Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Muse is Meta's subscription AI agent, and its new Apache 2.0 SDK lets hobbyists build physical front-ends for it on ESP32 boards or a Raspberry Pi, adding screens, mics, speakers, and sensors. Community-built skills can let Muse control lights, a TV, or a printer through a local network device. Meta's own Muse Home Link, which bridges Muse to home Wi-Fi devices, is free to US subscribers and ships in October.
Culpium reports TSMC is exploring building and operating a fab that would exclusively serve Tesla, SpaceX, and xAI, and Musk confirmed it on X, saying "just discussions, but something may come of it" while TSMC stayed silent. The two speculative structures Culpium floats are a JASM/ESMC-style joint venture where partners supply capital and demand while TSMC contributes process technology and operations, or a novel arrangement where SpaceX holds majority ownership and TSMC only operates the fab and licenses its process.
The ESPARGOS team found a hidden debug feature in Espressif's ESP32 modem that bypasses the fixed Wi-Fi modem and dumps raw baseband IQ samples straight into internal SRAM, using the chip's own ADC chain with no extra hardware. They reverse-engineered it, with help from LLMs, and it works across the original ESP32 plus the C3, C5, C6, C61, S2, S3, and S31, capturing up to 80 MSa/s across 2.2-2.7 GHz (and 4.8-6.0 GHz on the C5), which even reaches some LTE and 5G NR bands near 2.6 GHz.
Boston Dynamics has introduced a new four-finger hand for its Atlas robot, optimizing cost and reliability by excluding the pinky. The hand, with 13 degrees of freedom, uses dense tactile sensors and enables complex tasks like in-hand object reorientation and tool operation. Designed for simulation and reinforcement learning, this innovation aims to enhance Atlas' capabilities for industrial applications.
UC Berkeley and FuriosaAI researchers explored using high-bandwidth flash (HBF) to enhance memory capacity for LLM serving. They achieved up to 87% faster completion times and 55.8% energy savings compared to HBM-only systems. Buffered cache-aware scheduling extended HBF write lifetime from 4.77 to 14.82 years, highlighting the impact of data placement strategies on performance and efficiency.
A Nature Electronics paper from Sai Li, Zhang, and colleagues describes an Ising machine built from CMOS-integrated magnetoresistive random-access memory (MRAM) that switches magnetic states in under a nanosecond using electrical pulses. Ising machines map an optimization problem onto a network of coupled magnetic spins and let the system settle toward its lowest-energy state, which encodes the best combination of options, so the fast, controllable magnetic switching here is what does the searching.
Federal prosecutors arrested Greg (Yiu Kong) Lui, owner of San Gabriel Valley company Earthmade, on charges of exporting more than $300 million of servers loaded with export-controlled Nvidia AI GPUs to China between 2023 and 2024. The scheme worked by exploiting a licensing gap: Commerce doesn't require an export license to ship this hardware to Malaysia or Singapore, but does for China, so Lui allegedly sent servers to Malaysian and Singapore transshipment companies using false paperwork, then used freight forwarders to move them on to Chinese buyers.
The physical AI race will depend on owning IP rather than just building the best robot. Robotics companies need a strong IP strategy, focusing on key inventions that ensure their competitiveness. Utilizing AI in patent processes can reduce costs and improve efficiency, crucial for resource-limited startups.
Praxis-1 is a world action model that uses video pretraining to control robots.
Hello Robot secured nearly $3 million from the NIH to advance its Stretch robot for elder care, focusing on cognitive support, social engagement, and safety management.
Legato Frames pack the startup's hearing-assistance tech into the arms of eyewear, aimed at adults with up to moderate hearing loss and starting at $999.
Nvidia is adding a 64GB unified-memory DGX Spark, sold only through partners (Acer, ASUS, Dell, Gigabyte, HP, and MSI) starting October 23 at $4,999, built on the same GB10 Grace Blackwell chip as the 128GB model and able to run models up to 100B parameters on-device.
Dyna-2.1 introduces a semi-humanoid robot named Taku that is designed to autonomously complete long-horizon, complex physical workflows. The system utilizes a dual-model software architecture where a vision-language model monitors the environment while a separate model coordinates whole-body manipulation, allowing the robot to dynamically switch tasks. It highlights a major shift toward utilizing integrated foundation models and highly flexible physical platforms to automate complex, multi-step commercial and industrial service roles.
Kodiak AI is preparing to launch a driverless commercial shipping route for IKEA Supply along a 219-mile section of Interstate 45 in Texas. The milestone follows four years of collaborative testing during which they successfully completed over 1,300 runs to resolve operational complexities like scheduling, predictive maintenance, and dock assignment. It serves as a prime case study of how custom autonomous technology can successfully transition from controlled, simulated environments into the real-world logistics and shipping infrastructure.
SpaceX's Transporter-18 launched three payloads testing technologies for orbital data centers. Cowboy Space's Reason-1 tests beaming solar power via laser, Google's Project Suncatcher runs AI chips on solar in orbit to evaluate durability, and Star Catcher's Protostar aims to validate power-beaming to satellites. These tests are crucial for building infrastructure to support future AI computing and data centers in space.
The new Framework Desktop pairs AMD's Ryzen AI Max PRO 495+ with 192GB of LPDDR5X and 273GB/s of memory bandwidth, and it ships in November as a DIY kit, a Fedora prebuilt, or a bare Mini-ITX board. Its PCIe x4 slot is now open-ended, so full-size x8 or x16 cards fit without a riser, letting owners add 50Gbps networking to cluster several machines for faster inference. Framework expects memory prices to keep rising for six months and tells buyers to pick their capacity now.
PCIe 8.0 traffic can cross chiplet boundaries through UCIe 3.0, but engineers must coordinate both protocols' power, reset, and recovery states. Timing mismatches can cause completion timeouts, dropped traffic, and deadlocks. Siemens' verification tools trace transactions across both layers and inject faults before silicon bring-up.
Humanoid robots need different memory architectures for centralized AI compute and distributed joint control. Flash and EEPROM can delay writes by 5-10 milliseconds, which risks lost position data during power failures. The article maps F-RAM to continuous telemetry, safety-capable NOR Flash to boot storage, and high-speed NOR Flash to time-sensitive networking.
OpenArm 2.0 is an open-source, 7-degree-of-freedom robotic arm engineered by Tokyo-based Enactic for contact-rich manipulation and physical artificial intelligence research. The arm features backdrivable, quasi-direct drive joints for highly compliant physical interactions, a gripper with an integrated camera, and a motorless replica arm for bilateral force-feedback teleoperation. It delivers a highly customizable, low-cost platform that facilitates the training of robust manipulation models using modern imitation learning and simulation-to-reality pipelines.
Orion is a custom-built, 12-degree-of-freedom quadruped robot designed and programmed entirely from scratch. The system combines an NVIDIA Jetson Orin Nano for advanced stereo vision depth perception and LiDAR-based SLAM with an STM32-based microcontroller running FreeRTOS for low-level kinematic gaits and real-time stabilization. It provides a complete, open-source blueprint for constructing complex, dynamically stable legged systems using advanced embedded processors and modern spatial computing toolchains.
This open-source repository provides a ROS 2 hardware interface and differential drive controller designed specifically for the Raspberry Pi Build HAT. The package enables direct control of Lego Spike Prime motors and sensors, allowing developers to run modern robot communication middleware on lightweight educational hardware. It offers an accessible entry point for testing and prototyping mobile bases using popular, modular consumer components.
FedEx ordered 2,000 Harbinger electric delivery trucks for $300 million, with delivery due before the end of 2027. FedEx also led the startup's $160 million Series C last year. Harbinger projects $20,000 in annual savings per truck against diesel, but those figures are estimates rather than measured fleet results.
DeepSeek released six software modules for Huawei Ascend chips, including compute and communication libraries and TileLang support.
Aaeon launched the Up Xtreme PTL Edge Air mini PC with Intel Panther Lake processors and Frore's AirJet cooling.
Starship's 14th test flight reached orbit for the first time on Monday, despite one engine shutting down unexpectedly mid-flight, deploying 26 next-generation Starlink V3 satellites one by one from a Pez-dispenser-style mechanism at roughly 269km altitude before the mission ended early. Each Starship can now carry about 60 Starlink V3 satellites, roughly 20 times the capacity of a Falcon 9 launch, and the new satellites offer around 10x the downlink and 22x the uplink capacity of the previous V2 generation, capability SpaceX needs to support Musk's stated goal of Starlink eventually carrying a majority of the world's internet traffic.
Nvidia's Open Agent Safety Platform pairs OpenShell, a sandboxed runtime running on its Vera CPU, with Sentry, a monitoring layer that runs in silicon on the BlueField-4 DPU rather than on the same host as the agent it's watching. The architectural bet is physical separation as a security property: in a Vera Rubin POD, each compute tray's BlueField-4 sits on the node's only path to the model, meaning every agent "thought" has to pass through that chip, giving it both the best observation point and a hardware kill switch, isolated from the host and beyond the agent's own reach even if the host itself gets compromised.
VisionPower Semiconductor Manufacturing Company, a joint venture between Vanguard International Semiconductor and NXP, opened its first 300mm wafer fab in Tampines, Singapore, completed in just 22 months and now in risk production after its first sample lot cleared yields above 99%. The fab targets mature nodes between 130nm and 40nm for power management, mixed-signal, analog, and interposer chips serving automotive, industrial, consumer, mobile, and HPC customers, with volume manufacturing starting Q1 2027 and full capacity, roughly 44,000 wafers a month and 1,600 jobs, targeted by 2029.
Tesla ramped Optimus production from a few dozen units in Q2 to several hundred a week last month, but the company is hitting real engineering roadblocks on the path to its 20,000-units-a-week target for year-end. Most units built so far aren't shipping to customers at all, instead staying in-house for testing and data collection, and are still confined to simple, repetitive tasks rather than functioning as general-purpose machines. The core problem is durability and reliable assembly of the hands, a struggle Tesla has reportedly wrestled with for over a year, and rather than relying on touch sensors built directly into the hand itself, the company is now developing a hot-swappable "sensing glove" as a workaround.
Sponsored
After nearly a decade of R&D, Doroni unveiled a showroom model of its H1-X flying car. Morgan Stanley projects a $9T market by 2050. First
AMD has agreed to acquire World Labs, a startup specializing in spatial intelligence and deep learning models designed to understand the physical three-dimensional world. This merger aims to build an open, end-to-end AI ecosystem by combining frontier physical intelligence research directly with high-performance silicon development. This acquisition is exceptionally relevant to those building physical machines because it bridges the gap between hardware acceleration and the spatial awareness foundation models required for next-generation robotic manipulation and navigation.
Announced at ROSCon Global 2026 in Toronto, the NODL framework provides a metadata language to formally describe the ROS interface of a robotic node, including its topics, parameters, services, and actions. It allows developer tools to validate configuration layouts, auto-generate boilerplate code, and create automated testing setups directly from declarative node specifications. For anyone designing embedded control software or complex sensor integrations, this tool offers a standardized way to ensure subsystem nodes communicate reliably without manual, error-prone integration checks.
At TSMC's OIP North America symposium, VP Sajiv Dalal now projects the semiconductor industry hitting roughly $1.7 trillion by year-end, with AI alone crossing $1 trillion, dwarfing TSMC's own "aggressive" 2024 forecast of $1 trillion for the whole industry by 2030. TSMC's revenue is up 53.3% year-over-year through August, and new tape-outs have quadrupled across the N5, N3, and N2 nodes, with N2 seeing an especially sharp acceleration in new chip designs.
VAST Data unveiled DataEnclave, a confidential AI runtime built on NVIDIA Confidential Computing, aimed at industries like finance, healthcare, and government where sensitive data can't leave tightly controlled environments and outside AI models normally can't reach it. The system creates hardware-isolated secure enclaves using CPU and GPU trusted execution environments, requiring cryptographic attestation that verifies the environment and its enforced policy before any decryption key is released, so proprietary model weights and sensitive customer data get decrypted only inside a verified enclave and stay encrypted in both CPU and GPU memory while processing, invisible even to the infrastructure operator running the hardware.
Samsung Electro-Mechanics is investing a combined 6.78 trillion won ($4.9 billion) to expand FC-BGA substrate production, 4.27 trillion won for a new plant at its Sejong site, and 2.51 trillion won for its Vietnam subsidiary, marking the company's largest-ever investment in a single product.
CERN's EP department released colibri, a VHDL common library built up over four years after gateware designers found themselves relying on homebrew tools scattered across FPGAs from multiple vendors.
Red Hat Enterprise Linux has officially achieved Tier 1 support status in the Robot Operating System 2 development ecosystem, beginning with the Rolling Ridley release.
This repository contains the training materials, benchmarks, and hands-on exercises for a technical workshop focused on tuning ROS 2 middleware performance.
Qualcomm launched Snapdragon Sound Elite Gen 2, its first audio chip built for AI agents, with low-power Wi-Fi 6E so earbuds can reach cloud services without a phone. It claims up to twice the AI capability of the Snapdragon S7 Gen 1 in a package up to 30% smaller. Cleer showed a camera-equipped earbud concept that could answer questions about what the wearer sees.
Qualcomm and Apple have extended their global patent licensing agreement that was previously set to expire in 2027. The new deal will take effect on April 1. The companies did not disclose how long the renewed agreement will remain in effect. While Apple has not officially commented on its modem plans going forward, it is expected to eventually move away from Qualcomm modems altogether.
Apple has been ordered to pay more than $5.7 billion to Taction Technology, a small haptics company, for infringing on two of its patents. The verdict is the largest patent award in US history. The jury found that Apple's Taptic Engine infringed on patents from Taction, but that Apple did not infringe on purpose. Apple has rejected the verdict and said it would appeal.
Chinese companies have moved first to commercialize glass substrates, building production lines and accelerating product manufacturing and customer qualification. Korea is taking longer to build production lines as companies are trying to validate each process before making investment decisions. China is building the equipment first, making the products, and then having customers evaluate them, so it is faster in terms of commercialization speed. The country has built its own ecosystem spanning display, printed circuit board, and equipment suppliers.
Boston Dynamics has opened a new robotics center at Hyundai Motor Group's Metaplant America. The company began pilot operations at the robotics center in June and recently transitioned to full operations. The center serves as a test bed and training center for integrating Boston Dynamics' Atlas humanoids across Hyundai's automotive factories. The companies hope to have the robots performing component assembly by 2030.
This article looks at the current state of quantum computing, looking at different vendors and their progress.
Satellite navigation systems are surprisingly vulnerable to a variety of attacks, ranging from jamming to more sophisticated spoofing attacks. Europe's Galileo GNSS system aims to fix this. The system produces a cryptographically secured position to prevent spoofing. While it is complicated, it works, and tests show that a stable lock can be achieved on the receiver even during spoofing. The method can, in principle, be extended to other GNSS systems.
A GPU just needs to lose enough of the business its next owner expected to serve for its value assumption to fail. Increasingly capable inference on phones and laptops threatens that business. Edge adoption weakens the case for treating cheap inference as a dependable floor under GPU residual values. Once local execution meets quality, latency, memory, and battery requirements, end users may no longer require data center compute.
Most tend to think of caches strictly in terms of capacity, and that suffices for most purposes. This post looks at chip designers' motivation for a way predictor. It looks at how they work, when the hardware misbehaves, how to detect issues, and workarounds.
OS kernel compilation times have become a signature benchmark over the past two decades.
At the TSMC OIP Conference, SK hynix focused on showcasing its technological capabilities and product lineup while further defining and expanding its collaborative relationships with TSMC and other key partners.