Chris Ryan

Backdooring Electron Applications by Iván Cabrera Fresno looks at how locked-down Windows environments break old playbooks, where WDAC enforcement, signed execution, deep inspection, and constant monitoring make traditional C2 implants difficult to deploy and even harder to keep alive. When unsigned implants won’t run and there are no loaders or shortcuts, what’s your next move? https://lnkd.in/gN_yai7m

30

1 Comment

At Ori, we’ve seen many AI teams lean on Slurm as workloads grow more demanding. Slurm gives you a familiar, battle-tested way to schedule and track GPU jobs, but for teams encountering it for the first time, the initial setup can feel daunting. That’s why we’ve put together a practical, hands-on guide to running Slurm on Ori GPU Instances, with clear steps, configurations, and all the commands included. If you’re new to Slurm and want a smoother path to getting it up and running, this guide is a good place to start 👉 https://hubs.la/Q04082Sz0

20

#reconfigurable_space_computing_lab at CPP: Our latest infographic from the Reconfigurable Space Computing Lab at Cal Poly Pomona distills the sprawling world of drone flight-software into three digestible categories. By grouping well-known stacks by their scope and origin, we make it easy to see at a glance which platform fits your project’s needs—whether you’re chasing NASA-grade modularity or out-of-the-box flight readiness. First, there are four full-stack open-source autopilots: ArduPilot, PX4, Paparazzi UAV, and UAVobjects. These frameworks handle everything from low-level sensor drivers and real-time scheduling up through mission planning, telemetry, and logging. They’re ideal if you need a mature, community-driven codebase that you can customize down to the last line. Next, we highlight six lightweight, multirotor-focused forks and clean-sheet firmwares—Betaflight, Cleanflight, iNav, RaceFlight, LibrePilot, and KISS. Born from the FPV racing scene, they emphasize high-rate control loops, minimal latency, and straightforward configuration, making them a favorite for hobbyists and anyone pushing performance on microcontrollers. Finally, there are roughly eight to ten proprietary and OEM stacks, including DJI’s Naza/A3/Mavic series, Yuneec, Skydio, Auterion’s enterprise PX4 distribution, MicroPilot, 3DR Solo (legacy), FreeFly Systems, and Microdrones. These closed-source solutions often come locked to specific hardware but offer polished user experiences, integrated ground-control software, and vendor-backed support. Whether you need the ultimate flexibility of a fully open framework or the rapid deployment of a turnkey commercial system, this breakdown helps you navigate the choices. Dive into the infographic to pinpoint the flight-software family that will launch your UAV project to new heights!

10

#reconfigurable_space_computing_lab at CPP: I’m excited to announce the launch of our new Ptah website—Secure Edge-AI & Post-Quantum Crypto in Space Systems—at Cal Poly Pomona’s Reconfigurable Space Computing Lab! 🚀🔐 Explore how Ptah empowers students to: Build resilient space-grade and terrestrial edge-AI platforms, combining RISC-V accelerators, Raspberry Pi clusters, and NVIDIA edge GPUs Integrate Post-Quantum Cryptography (CRYSTALS-Dilithium & Kyber) and lightweight AEAD (ASCON) to safeguard firmware, telemetry, and command streams Orchestrate heterogeneous clusters with K3s, monitor performance with Prometheus & Grafana, and deploy containerized microservices under real-world constraints Over a 15/30-week curriculum, learners dive into hands-on labs—benchmarking PQC schemes, configuring Kubernetes on Pi/Jetson nodes, and building self-healing drone and rover networks. You’ll find architecture diagrams, performance charts, code snippets, and video walkthroughs to guide every step. A huge thank you to Dr. Wen-mei Hwu (UIUC) for your mentorship, Dr. Benny Cheng (Naval Surface Warfare Center) for your partnership, and our incredible RSCL students whose creativity drives Ptah forward. 👉 Check out the site and let me know if you’d like to collaborate or integrate Ptah into your curriculum: https://lnkd.in/guKqU765 #EdgeAI #PostQuantumCrypto #ReconfigurableComputing #SpaceSystems #Education #Kubernetes

22

1 Comment

Radiance field like NeRF and 3DGS are amazing to capture a specific moment in time with a given geometry, lighting and context (static but now also dynamic). What if the lighting context could be changed post training? Carnegie Mellon University presents "💡LightSwitch💡: Multi-view Relighting with Material-guided Diffusion" They propose Lightswitch, a novel finetuned material-relighting diffusion framework that efficiently relights an arbitrary number of input images to a target lighting condition while incorporating cues from inferred intrinsic properties. By using multi-view and material information cues together with a scalable denoising scheme, our method consistently and efficiently relights dense multi-view data of objects with diverse material compositions. They show that our 2D relighting prediction quality exceeds previous state-of-the-art relighting priors that directly relight from images. They further demonstrate that LightSwitch matches or outperforms state-of-the-art diffusion inverse rendering methods in relighting synthetic and real objects in as little as 2 minutes. We will publicly release our model and code. #gaussiansplatting #3DGS #relighting #reflectance #highdynamicrange #HDR #vfx #light #radiancefield #multiviewgeometry #futurism

92

4 Comments

DeepPolisher is a new open source deep learning tool from Google Research and the UC Santa Cruz Genomics Institute that fixes errors in genome assemblies and makes them much more accurate. It was recently used to improve the Human Pangenome Reference, which is a high quality reference of the human genome. The human genome has about 3 billion base pairs (A, T, G, C). Even small error rates mean a huge number of mistakes that can hide important genes or cause problems for disease diagnosis. Genome assembly takes many short or long DNA reads from sequencing machines and puts them together into a complete genome. Errors still remain, especially insertion and deletion errors (indels), which are dangerous because they can shift how genes are read. DeepPolisher builds on an earlier tool called DeepConsensus, which already reduced long read sequencing error rates to less than 0.1%. DeepPolisher uses a transformer model trained on one of the most accurate human reference genomes ever made (~99.99999% correct, only a few hundred to a thousand errors in all 6 billion bases from both parents). It takes in several kinds of input for each read: the base call (A, T, G, C) the quality score from the sequencer how uniquely that read maps to the draft genome annotations about mismatched bases The model is trained by aligning reads to a draft genome assembly, finding the mistakes, and learning how to correct them. For training, chromosomes 1–19 are used. Chromosomes 20–22 are held out to test and pick the best model. DeepPolisher works in a single pass, classifying each base as correct or wrong and suggesting a fix. It reduces total genome assembly errors by about 50% and indel errors by more than 70%. The quality score (Q-score, a log scale measure of error probability) improved from Q66.7 to Q70.1 in tests on Human Pangenome Reference Consortium data. Q70 means a 99.99999% chance that a base is correct. This improvement means less than one wrong base in half a million bases. DeepPolisher has already been used in the HPRC’s second release of data, which included assemblies for 232 individuals (five times more than the first release). All assemblies went through an extra polishing step with DeepPolisher, cutting single nucleotide and indel errors in half. By releasing DeepPolisher as open source, Google and UCSC want to make these improvements available to scientists everywhere. This will help create more accurate reference genomes, improve the detection of genetic diseases, and support research and drug discovery for people from all ancestries.

7

In 2021, we started investigating uses for neural networks in our rendering engine. 𝗢𝘂𝗿 𝗳𝗶𝗿𝘀𝘁 𝘀𝘂𝗰𝗰𝗲𝘀𝘀𝗳𝘂𝗹 𝗮𝘁𝘁𝗲𝗺𝗽𝘁 𝘄𝗮𝘀 𝗡𝗲𝘂𝗿𝗮𝗹 𝗠𝗮𝘁𝗲𝗿𝗶𝗮𝗹𝘀. Getting hardware-accelerated inference to work cross-platform inside shaders turned out to be the hard part. NVIDIA Tensor Cores were only accessible through CUDA. Intel's XMX extensions were vendor-specific. AMD used regular shader cores for WMMA instructions. Wave Matrix existed as an abstraction, but it was designed for non-divergent work, meaning every thread in a wave had to operate on the same weights. That breaks down in rendering. In Neural Materials, adjacent pixels might belong to completely different materials, each requiring different network weights. So we bucketed queries per material and ran multiple dispatches. It worked, but it was cumbersome. Cooperative Vectors address this by shifting from matrix-matrix to vector-matrix operations. Each thread holds its own long vector, a transparent data structure that can be much longer than the usual 4-element vectors in shaders. The hardware then combines individual vector-matrix requests from threads in a wave into a collectively accelerated operation. For NRC, where every pixel has different inputs but feeds into the same network, this is the best-case scenario performance-wise. 𝗙𝗼𝗿 𝗶𝗻𝗳𝗲𝗿𝗲𝗻𝗰𝗲, MulOptimal layout is what you want. Write data in row/column first, convert, and query destination size on the CPU before allocating memory, since it's implementation-dependent. 𝗙𝗼𝗿 𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴, gradients are computed using OuterProductAccumulate and VectorAccumulate. The weight update step is one thing Cooperative Vector intrinsics can't currently handle and still needs to be done on the CPU or in regular shader code. One note on API status: DirectX has withdrawn the Cooperative Vector proposal, folding those features into a new Linear Algebra feature. On Vulkan, Cooperative Matrix was promoted to KHR while Cooperative Vector remains NVIDIA-only. Luca Quartesan from our team wrote the full technical breakdown: https://lnkd.in/e6yw-sbB If you're working on neural inference in shaders and want to understand how we're approaching this across platforms, DM me.

48

From rockets to AI. Nine years after the original NVIDIA DGX-1 handoff, Jensen Huang delivered a brand-new DGX Spark to Elon Musk at Starbase. The DGX Spark is a petaflop powerhouse built for creators, researchers, and developers — a desktop-sized supercomputer with five times the performance of the first DGX-1. 🚀✨ Filmed at SpaceX rocket factory in Texas, this moment captures a conversation between the two visionaries. ➡️ https://bit.ly/4o7oqGw 📹 https://bit.ly/42GxidS

26

Quantum, AI, fusion. Every breakthrough comes with its global race. Europe has long been a leader in fusion research, but public projects aren’t expected to deliver results before the late 2030s. This is why private companies are trying to move faster to close the technology-time gap between research and demos. Among them are two of the most exciting fusion tech startups: - Proxima Fusion, founded in 2023 by Francesco Sciortino, Jorrit Lion, Lucio Milanese, Martin Kubie, and Jonathan Schilling. - Marvel Fusion, founded in 2019 by Moritz von der Linden, Dr. Karl-Georg Schlesinger, Pasha Shabalin, and Georg Korn. We analysed their talent structures from Manager to C-level and identified 33 current professionals at Proxima Fusion and 32 at Marvel Fusion. We were curious to know whether the two companies differ only in their approach to fusion, or whether those differences carry over into how they build their teams. Read the full piece for the full analysis 👉 https://lnkd.in/eEfTxJHB

14

New AI cameras are watching the outdoors to spot endangered, rare bees. 🐝 The AI-based imaging system is a collaboration among researchers at the University of Oregon and Oregon State University, funded by the National Council for Air and Stream Improvement Inc. Designed to monitor bee populations without harm to the insects, the technology potentially solves a key paradox in bee conservation efforts. “California and Oregon are global hotspots for bee diversity. So what we do here means a lot for the rest of the world,” said Lauren Ponisio, an associate professor of biology at the UO and a lead researcher on the project. Read more 🔗 https://bit.ly/45hlaS1

66

1 Comment

Caltech engineers have introduced ATMO (Aerially Transforming Morphobot), a groundbreaking robot that shifts mid-air from a flying drone to a wheeled rover. Advancing their earlier M4 model, ATMO addresses the challenge of seamless transitions on real-world terrain. Unlike other hybrid robots, it folds its propeller-wheels downward before landing, enabling stable “dynamic wheel landings” on uneven surfaces. A central motor and joint system, paired with an advanced algorithm, adjusts propeller thrust in real-time for flight stability, while belt drives and differential steering power its rover mode. Published in Communications Engineering, this innovation could transform exploration, search and rescue, and planetary missions by enhancing multi-modal robotics. #Robot #airobots #Robots #ATMO #TransformingRobot #CaltechInnovation #MultiModalRobotics

110

4 Comments

TGen - Part of City of Hope partners to release new technology with Google and University of California, Santa Cruz: DeepSomatic 🧑🏽‍💻 This machine learning method is a major leap forward for cancer mutation detection. DeepSomatic is the first of its kind, designed to work with both short-read and long-read genetic sequencing data. Why it matters: - It improves the precision and completeness of information from long-read data. - It increased the confidence of known cancer mutations by 7x in initial tests. - It works well even on challenging archival patient samples. TGen, part of City of Hope, is proud to have contributed to this landmark study alongside UC Santa Cruz Genomics Institute and Google Research. The team also released the high-quality CASTLE dataset to accelerate research. Read more at https://bit.ly/4nYDDJY and https://lnkd.in/gQA4frVa #CancerResearch #Genomics #MachineLearning #Bioinformatics #DeepSomatic #TGen

45