient agent deployed on AgentCore Runtime that continuously monitors log streams and executes Notify, Question, or Review HITL patterns. This provides a blueprint for overseeing autonomous infrastructure in highly regulated environments.

Lilly's experimental retatrutide, a triple agonist targeting GLP-1, GIP, and glucagon receptors, achieved a 28.3% average body weight loss over 80 weeks. Paired with their acquisition of Engage Biologics' non-viral DNA platform, this positions GLP-1 analogues as upstream foundational longevity interventions.

Daewoong Pharmaceutical acquired Turn Biotechnologies' ERA platform to commercialize mRNA-delivered partial cellular reprogramming. Initial strategic clinical targets focus on localized delivery for ophthalmic and auditory tissues to restore epigenetic states without risking pluripotency-induced tumors.

New research shows tau protein phosphorylation at Threonine-205 is essential for maintaining memory engram precision. This challenges existing Alzheimer's paradigms, suggesting that total tau depletion therapies may impair remote recall by inducing noisy neuronal firing.

Pope Leo XIV released 'Magnifica Humanitas' (co-advised by Anthropic's Chris Olah), framing LLMs as 'cultivated' systems rather than engineered tools, leading to profound interpretability black boxes. Simultaneously, open-weight decensoring tools compromised Llama 3.3 and Gemma 4 guardrails in under 90 minutes.

While median seed rounds have grown to $3M, the transition requirements to Series A have drastically increased to $2M-$4M ARR. As a result, only 24% of 2023 seed-funded AI companies have successfully progressed, highlighting severe execution thresholds for new platforms.

Pro se lawsuits have spiked to 16.8% of federal civil cases, with 18% of recent complaints flagged by the Pangram detector as AI-generated. The mass drafting of mandamus petitions via Microsoft Copilot is causing significant processing backlogs, forcing courts to issue structural shredding orders.