Daily Digest

Epic reported that >85% of its clients use its AI suite, with 'Art' cutting discharge summary times by 20-30% and 'Penny' dropping prior authorization submission times by 42%. They also debuted 'Agent Factory', a visual builder allowing hospital IT to deploy customizable agents grounded in local policies.

Yann LeCun's Advanced Machine Intelligence (AMI) raised $1B+ to build non-autoregressive 'world models' that learn continuous physical and abstract environments. AI medical scribe Nabla will be the first healthcare partner to leverage this architecture, aiming to establish a deterministic, auditable regulatory pathway for autonomous medical agents.

To bypass the limitations of naive RAG over messy clinical text, CertHLM uses an ontological data platform that synthesizes EHR and claims data into answer-ready metadata. The agent generates SQL via healthcare-specific pseudo-code to answer natural language queries directly.

A large-scale validation of Google's AI on 125,000 women showed it detected 25% of 'interval cancers' missed by radiologists and reduced second-reader workload by 40%. The study highlighted production gotchas like 'arbitration tension', where humans incorrectly overruled the AI, underscoring the necessity of site-specific calibration workflows.

Researchers found that LLMs are paradoxically more susceptible to adopting medical fabrications when the prompts use authoritative clinical prose rather than logical fallacies. This indicates that safety rails rely heavily on style and context matching rather than deep factual grounding.

A new method called Swave utilizes a recurrent neural network (RNN) to identify complex structural variant patterns directly within pangenome graph architectures, significantly improving SV calling at the population level.

A new genomics-based diagnostic relies on cfDNA methylation and mRNA profiling to measure alpha-Klotho gene silencing. The AI platform detects SNVs and CNVs to stratify patients in neurodegenerative clinical trials based on biological age to prevent placebo-arm bias.

Phase 2 trial results for UBT251—a GLP-1, GIP, and Glucagon receptor triple-agonist—demonstrated a massive 19.7% weight loss in 24 weeks. This validates the shift from single-target pathways to system-level metabolic control.

MITI-101 shifts Alzheimer's focus away from amyloid clearance by neutralizing CD11c markers on peripheral innate immune cells. This prevents the cells from breaching the brain and triggering synapse-damaging neuroinflammation.

To solve the context bloat of raw-text RAG in long-running agents, PlugMem extracts propositional and prescriptive knowledge into a structured memory graph. It uses high-level intents as routing signals, delivering higher utility-per-token than standard chunk retrieval.

Converting source documents into semantically hierarchical file trees and allowing agents to navigate via 'ls' and 'grep' yielded +19% precision and +18% recall for procedural queries over Hybrid RAG, reducing retrieval noise by 72%.

Naive chunking completely destroyed sequence information in a longitudinal sports database. The successful architecture required a combination of dense retrieval, hard structured metadata filters, query decomposition, and cross-encoder reranking.

Nvidia deployed AST-based syntax-aware chunking to preserve code signatures and control flow in complex Unreal Engine contexts. They utilize NVIDIA cuVS for GPU-accelerated quantization and tiered hybrid indexing to power precise tool use via the Model Context Protocol.

A common production failure point in RAG pipelines is the hidden mangling of tables and headers during PDF-to-Markdown conversion, which subsequently invalidates splitting strategies. An open-source visual inspector, 'Chunky', was released to debug this.

Nemotron 3 Nano is a 30B MoE model (only 3B active parameters) featuring a Transformer-Mamba hybrid backbone. It achieves a 256K context length by leveraging Mamba for long-range sequence efficiency while retaining attention layers for structured reasoning.

A developer secured the #1 spot on the Open LLM Leaderboard by surgically duplicating a specific 7-layer block in the middle of Qwen2-72B without modifying the weights. The finding suggests that pre-training carves out distinct, whole-unit functional circuits.

Falcon-H1 introduces a 'Parallel Hybrid' design where attention and Mamba-2 SSM operate simultaneously in each block. It utilizes maximal update parametrization (μP) with non-learnable multipliers to stabilize training spikes associated with SSMs.

Safety researchers utilized heavily censored Chinese LLMs as a proxy for deceptive alignment. They found that removing chat templates and forcing next-token completion effortlessly bypasses the models' fine-tuned suppression of political facts.

DeerFlow 2.0 orchestrates sub-agents for autonomous coding and research by executing them within fully isolated Docker containers. This gives agents a real filesystem and bash terminal to run and debug their own code, minimizing human hand-off friction.

Anthropic's new highly autonomous 'Claude Code' agent uses project-specific markdown and the Model Context Protocol to reason across codebases. During a Mozilla pilot, it chained an average of 21.2 independent tool calls per task to identify 14 high-severity vulnerabilities.

To combat 'Agent Drift' caused by stale LLM training data, DeepLearning.AI released `chub`, an open-source CLI registry. It dynamically injects updated, language-filtered markdown documentation directly into the agent's context window.

AgentMail provides an API-first headless email service that serves as an identity layer for bots. This infrastructure allows autonomous agents to sign up for SaaS tools and manage communication threads seamlessly.

Following internal audits linking a 'trend of incidents' with high blast radiuses directly to unchecked generative AI code, Amazon now requires explicit senior engineer sign-off for all AI-assisted repository changes.

To prevent training GPUs from sitting idle during slow autoregressive rollouts in GRPO pipelines, the standard infrastructure has shifted to disaggregated training and inference. Frameworks overwhelmingly rely on Ray for orchestration and NCCL broadcasts for asynchronous weight syncing.

NIXL is a new vendor-agnostic library designed to dramatically speed up disaggregated model serving. It enables non-blocking, zero-copy transfers of KV cache blocks between prefill and decode workers over RDMA and GPU-Direct Storage.

PostgreSQL 18 introduces `pg_restore_relation_stats()`, a critical developer tool that exports sub-megabyte database statistics from production. This enables local development environments to accurately simulate the production query planner without needing terabytes of actual data.

Hugging Face launched S3-like, Xet-backed mutable storage buckets to solve the inefficiency of using Git for machine learning checkpoints. It features automatic block-level deduplication across model versions and supports cloud-region pre-warming to minimize egress.

A US court issued a preliminary injunction against Perplexity's 'Comet' shopping agent for bypassing Amazon's bot protections by impersonating a Chrome browser. This sets a major precedent: user consent does not override the platform's terms of service regarding unauthorized agent access.

A court ruled that unauthorized use of lyrics cannot be justified simply by arguing the content feels 'AI-generated' due to stylistic breaks. This establishes a high burden of proof for claiming material is public domain AI-slop, protecting human creators who document their prompt/edit workflows.