Daily Digest

GPT-5.4 integrates native computer use, a 1M-token context window, and a Tool Search API that reduces tool-calling token consumption by 47%. The model scored 83.0% on the GDPval professional knowledge benchmark and introduces an experimental /fast mode for Codex generation.

AI2's new 7B model utilizes a 3:1 ratio of Gated DeltaNet to standard Attention layers, compressing computation into hidden states to bypass quadratic KV cache expansion. This specific hybrid scaling achieved a 2x pretraining efficiency gain over pure Transformers and Mamba2.

This 1T parameter MoE model implements Layer-Adaptive Expert Pruning to drop 500B underutilized parameters and features a Reflection Inhibition Reward Mechanism to halt overthinking. Penalizing reflection beyond three steps reduced output length by 14.38% while boosting reasoning accuracy by 16.33%.

Unsloth released GGUF updates for Qwen3.5 MoE models achieving 99.9% KL divergence. A novel MoE quantization method reduces maximum KLD by 51%, preserving coding and tool-calling fidelity critical for VRAM-constrained production environments.

Argues Mean Reciprocal Rank (MRR) is deceptive for agentic pipelines where context coverage matters more than top-1 clicks. Testing revealed that context prefixing at ingestion and chunk expansion windowing increased content match by 17.9 percentage points.

A local 3B parameter model scored 85.0% on the SimpleQA factual benchmark when paired with a research retrieval API, outperforming DeepSeek-R1 671B. This confirms factual recall is primarily a retrieval optimization problem rather than a parameter-scale reader problem.

RapidFireAI released an evaluation suite for structured benchmarking of chunking, retrieval parameters, and cross-encoder reranking. Includes targeted datasets for healthcare support and PII redaction to replace guesswork prior to vector DB productionization.

Microsoft launched Dragon Copilot integrations using the PDSQI9 standard for provider summarization, embedding third-party hooks for real-time comorbidity identification. Simultaneously, Google rolled out Gemini agents for CVS Health and Highmark Health, processing 6M prompts annually.

Replaced manual OCR with Bedrock Multimodal LLMs and Textract to process 70,000 highly variable healthcare packets monthly. An automated classification layer routes section types prior to extraction, maintaining 98-99% accuracy via configurable confidence thresholds and HITL fallbacks.

Research reveals LLMs are paradoxically more susceptible to absorbing and repeating medical fabrications when prompts are formatted in authoritative clinical prose rather than logical fallacies. Highlights the need for robust fact-grounding guardrails in health CDS platforms.

An analysis of over 940,000 ED visits via Epic's Cosmos DB showed a 10-16% drop in Tylenol orders for pregnant patients following administration misinformation. This presents a critical CDS risk where patients might substitute safe acetaminophen for hazardous NSAIDs during pregnancy.

The CUB library in CCCL 3.1 introduces a gpu_to_gpu determinism mode utilizing Reproducible Floating-point Accumulator (RFA). This ensures bitwise reproducibility across disparate GPU architectures, a vital requirement for FDA validation of deterministic clinical AI models.

ORION bypasses CoreML abstractions to compile multi-step training graphs directly to ANE-native MIL, overcoming a 32MB SRAM cliff. The system hits 170+ t/s for GPT-2 124M inference on M4 Max, opening the door for on-device incremental learning without cloud telemetry.

Blackwell hardware running NVFP4 (4-bit floating point) quantization set STAC-AI LANG6 records for long-context RAG summarization. The GB200 NVL72 showcased vastly superior Interword Latency and Reaction Time against Hopper architecture in both batch and interactive inference.

Solves runtime errors when coupling high-throughput serving engines like SGLang and vLLM with agentic frameworks. Demonstrates extending the SageMakerAIModel class to intercept and parse non-standard streaming outputs into the required Bedrock Messages API schema.

Shifts agent design from static prompt-stuffing to progressive disclosure of instructions (skills). Highlights using Docker scaffolds and LangSmith to track invocation accuracy and step completion metrics, treating dynamic tool retrieval with the same rigor as standard RAG evaluations.

Built on Elixir and the Erlang/BEAM runtime, Symphony orchestrates highly concurrent agent implementation runs. It enforces hermetic testing and requires agents to provide CI pass proofs before merging against a WORKFLOW.md technical contract.

Developed to address the malware risks of unconstrained OS-level agents like OpenClaw, IronClaw wraps tool execution in WASM sandboxes. It virtualizes filesystem access and enforces credential shielding so tokens never surface in model logs.

Implements a production-ready Tree-of-Thoughts pipeline over a FLAN-T5 model utilizing beam search and local depth pruning. Uses a custom heuristic scoring function to evaluate numeric state proximity, significantly improving goal-reaching probabilities over linear Chain-of-Thought.

Discovered a toxic lipid profile in the extracellular vesicles of ALS patients that damages motor neurons from the periphery. This upstream blood biomarker claims >90% detection accuracy and enables rapid pharmacodynamic assessment for clinical trials.

Researchers successfully engineered astrocytes with Chimeric Antigen Receptor (CAR) technology to target and clear amyloid-β plaques. This breakthrough translates CAR-based immunotherapies from oncology into neurodegenerative glial cell modulation.

Introduction of CLIM-TIME, a spatially resolved in vivo CRISPR screening platform. It maps the loss of specific tumor suppressor genes to metastatic immune architectures, providing crucial links for predicting immunotherapy resistance.

The maintainer of the LGPL library chardet utilized Claude Code to engineer an MIT-licensed rewrite with no direct source access. With JPlag source similarity maxing at 1.29%, this challenges traditional clean room definitions and copyleft software economics.

The DOD labeled Anthropic a supply chain risk under 10 USC 3252 due to the company's refusal to authorize Claude for mass surveillance and fully autonomous lethal weapons. An internal memo from Dario Amodei criticized OpenAI's resulting defense contracts as "80% safety theater."

The FDA expanded a Class I recall for Medline ReNewal reprocessed electrophysiology and ultrasound catheters due to residual particulates. A critical inventory alert for clinical CDS systems, the recall requires the immediate quarantine and destruction of the high-risk devices.