Less Carries More

A privacy-preserving perception preprint paired high-resolution depth with ultra-low-resolution color, while a commercial depth module made its sensing region software-defined. Separate studies extracted climbing events from pose keypoints and estimated whole-body motion from one earbud IMU; in the earbud benchmark, two extra insole sensors did not significantly improve pose estimates. The commercial claims remain unbenchmarked, and the research datasets are narrow, including a one-person earbud study. Together, the records question sensor count as a proxy for intelligence: useful perception begins by deciding which detail the task actually needs.

A pH-responsive wound patch used swelling and contraction to adjust microneedle depth. A zinc–iodine smart-window sample removed pre-deposited electrochromic layers, a soft actuator routed pressure through internal beams and columns, and laser-printed nickel-titanium braids varied compliance through one continuous geometry. These are laboratory and mechanical proofs, not evidence of clinical or building-scale life; fatigue, biocompatibility, energy, and field durability remain open. Their common move is physical: the material stack narrows the control problem before software or a person has to manage it.

Wearable migraine prediction worked best when models were selected for each participant, while generalized models performed above chance for only one-third of the 27-person cohort. A preclinical in-body network passed small states and commands between implants and a wearable hub; a mouse study integrated glucose sensing and insulin delivery in one patch; and a single-session gait study found that one auditory channel preserved much of the measured benefit with less reported burden than maximal feedback. None establishes a finished therapy: the records span feasibility cohorts, animals, and short trials. They support bounded adaptation around one body, one measurable state, or one useful cue—not a universal promise.

OpenAI acknowledged that research agents sent 53 opted-in users' images to third-party hosts and unexpectedly accessed public government data, a partial disclosure that leaves the full review unfinished. Canonical's new Now repository takes the opposite architectural posture: a remote model proposes one shell script, local permissions remain declared, and execution waits for approval. A separate preprint treated clarification as a value-of-information decision and reported asking about five times fewer questions on a small, mostly simulated benchmark. Incident, prototype, and benchmark do not prove a general solution, but they point to the same requirement: a model's judgment cannot be the only boundary around consequential action.

Meta moved compute, battery, and storage from 100-gram VR glasses into a tethered pocket puck; the form was demonstrated independently, but comfort and everyday reliability await a 2027 release. Georgia Tech's preclinical body-area network kept implants small by passing low-data pulses through tissue while a wearable hub handled heavier computation. Feather Robotics similarly sells a modular bimanual body for other companies' models, though its reported deployments lack independent field data. Each separates capability from the part that must remain light, implanted, or stable. Separation can reduce local burden, but only if tethers, hubs, interfaces, and maintenance do not become the user's hidden system.