All news

vLLM TorchTPU is a new backend that runs existing vLLM model code on Google Cloud TPUs without changes.

Google pairs AI with low-cost ultrasound so frontline health workers can do prenatal scans.

Google open-sourced ML Drift under Apache 2.0, a cross-platform GPU compute engine for on-device inference over OpenGL ES, OpenCL, Metal and WebGPU.

IQuest-Q1 is a 320B sparse MoE model built for agentic CLI tasks.

JetBrains open-sourced ThinkRail, an agentic IDE built on the pi coding agent with git-worktree workspaces per task, living specs and reusable skills.

Compares pipelining and combining for convergecast aggregation when values exceed one message.

SplatLab automates reproducing and extending 3DGS research.

Self-evolving harnesses for history-dependent robot manipulation.

Anthropic researchers report LLMs now make formal software verification feasible at much larger scale.

POSIX wordexp(3) can run command substitution, so using it for tilde expansion is a code execution risk.

An essay on two aspects of compiler warning design: machine-applicable fixits and how code can suppress a warning, using GCC and Clang -Wparentheses as the example.

Cloudflare argues login pages are a key attack surface after the NetScaler flaws and shows how to shield gateways beyond patching.

RealMe builds immersive volumetric video from sparse captures.

A framework that detects and reduces QoS violations caused by noisy neighbors in multi-tenant public clouds.

Diffusion-guided contrastive learning for cross-domain sequential recommendation.

Information-bottleneck VAE with gating for top-N recommendation.

Pull request to remove pgrust from ClickBench because it embedded one hand-written kernel per query and cached results.

Noise-resilient priors to denoise performance models of parallel codes at scale.

Five Qwen3.8-27B models run through the GVS5H harness match Fable 5 on a benchmark, showing harness design matters as much as model size.

How RL papers get into seven open-source LLM training frameworks.

Agents create physically feasible brick builds.

eMudhra exits the TLS CA business after issuing intermediates with keys so weak anyone could mint certificates.

ESR-HGNN removes redundant semantic computation to speed up mini-batch heterogeneous GNN inference.

AgenticEvolve uses persistent agents that design and run experiments inside evolutionary program search.

OLMo-3 checkpoints show little evidence of instrumental convergence in RLVR.

Applies coding-agent post-training to agentic robotics.

Vercel Labs open-sources fx, an embeddable coding agent written in Zig.

Survey of multimodal embodied agents from computer use to robot use.

Trail of Bits releases patching-bugs and regression-hunt, agent plugins that write fixes and then check what else the fix changed.

Tests whether frontier agents can automate ML for science in the wild.