Rosalind: A genomics toolkit in Rust running whole-genome pipelines on a laptop
Rosalind is a Rust-based genomics engine that enables whole-genome analysis with a memory footprint under 100 MB, making it suitable for resource-constrained environments like hospitals and fieldwork.
Matrix Multiplications on GPUs Run Faster When Given "Predictable" Data
Matrix multiplications on GPUs exhibit faster performance with predictable data, such as zeros, due to reduced dynamic power consumption, which minimizes transistor switching. This phenomenon challenges conventional understanding, as the content of matrices typically should not influence computational speed.
AI-generated CUDA kernels silently break training and inference [R]
AI-generated CUDA kernels from NVIDIA's SOL-ExecBench exhibit critical failures in production, notably the embedding-gradient + RMSNorm backward pass, which caused loss divergence during transformer training despite passing initial benchmarks.
Multi-Agent LLM System for Automated Vulnerability Discovery and Reproduction
FuzzingBrain V2 is a multi-agent system that enhances automated vulnerability discovery by achieving a 90% detection rate on the AIxCC 2025 dataset and identifying 29 zero-day vulnerabilities in real-world applications, all confirmed by maintainers.
The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence
The MiniMax-M2 series introduces a groundbreaking approach to language models, leveraging mini activations to achieve maximum real-world intelligence with the flagship M2 model featuring 229.9B total parameters and only 9.8B activated per token. Link to article
QUACK: Questioning, Understanding, and Auditing Communicated Knowledge in Multimodal Social Deduction Agents
QUACK is an innovative open-source framework designed to audit the grounding of language in multimodal social reasoning, addressing the limitations of traditional scoring methods in social deduction games.
EMA-Gated Temporal Sequence Compression in Vision Transformers [P]
NeuroFlow enhances Vision Transformers by eliminating 90% of compute waste on stationary elements, achieving a 55.8x speedup in processing high-resolution video without requiring fine-tuning.
noisekit - CLI for generating realistic degraded speech datasets for ASR benchmarking [P]
noisekit is a CLI tool that generates realistic degraded speech datasets for ASR benchmarking, allowing teams to evaluate STT models under conditions that mimic actual production environments.
EAMS (Equivariant Anatomical Mesh Segmentor) leverages Equivariant Mesh Neural Networks to achieve robust anatomical mesh segmentation, outperforming traditional methods by maintaining stability under geometric perturbations while using fewer than 2M parameters.
Cross-species RSA: same learning rules (BP, PC, STDP, FA) tested against both human fMRI and macaque electrophysiology [P]
Cross-species RSA reveals that early visual alignment is conserved, with STDP (ρ ≈ 0.30) and PC (ρ ≈ 0.28) leading in macaque V1/V2, mirroring human V1 results and indicating that this pattern is not an fMRI artifact.
Building self-improving tax agents with Codex
Warp’s big bet on building open source with GPT-5.5
Cisco and OpenAI redefine enterprise engineering with Codex
NVIDIA Vera CPU Is ‘Packing a Heavy-Hitting Punch’ Against Competition
The NVIDIA Vera CPU showcases exceptional performance for agentic AI, featuring 88 custom Olympus cores and 1.2TB/s memory bandwidth, making it a formidable competitor against Intel and AMD processors.
AI Factories: The New Infrastructure of Intelligence
AI factories necessitate a new class of CPUs that deliver fast cores and massive memory bandwidth, essential for sustaining high performance in agentic AI applications.