Hey there, fellow future-addicts!
Welcome to this week's edition of Rushing Robotics! If you blinked this week, you missed at least three world-firsts. A full-stack MRI brain-computer interface that reads and writes the brain. The first AI-assisted live brain surgery, which ended with a man keeping his sight. A one-dose gene therapy for inherited heart failure. Also: OpenAI's chip outmuscled Nvidia, IBM taught a mainframe Arm, and humanoid robots took up tennis. Here's everything that mattered.
🤯 Mind-Blowing
From a machine that reads and writes the brain in one closed loop, to an AI second pair of eyes watching over a neurosurgeon's shoulder mid-operation — these are the stories that made us do a double-take this week. Add a gene therapy that restores broken hearts at the cellular level, a custom chip that beats the industry's best on performance per watt, and robots that now come with their own automated training coach, and you've got a week the future won't forget.
🔊 Industry Insights & Updates
The plumbing of the future got serious upgrades this week. IBM's 2-nanometer dual-architecture processor brings Arm's 22-million-developer ecosystem to the mainframe for the first time, a whisper-thin layer of salt pushed perovskite tandem solar cells past 30% efficiency, a nine-seat hybrid aircraft that lands on a soccer field is heading into production, and New York now hosts America's longest wireless quantum link — 13 miles of entanglement-ready light through open air.
🧬 BioTech
Healing is becoming programmable. This week's lineup: a nanoparticle hydrogel that nearly sealed a chronic wound in 72 hours, engineered tRNA that reads through faulty stop signals and could one day treat thousands of genetic diseases with a single strategy, and a pill doctors already prescribe that may halt the spread of one of the deadliest breast cancers.
💡 Products/Tools of the Week
Your workflow just leveled up. Build a cinematic 3D website from a single sentence, plug AI agent harnesses into your product with one API call, get a clear ship-or-block security verdict on AI-built apps, and hand everyday busywork across 1,500+ apps to plain-English AI agents.
🎥 Video Section
Grab the popcorn — the 2026 World Humanoid Robot Games delivered everything this week: an intense 1-on-1 humanoid boxing match, the world's first fully autonomous robot tennis rally, and a glorious highlight reel of fails proving the robot revolution still has a few bugs left to iron out.
Weeks like this are a reminder that the future rarely arrives with fanfare — it compounds quietly, one brain interface, one gene therapy, and one improbable robot tennis match at a time. The tools being built today are collapsing the distance between what we imagine and what we can actually do, and that pace is only accelerating. Stay hungry, stay futurish!
🤯 Mind-Blowing
Reading the brain and writing to it now happen in one machine. China has released uMR Shenguan, described as the world's first full-stack MRI-based brain-computer interface, combining signal acquisition, decoding, neuromodulation, and evaluation in a single closed-loop system. UIH and TJU jointly built the platform, in which the MRI scanner serves as the loop's perception and feedback hub, capturing neural signals at millisecond speed and brain structures at sub-millimeter precision while compensating for interference from implanted devices. The developers say the system makes neural plasticity measurable and controllable, with clinical uses emerging in rehabilitation and psychiatry and consumer applications expected to follow.
Brain surgery has a new second pair of eyes. For the first time, real-time AI assisted a neurosurgeon during a live operation, helping remove a pituitary tumor while sparing the patient's vision. UCL and NHNN surgeons ran the clinical-trial procedure in London using UCL-built software that analyzed the endoscopic video feed as it happened, flagging risky anatomy in a space where a millimeter's error can cause blindness or stroke. Hani Marcus led the operation with Danyal Khan, the surgeon fully in control throughout. Patient Rhys Hibbert, 48, whose tumor had stolen much of his peripheral sight, walked without glasses or sticks within a week. Backed by the NIHR and Google, the UCL team next wants the system to track instruments and monitor tissue interaction in real time.
Genetic heart disease has met its first real challenger. A single-dose gene therapy restored normal heartbeat strength and rhythm in patient-derived heart tissues and reversed disease in mice, according to work published in Nature Cardiovascular Research. MCRI scientists replaced faulty copies of the ALPK3 gene with healthy ones, fixing the internal scaffolding of heart cells; AI-assisted analysis suggests the same strategy could help patients with MYH7 and TTN-related cardiomyopathies, among the most common genetic causes of heart failure. With about 30 million people affected by cardiomyopathy worldwide, the team is preparing the safety studies needed before human trials and seeking commercial partners to get there.
A custom chip just outmuscled the industry's best on efficiency. OpenAI's first in-house inference chip, Jalapeño, delivered 1.5 to 1.9 times more AI work per watt and 1.7 to 3.6 times lower latency than Nvidia's top commercial systems in new benchmark tests. OpenAI measured the chip on the public InferenceX benchmark from SemiAnalysis across three large models — its own GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T — and reported 2.1 to 4.1 times higher performance on highly interactive workloads. Rated at 700 watts but sustaining 550 watts or less in testing, Jalapeño is built to erase the usual tradeoff between throughput and latency: it handles both the compute-heavy prefill phase and the bandwidth-hungry decode phase in one architecture, keeping model state close to its processors and integrating networking to cut data movement between chips. OpenAI designed it with AI's help, moving from initial design to tapeout in nine months, and plans deployment in its compute infrastructure by the end of 2026 as the first of a multigenerational roadmap.
Robots learning to navigate new places now have an automated coach. A new framework called COMPASS reduces the rework normally required when robots, scenes, or conditions change. NVIDIA researchers paired the company’s pretrained X-Mobility policy with residual reinforcement learning, training a specialist that adjusts the base policy for each robot-environment pair rather than starting over. AI coding agents run the pipeline, preparing simulation assets, testing, training, debugging failures, and comparing checkpoints, while humans approve the critical decisions. NVIDIA validated the workflow on the Spot quadruped, measuring goal success, falls, and route time, and says approved policies can generate movement commands directly from camera images, odometry, and a target goal.
🔊 Industry Insights & Updates
Arm is moving into the mainframe. A new IBM processor built on a 2-nanometer process can run Arm-native Linux applications alongside traditional z/OS and IBM workloads on the very same cores, a first for enterprise computing. IBM and Arm gave the chip 11 cores running above 5.7 GHz, each able to execute both instruction sets natively, plus AI inference accelerators for real-time fraud detection and a dedicated on-chip unit for I/O acceleration. The design, the first hardware result of the companies' April collaboration, targets future IBM Z and LinuxONE systems and opens the door to Arm's 22-million-developer ecosystem.
A tiny salt layer just fixed a big solar problem. Cesium chloride, deposited as an ultra-thin seed layer, has pushed perovskite-silicon tandem solar cells to 30.3 percent efficiency while making them far easier to mass-produce. HZB scientists had faced years of manufacturing frustration: on textured silicon, the self-assembling monolayer molecules that transport positive charges collected in the surface valleys and left the peaks exposed, and perovskite grown on top developed defects that sapped power. Nanoscale imaging at the BESSY II synchrotron revealed the failure, and the seed layer corrected it by leveling the microscopic terrain and guiding uniform crystal growth, as lead author Viktor Å korjanc explains.
Regional air travel may soon skip the airport entirely. A nine-passenger hybrid-electric aircraft able to take off and land on a soccer field is heading into production in Springfield, Ohio, through an $850 million investment creating nearly 2,000 jobs. MIT-born startup Electra pairs a nose-mounted gas generator with under-floor batteries and eight electric wing motors, using blown lift for ultra-short takeoffs and quiet operation in noise-sensitive areas. The plane flies 1,200 miles at around 200 mph, targeting the 50-to-250-mile trips where cars and airlines both fall short. Electra's EL2 demonstrator has logged over 200 test flights since 2023; the new factory breaks ground next year, aiming first at 400 aircraft a year, then 800.
America's longest quantum network just went wireless. Quantum states of light — and later entangled photons — traveled 13 miles through open air between rooftops in Stony Brook and Upton, New York, a US first. Brookhaven Lab (BNL) and Stony Brook University (SBU) researchers built the link around a laser source at SBU's Quantum Watchtower, an ultrafast camera at BNL's Quantum Lighthouse, and adaptive optics from astronomy that kept atmospheric turbulence from destroying the signal. The free-space optical connection won't replace fiber; it extends the existing 161-mile, eight-node network to places cables can't reach and to wavelengths fiber doesn't favor. A Yale facility comes next, enabling a 30-mile hop across Long Island Sound, with satellite links as the long-term goal.
🧬 BioTech
A tiny particle is taking on wounds that won't close. In lab tests, a hydrogel loaded with hybrid nanoparticles nearly sealed a simulated chronic wound in just 72 hours. USP researchers designed the particles to deliver two agents at once: siRNA that switches off production of MMP-9, an enzyme that eats away at skin-repair proteins in chronic wounds, and resveratrol, which calms inflammation and spurs collagen-making fibroblasts. The nanoparticle protects both fragile molecules and releases them slowly at the wound site. The FAPESP-funded team says patients with diabetic and pressure ulcers, who face heightened risks of infection and amputation, could eventually benefit, though 3D skin models and animal studies must come first.
One drug strategy could soon serve thousands of genetic diseases. Engineered transfer RNA has read through the premature stop signals known as nonsense mutations, restoring functional protein in cystic fibrosis models where today's drugs fail. U of T researchers paired chemically enhanced tRNA, made longer-lasting with a modification borrowed from nature, with a tailor-made lipid nanoparticle delivery system, publishing the results in Science. The restored CFTR protein kept working for over 40 days in human airway cells, and in patient-derived organoids the tRNA combined with Trikafta succeeded where neither worked alone. Study lead Bowen Li says the goal is tRNA medicines that recognize shared stop signals across many diseases, with an inhalable version already in sight.
A breast cancer killer may be stopped with a pill doctors already prescribe. Restoring levels of a natural molecule called miR-342 sharply slowed the spread of triple-negative breast cancer to the lungs and bones in pre-clinical models, and the existing CDK4/6 inhibitor palbociclib achieved similar results where miR-342 was low. AU and ONJCRI researchers identified the miR-342–E2F network as the switch controlling metastatic growth, reporting in EMBO Molecular Medicine that low miR-342 combined with high E2F activity flags the patients most at risk. That same signature could serve as a test to select triple-negative patients for CDK4/6 inhibitor treatment, repurposing an existing therapy for one of the deadliest breast cancers. Validation in patient-derived models comes next, before clinical trials.
💡Products/tools of the week
A full cinematic 3D website now takes a single sentence. Draftly, an AI-powered 3D website builder, turns one text prompt into a scroll-driven web experience, requiring no coding or design skills from the user. The platform handles the entire pipeline on its own: it generates visuals, animates motion, extracts video frames, and assembles everything into a scroll-reactive 3D experience, then exports production-ready HTML, CSS, and JavaScript. Users who want changes can chat with the AI to adjust copy, colors, and layout live, or switch to a node-based visual studio for deeper customization of the build. Finished projects publish directly to a Draftly subdomain or a custom domain, or download as a ZIP file for self-hosting. A free plan is available, with paid plans starting at $25 per month.
Agent-powered features just became an API call away. HarnessRouter's unified platform lets developers plug AI agent harnesses such as Codex, Claude Code, and Hermes into their products, running each session in an isolated serverless sandbox while it manages provisioning, streaming, retries, and artifact delivery. A benchmarking tool compares harness and model combinations on cost, latency, and success rate, and one API key is all it takes to start shipping.
Shipping an AI-built app now comes with a clear go or no-go signal. AtRisk, a continuous security tool, scans live URLs and connected GitHub repositories for vulnerabilities across 100+ checks, then returns a Ship/Block score indicating whether an app is safe to deploy. Findings from the live site and source code land in a single inbox, paired with paste-ready fix prompts that work with Cursor and Claude Code via MCP. Automatic pull request reviews come built in, and the Pro plan can block risky deployments via a GitHub Action with SARIF output.
Everyday work across an entire app stack can now be handed to AI agents. DeskFerry, an AI automation platform, turns plain-English instructions into executed tasks across more than 1,500 connected apps, offering DeskFerry Assist for quick, repeatable jobs and an Agent Builder for multi-step workflows spanning several tools. Connections include Gmail, Slack, HubSpot, Notion, and Salesforce, with AI models from OpenAI, Anthropic, Google, and others available without separate API keys. Custom skills keep agents on-brand, format outputs, and file results into the right tools automatically.
