Daily Digest

Lead candidate REP-0004 targets liver cells with mRNA/LNPs to degrade intracellular free cholesterol, triggering reverse cholesterol transport. This novel active-removal approach moves beyond statin-based suppression, with Phase 1 trials aimed for mid-2027.

A clinical decision support framework differentiating deterministic and probabilistic genetic findings. Attia argues that testing is only clinically valuable when it alters actionable CDS paths, prioritizing pharmacogenetics over generic functional medicine 'detox' panels.

Madrigal licensed RNAi asset ARO-PNPLA3 for genetically-driven MASH in a $1B deal. The asset targets PNPLA3 I148M homozygous patients and achieved up to a 46% liver fat reduction at 12 weeks in Phase 1 data.

As 2025 U.S. Acute Care EHR purchasing dropped 40% YoY, Epic captured 100% of enterprise-wide decisions for systems with more than 10 hospitals. The consolidation secures their ecosystem dominance ahead of Oracle Health's impending AI-enabled EHR release.

ChatRx integrated AI-assisted intake with DoseSpot e-prescribing to slash prescribing time from 15+ minutes down to 6-10 minutes, with a 0% prescription failure rate for low-acuity clinical workflows.

Aderant unified six disconnected systems (Confluence, Jira, Git, etc.) using Model Context Protocol (MCP) servers and Okta SSO. The implementation dropped cross-platform search times from 45 minutes to under 5 minutes while maintaining strict IAM separation.

A governed on-premise agentic architecture on Dell/NVIDIA hardware featuring 'ACL Hydration', which preserves Access Control Lists from source documents (like SharePoint) directly into the vector database to prevent unauthorized RAG access.

NVIDIA validated NVFP4 on a 12B hybrid Mamba-Transformer trained over 10T tokens, achieving a 62.58% MMLU-Pro score. The method delivers 2x-3x speedups over FP8 using 16x16 2D block scaling and Random Hadamard Transforms (RHT) to manage weight-gradient outliers.

Addressing the 'neuron death' problem in tall MLP matrices caused by the Muon optimizer, the new Aurora leverage-aware optimizer improved MMLU scores by 10 points on 1.1B-parameter transformers.

Notes that 95% of prototypes fail due to 'Vibe Check' evaluation debt. For production tool use, giving an agent a terminal (CLI) with pipe capabilities is highly context-efficient compared to stringing together numerous granular MCP tool calls.

Research indicates curated, hardcoded skills improve task completion by 16.2%, whereas self-generated LLM procedures offer no consistent benefit. The optimal pattern treats skills like declarative YAML configurations while utilizing MCPs as the execution runners.

IBM Research released a framework for evaluating end-to-end agent systems. The data shows failed agent runs burn 20-54% more tokens due to retry loops, with 'tool shortlisting' identified as the highest-impact architectural optimization.

Built on the open-source Kimi K2.5 checkpoint with heavy RL and synthetic data, Composer 2.5 achieves frontier coding performance at just $0.50 per 1M input tokens. The release coincides with rumors of a $60B acquisition by SpaceX.

OpenAI partnered with Plaid to offer real-time read-only sync of 12,000+ financial institutions within ChatGPT. This data grounding strategy strongly mirrors the secure third-party aggregation architectures needed for clinical CDS platforms.

Shifts away from LLM-as-a-judge patterns by enabling deterministic, Lambda-based evaluators running over OpenTelemetry session traces for high-stakes schema validation and numerical drift analysis.

Cloudflare's testing of Anthropic's Claude Mythos Preview revealed the model can autonomously construct and compile exploit chains. However, memory-unsafe languages (C/C++) trigger high false-positive vulnerability reports unless the model is explicitly forced to generate a PoC.

A new 'AudioHijack' vulnerability demonstrates a 79-96% success rate injecting imperceptible adversarial commands into LALMs. The attack disguises itself as natural reverberation, easily bypassing standard reflection defenses by exploiting the model's attention mechanisms.

A conservative coalition has called for executive branch oversight and mandatory safety testing for frontier AI models prior to release, rejecting current corporate self-policing standards.

Following the UK NHS's move to close open-source repositories over vulnerability fears, the Government Digital Service issued a firm recommendation to remain 'open by default' to prevent spiraling delivery and maintenance costs.