Daily Digest

Combines deep learning for MRI feature prediction with Conformal Prediction for rigorous uncertainty quantification, feeding only high-confidence outputs into a Longitudinal Latent Class Mixed Model (LCMM). This approach drastically improved Matthews Correlation Coefficient (MCC) scores for structural abnormalities like Bone Marrow Lesions (0.69 to 0.91) and Cartilage Loss (0.45 to 0.80).

A multi-ancestry GWAS analyzing 105,869 cases identified 80 associated genetic variants, including 37 novel targets, and validated the reliability of self-reported medical data (84% alignment with clinical records). The findings led to the identification of 5 candidates for drug repurposing and a new Endometriosis PRS report utilizing over 13,000 variants.

A Bundibugyo virus Ebola outbreak in DRC and Uganda has spread to 25 health zones, with CDC computer modeling predicting up to 20,000 cases and 4,000 deaths if isolation rates remain low. The situation exposes critical vulnerabilities in global health surveillance, exacerbated by the CDC losing roughly 3,000 specialists over the past 18 months.

Preclinical data comparing GLP-1 therapy (semaglutide) with PTC-2105, a mitochondrial-targeted geropeptide, shows that while semaglutide drives a 25–40% loss of lean mass and a sarcopenic phenotype upon discontinuation, PTC-2105 simultaneously reduces visceral fat and increases lean mass. PTC-2105 operates by restoring mitochondrial membrane potential and triggering apoptosis in senescent cells.

Stanford spinout Adaptyx Biosciences has demonstrated a wearable capable of continuous, multi-day tracking of free cortisol using synthetic DNA-based molecular switches beneath the skin. Validated against paired blood labs, the sensor successfully captured both overnight nadirs and the Cortisol Awakening Response (CAR).

Preliminary data for Eli Lilly’s VERVE-102 shows profound LDL-C lowering via a one-time gene editing intervention utilizing an Adenine Base Editor delivered via Lipid Nanoparticles to the liver. This permanently disables the PCSK9 gene, moving cardiovascular protection from chronic monoclonal antibody therapies to targeted genomic modification.

Researchers identified "neuroproteasomes" that eject peptide fragments into intercellular space; blocking them induces tau filaments identical to those found in Alzheimer's. The mechanism links directly to genotype, with ApoE4 neurons reducing surface levels and requiring only 20% inhibition to trigger aggregation, establishing a proteostatic reserve model of aging.

A new Postgres C extension, `pgtoken`, eliminates repetitive RAG tokenization by storing token IDs as rank-varint compressed bytea adjacent to `pgvector` embeddings. Compressing down to an average of 1.7 bytes per token, it enables O(1) context window filtering directly in SQL without application-layer tokenization overhead.

Addresses "context poisoning" in patch-based Time Series Foundation Models using a Gate, Router, and Critic framework that suppresses structurally anomalous patches at inference without parameter updates. Testing on TimesFM 2.5 achieved a 1.95% MASE reduction across 53 datasets by predicting when context intervention improves zero-shot forecasts.

Fine-tunes LLMs using Contrastive Preference Optimization (CPO) on synthetic datasets featuring both positive and negative reasoning traces for relational data tasks. This methodology yielded 9.7%–16.3% absolute improvements over standard Supervised Fine-Tuning on the MMQA benchmark.

Community discussions highlight systemic RAG failures, including shallow retrieval limits on dense corpora (e.g., missing cross-document connections in 300+ PDFs) and "missing image" errors where semantic matching of image metadata becomes decoupled from the most relevant text chunks during final assembly.

Early llama.cpp support for DeepSeek V4 reveals a native FP4-FP8 hybrid architecture that is highly resilient to quantization degradation. Utilizing the DSA (DeepSeek-V3/V4 specific architecture), it features exceptionally efficient context window scaling that minimizes KV cache consumption.

Nvidia dropped Nemotron 3.5 ASR, a 600M-parameter cache-aware FastConformer-RNNT model that performs real-time streaming transcription across 40 language-locales. Developers can tune an `att_context_size` parameter at runtime to adjust chunk sizes (80ms to 1.12s) to navigate latency-accuracy tradeoffs without retraining.

Cohere is soft-launching its first dedicated coding model, featuring a 30B total / 3B active parameter Mixture-of-Experts architecture. Built for local execution, weights are currently available on Hugging Face and deliver inference speeds comparable to smaller 3B-dense classes.

Tokyo-based Sakana AI launched the RSI Lab, transitioning from agentic tasks to systems that natively redraft their own code and training architectures. Highlighted projects include LLM-Squared (models designing training methods) and Darwin Gödel Machine for codebase iteration.

Google released Quantization-Aware Training (QAT) checkpoints for Gemma 4 (E2B and E4B), shrinking the E2B model from a 9.6GB BF16 footprint to around 1GB via a new mobile schema. This channel-wise, 2-bit compressed format is specifically targeted at token-generation layers while modularizing encoders to support ultra-low memory hardware.

Nvidia introduced RTX Spark based on the Blackwell GB10 'superchip', integrating 20 Arm CPU cores, 6,144 GPU cores, and up to 128GB of LPDDR5X unified memory for local Windows devices. It leverages Microsoft Execution Containers (MXC) to securely sandbox local autonomous AI agents from core OS functions.

A technical blueprint showcasing the distillation of structured reasoning from DeepSeek V4 Pro into a Qwen3-8B edge model utilizing separate LoRA SFT runs on a single A100. Using dedicated, hot-swappable adapters for generation versus evaluation successfully prevented the formatting leakage typically seen when merging prose and JSON outputs.

Kernel developers are debating the phasing out of `fork()` overhead via proposed Spawn Templates, shifting toward a 'pristine process' model. Utilizing a new `pidfd_config()` API, the change would create empty processes natively to realize a highly efficient `posix_spawn()` mechanism.

Proposes the Protocolized Action-state Communication and Transmission (PACT) framework, reducing token inflation in MAS by substituting raw natural language with compact action-state records. Benchmarks showed SWE-agent maintaining its resolve rate while halving input tokens.

The release of `micropython-wasm` 0.1a2 offers a 362KB WASM-based sandbox for executing LLM-generated Python safely. Crucially, it manages state persistence across executions using a host-blocking `__session_next__()` function, preserving in-memory variables while enforcing rigid hardware constraints.

A zero-dependency Python standard library implementation of the Model Context Protocol (MCP) using JSON-RPC 2.0 over stdio/HTTP. Replaces recursive globbing with shallow search logic to keep execution under 30ms, enforcing strict path resolution to block directory traversal attacks.

A TypeScript-native terminal AI coding agent leveraging a continuous Plan-Modify-Test-Report execution loop with isolated subagents. It features native Model Context Protocol (MCP) integration, branching contextual `/fork` commands, and the ability to process screen recording video inputs to reproduce bugs.

Explores how RLHF induces input-dependent overconfidence in generative models, rendering single global temperatures ineffective. Implementing Adaptive Temperature Scaling (ATS) for per-token temperature modification based on hidden features reduces Expected Calibration Error (ECE) by up to 50%, while Isotonic Regression mitigates overconfidence in massive datasets.

Introduces the Evaluation Robustness Score (ERS) to measure how easily an LLM judge's outcome can be overturned via follow-up prompts. The research identifies that standard judges engage heavily in 'post hoc rationalization' and flip decisions when subjected to authority framing or targeted challenges.

OpenAI introduced a 'Lockdown Mode' targeting prompt injection vulnerabilities by severely limiting outbound network requests. By severing the exfiltration vector, it stops attackers from transmitting hijacked private data to untrusted endpoints.

Google signed a massive infrastructure agreement with SpaceX to utilize 110,000 Nvidia GPUs at the Colossus data center from Oct 2026 to June 2029, dwarfing Anthropic's earlier $1.25B/month commitment. The move operates as intricate financial engineering given Google's ~5% ownership in SpaceX, artificially bolstering revenue ahead of the space company's target $1.75 trillion IPO.

Meta plans to launch a proprietary, paid counterpart to OpenClaw named 'Hatch' to offset its staggering infrastructure overhead. Expected to arrive in July, the tool automates workflow construction and scheduling, with premium tiers ranging up to $200/month.

S&P Dow Jones Indices formally rejected a request to waive the 12-month seasoning and profitability rules for rapid index inclusion. This ruling currently blocks immediate entry for cash-burning operations like SpaceX, Anthropic, and OpenAI into major passive index tracking funds.

Transitioning to Diffrax, a JAX-based ODE solver, eliminated Python interpreter overhead via XLA compilation and reduced execution from 404 μs to 59 μs. Reverse-mode autodiff support allowed for exact log-likelihood gradients at a fraction of the traditional computational cost.

An interactive, browser-based 4x4 weight-stationary systolic array simulation, compiled directly from NumPy-verified SystemVerilog into WebAssembly to expose hardware-level matrix math mapping.

RedNote open-sourced an Apache 2.0 2B-parameter text-to-speech model utilizing a fully continuous architecture. By avoiding intermediate discrete phoneme pipelines, it drives zero-shot voice cloning directly to 48kHz audio with remarkably low latency.

A highly localized Bash agent targeting llamacpp/koboldcpp stacks. Features automated PII redaction and a symbol-targeting code engine that minimizes contextual load by editing isolated logic blocks rather than utilizing broad search-and-replace strings.