Meta gives Muse DIY HW 🔌, Musk confirms TSMC talks 🤝, Boston Dynamics' 4-finger hand ✋

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 ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌  ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ 

Sign Up |Advertise|View Online TLDR TLDR Hardware 2026-10-05 📱 Headlines and Launches Meta opens its Muse agent to DIY hardware with an open-source gadget SDK (3 minute read) 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. Musk confirms TSMC talks on Terafab, which could sideline Intel's 14A node (3 minute read) 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. 🔨 Engineering and Applications Researchers Find an Undocumented Raw IQ Capture Mode in ESP32 Chips, Turning $5 Boards Into Software-Defined Radios (4 minute read) 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 unveils new four-finger hand for Atlas humanoid robot (4 minute read) 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. 🧠 Research and Prototypes HBF for High-Throughput LLM Serving (UC Berkeley, FuriosaAI) (2 minute read) 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. Sub-nanosecond MRAM Ising machine tackles chip-design optimization problems (3 minute read) 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. 🎁 Miscellaneous California man charged with smuggling $300M of Nvidia AI servers to China through Malaysia and Singapore (3 minute read) 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 be won in the patent office (8 minute read) 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. ⚡ Quick Links Runway introduces Praxis-1 world action model for robotics (4 minute read) Praxis-1 is a world action model that uses video pretraining to control robots. Hello Robot awarded $3 million to develop Stretch robot for elder care (5 minute read) 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 launches $999 AI hearing glasses that amplify voices and cancel noise on-device (3 minute read) 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 launches a $4,999 DGX Spark with 64GB of memory that clusters with a second unit to run 200B-parameter models (3 minute read) 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. Want to advertise in TLDR? 📰 If your company is interested in reaching an audience of hardware professionals and decision makers, you may want to advertise with us. Want to work at TLDR? 💼 Apply here, create your own role or send a friend's resume to jobs@tldr.tech and get $1k if we hire them! TLDR is one of Inc.'s Best Bootstrapped businesses of 2025. If you have any comments or feedback, just respond to this email! Thanks for reading, Kshitij Sudan & Adi Nikumbh Manage your subscriptions to our other newsletters on tech, startups, and programming. Or if TLDR Hardware isn't for you, please unsubscribe.