Daily Digest

CNA enables sparse MLP circuit steering without SAE training or weight modification by ablating just 0.1% of MLP activations. The method reduced refusal rates by over 50% in instruct models while maintaining high output quality, proving alignment fine-tuning primarily acts as a refusal gate on existing base model structures.

Addressing the memory-bound limits of autoregressive decoding, Nemotron-Labs' new models use block-by-block denoising and self-speculation. The 8B model achieves up to 865 tok/s on Blackwell architectures, representing a 4x-6.4x throughput improvement over AR baselines.

Operating autonomously for 35 hours, Qwen3.7-Max optimized a hardware-based attention kernel for SGLang on T-Head accelerators. It executed over 1,158 tool calls with a split training architecture, achieving a 10x speedup over the reference Triton implementation.

DeepSeek permanently priced its V4 Pro model at $0.435/1M input and $0.87/1M output tokens. At 34.5x cheaper than GPT-5.5 for output, this shifts the unit economics for high-volume agentic systems relying on massive token generation.

Garry Tan open-sourced GBrain, a WASM-backed local Postgres memory layer that auto-wires a knowledge graph without LLM extraction. It utilizes a hybrid search pipeline with HNSW, BM25, and Reciprocal Rank Fusion, achieving a 49.1% P@5 on BrainBench.

Unconstrained agentic loops are economically unsustainable in production. Implementing early commitment classification and compiling LLM traces into branch-free, deterministic recipes via a 'LOOP Skill Engine' can reduce token consumption by over 93%.

Standard RAG often fails in enterprise deployments because models blend parametric and retrieved data unpredictably. A grounded approach forces models into a procedural role, first extracting structured Pydantic schemas with line-level citations before composing any narrative.

Independent researchers questioned Google's claim that AI agents built an OS from a single prompt, noting the instruction set was thousands of lines long. The critique highlights the lack of log transparency and advocates for rigorous open-world evaluation norms to verify long-horizon agent autonomy.

NV-Generate-MR-Brain uses Latent Rectified Flow to synthesize high-resolution 3D MRI volumes and segmentations at 33x the speed of DDPM architectures. Trained on the massive MR-RATE dataset, it provides anatomically consistent synthetic data to bypass HIPAA constraints in model training.

Children's Hospital of Orange County implemented a generative AI RAG system to process unstructured patient charts, cutting review times from over 60 minutes to minutes. The system surfaces source-linked annotations while strictly maintaining a human-in-the-loop governance model.

Research identifies PC decline as a primary driver of age-related mitochondrial fragmentation. Modulating the SAMS-1 pathway and supplementing with choline restored PC synthesis and rescued mitochondrial morphology in both C. elegans and human fibroblasts.

Oligomerix successfully completed a 76-person Phase 1a trial for OLX-07010, an oral small-molecule inhibitor. Unlike traditional clearance therapies, this drug prevents toxic tau proteins from self-associating, offering a logistical advantage over clinical infusions.

Oura acknowledged receiving government demands for user data, highlighting that its biometric metrics are not end-to-end encrypted. Sensitive biomarker data is unscrambled on Oura's servers, exposing continuous health monitoring to potential third-party access.

Testing against gpt-oss-20b reveals that while rate limiting acts as a firewall against GPU OOM errors, quota reservations are essential for multi-tenant throughput. Load tests at 20,000 RPM against a 1,000 RPM capacity proved the necessity of reserving a 10-20% unreserved pool to prevent priority starvation.

The Claude Mythos Preview is surfacing critical software vulnerabilities at a rate that breaks standard 90-day disclosure cycles. Outperforming GPT-5.5 on UK AISI cyber ranges, the model flagged thousands of critical bugs with lower false positive rates than human testers.

The latest FastAPI release introduces strict Server Sent Event (SSE) field validation to prevent malformed data structures. It also delivers a 24% speedup in the test suite through optimized fixture scoping and bumps dependencies for pydantic-ai and sqlmodel.

Written in Go with zero non-stdlib dependencies, Bumblebee is a read-only inventory collector that scans developer environments for malicious packages. It safely inspects local state, including MCP JSON configs and editor extensions, without triggering postinstall execution hooks.