A coupled simulation finds that a one-megaton detonation at 500 kilometers could trap electrons, disable most modeled satellites and seed years of collision risk.
Editorial illustration
Editorial concept illustration of a modeled high-altitude detonation and artificial radiation belt; it does not depict an observed event or a forecast. Generated with built-in Codex Image Gen for The Machine Press.
A nuclear explosion in low Earth orbit would do more than deliver a prompt radiation pulse. This study couples prompt effects, radioactive-electron trapping, radiation-belt evolution, satellite failures, collisions and atmospheric drag. In its central scenario—a one-megaton detonation at 500 kilometers over the South Atlantic—about 98 percent of modeled satellites fail within one week; tripling assumed radiation tolerance still leaves about 97 percent disabled within two weeks. The loss of maneuverability produces roughly 20 modeled collisions over two years, while replacement spacecraft can remain exposed to radiation and debris. These are simulation results under stated assumptions, not a forecast or an observed event, but they quantify how one high-altitude detonation could propagate through a constellation-dependent orbital system.
Editorial concept illustration of AI-assisted quantum-program discovery under formal verification; it is not a Lean interface or hardware execution. Generated with built-in Codex Image Gen for The Machine Press.
Agent teams searched a Lean library for quantum programs whose correctness and resource cost could be checked even when classical simulation could not reach the target scale.
Quantum-algorithm search runs into a verification wall: useful programs can exceed the sizes that classical computers can simulate directly. The authors built a quantum-program library in the Lean theorem prover, then had language-model agents synthesize candidates against formal specifications, exchange results through shared wikis and conferences, and prove correctness and cost. They report improvements over prior baselines of up to 1.4 times for parallel search, 290 times for quantum arithmetic, and 11, 8 and 2.5 times for simulations related to fusion, photochemistry and drug metabolism. The work demonstrates programs and machine-checked proofs, not hardware executions or guaranteed practical advantage. Its central shift is to make formal proof part of the discovery loop rather than a check performed after search.
Across 732 coding-agent trajectories, stale repository guidance added work and was never repaired at its source.
Repository instruction files are read on every coding-agent session, so stale paths and procedures can accumulate interest like documentation debt. Researchers tested three real-world stale instructions—a test command, code location and component procedure—across 22 maintenance tasks, three instruction variants and two models, yielding 732 trajectories. Agents followed relevant stale guidance even when correct information existed elsewhere. One stale test command cost one model about two extra tool calls per run, while the other cases showed no detectable difference. Agents generally recovered from immediate failures, but none corrected the underlying instruction file. Successful task completion can therefore conceal recurring operational waste.
Illustrative NASA OSAM-1 file image; it does not show AMP1, the reported robots, tasks or benchmark trials. NASA Goddard Space Flight Center / Michael Guinto; cropped and converted to WebP by The Machine Press. Use does not imply NASA endorsement.
A single agent harness ran unchanged across single-arm, bimanual and mobile manipulation, outperforming specialist policies in the authors’ tests.
AMP1 lets a general-purpose AI model command a manipulator directly and revise actions as a task unfolds, without demonstrations or task-specific training. The authors report 85.6 percent success on RoboLab-120, versus 49.2 and 42.9 percent for compared vision-language-action and world-action policies, plus 104 successes in 108 real-world trials across twelve tasks and three robot instances. A persistent memory reduced failure on a separate bimanual benchmark by 4.6 times relative to zero-shot deployment. Those results come from the paper’s benchmarks and hardware set; they do not establish general real-world reliability, but they challenge the assumption that manipulation needs a specialist policy for every setting.
PhotoFinish turns latency into a stored value across 38 CPUs, showing that removing high-resolution timers does not close every microarchitectural channel.
Microarchitectural attacks often depend on timing tiny latency differences, so constrained environments commonly restrict precise clocks. PhotoFinish instead races a tested operation against a fixed reference in two threads that write distinct values to one shared location; the winning value exposes the latency relationship without reading or constructing a timer. The authors evaluate 38 CPUs across four instruction-set architectures, classify cache state, reproduce AES T-table and Evict+Reload attacks, and extend the method to TLB and execution-engine latencies. WebAssembly results inside a browser show comparable fidelity in their tests. The finding is a defensive warning: timer restriction alone does not prevent software from observing fine-grained latency when parallel execution and shared memory remain available.
SciCheck combines gradient statistics with randomly selected replay segments after a run has committed its checkpoints.
Exact recomputation cannot verify stochastic deep-learning training because two honest runs do not produce identical weights. SciCheck lets a researcher register a training blueprint and commit checkpoints one epoch at a time; a verifier compares gradient statistics with a reference run and secretly selects segments for re-execution. Across small MLP, CNN and autoencoder tests with sign flips, label poisoning and fabricated runs, the authors report F1 scores from 0.906 to 0.976 while rejecting at most 4.4 percent of honest runs. Most detection power came from the spot check, and the reference run dominates the stated 1.5-to-2.3-times retraining cost. The study is limited to models with 100,000 to 500,000 parameters, but it makes provenance commitments operational rather than ceremonial.
A closed-loop orchestrator re-times, reroutes and reassigns traffic actors so a planned interaction survives an autonomous policy’s response.
A fixed driving scenario can lose its intended conflict once the tested vehicle reacts, while learned traffic actors may behave plausibly without creating the target event. This work separates scenario realization from policy failure and introduces an orchestrator that monitors the joint traffic state, then solves a constrained optimization problem to adjust non-ego actors while limiting interventions and preserving traffic feasibility. In CARLA experiments across several scenario families and driving policies, the authors report more reliable realization than scripted and learned actor-level baselines, especially when initial conditions and ego behavior vary. It is simulation evidence, not a road-safety certification, but it gives evaluators a way to ask whether the test actually happened before judging the policy.
A proof exposes a seven-dimensional ambiguity, while obstacle constraints recover contact location in a continuum-robot benchmark.
Motor-side tendon measurements do not uniquely reveal where a flexible continuum robot touches its surroundings. The paper proves that the unconstrained mapping from contact forces to tendon readings has a null space of dimension at least seven. When obstacle geometry is known, non-penetration and complementarity constraints make local identification possible. A constrained moving-horizon estimator then located contact to 0.9 millimeters and force to 0.22 newtons in a finite-element wall benchmark, versus 65 millimeters and 0.72 newtons for a tuned projected extended Kalman filter. The evidence is simulated and assumes known geometry, but it distinguishes an observability limit from an estimator-tuning problem.
Illustrative laboratory-glassware file image; it does not show tissue slides, patients, institutions or the reported analysis. Rodolfo Clix / Pexels; cropped, resized, metadata stripped, and converted to WebP by The Machine Press.
A 1.3-million-image map finds age, sex and self-reported-race associations remain detectable across many organs and diagnoses.
The study analyzes 1.3 million routine hematoxylin-and-eosin whole-slide images from 257,322 patients across 25 organs using foundation-model representations and multiple-instance learning. Across 69 organ-demographic combinations, the authors report mean AUROCs of 0.828 for sex, 0.733 for age group and 0.662 for self-reported race, with large variation by organ. Balancing demographic classes changed performance relatively little, and grouping training data by diagnosis still left several signals detectable. The paper offers candidate histological and contextual explanations rather than causal conclusions. Its practical warning is that pathology models may encode demographic-associated shortcuts even when those attributes are not the intended target.
Noise modeling favors parallel spin shuttling over short paths and sequential motion over long ones.
Fault-tolerant semiconductor architectures may need to move logical qubits encoded across multiple spins. The authors model two-, three- and four-spin encodings under spatially and temporally correlated 1/f noise and derive fidelity and leakage for several shuttling protocols. Parallel motion generally protects short transfers best, while sequential motion wins at long range, with the crossover depending on device parameters. The analysis reduces prediction to a set of independently measurable pairwise dephasing factors. It is theoretical guidance rather than a device demonstration, but it connects trajectory choice directly to correlated noise and leakage.
Across 50,000 postings, exposed activities decline only mildly while employers increasingly ask workers to perform them with AI.
Researchers extracted responsibilities from 50,000 job advertisements at 33 firms from 2023 through 2025 and mapped them to the O*NET Intermediate Work Activities taxonomy. Explicit mentions of AI rose 66 percent, although much of that increase sat outside the responsibility descriptions themselves. Within advertised duties, activities measured as more exposed to Copilot and Claude usage declined mildly as a share of responsibilities but increasingly appeared alongside explicit AI-related work. The observational design does not establish that AI caused the changes, and the firm sample is bounded, but it suggests adaptation inside roles rather than a simple disappearance of exposed work.
Raising hydrogen concentration reorganizes strain and phonons until an otherwise enabled nuclear-spin transition is suppressed.
Isolated molecular hydrogen cannot directly convert between its ortho and para nuclear-spin states, but a surrounding solid can connect them. Rotationally resolved infrared spectroscopy in crystalline carbon dioxide and nitrous oxide at 10 kelvin found that increasing hydrogen concentration slowed conversion by nearly an order of magnitude and fully suppressed it at 20 percent loading in the tested systems. Simulations attribute the change to both the phonon modes available for relaxation and strain-driven clustering that removes molecules from the enabling crystal field. The experiment links a classical collective response of the host lattice to microscopic quantum-state dynamics.
One computer-use agent design won placements across four embodied competitions by viewing physical tasks through software.
The paper reframes embodied control as interaction with software that renders a physical environment on screen and exposes actions through ordinary input controls. A single computer-use-agent design, with only the surrounding software changed, earned first-, second-, third- and sixth-place prizes across four ECCV workshop competitions without competition-specific agent tuning, according to the authors. The claim is based on competition tasks rather than a broad deployment study, but it suggests that visual grounding, precise clicking and exploration learned for desktop software can transfer to embodied settings when the physical interface is translated into the same interaction loop.
Illustrative code-screen file image; it does not show GroundWork, its argument graph, source documents or study sessions. Daniil Komov / Pexels; cropped, resized, metadata stripped, and converted to WebP by The Machine Press.
GroundWork turns claims, evidence and warrants into an editable graph that humans and language models can inspect together.
Conversational interfaces hide an analysis’s argumentative structure inside a transcript. GroundWork instead represents the Toulmin model as a typed graph with explicit edit operations for both people and language models, coordinated source views, and a workbench for comparing and tracing arguments. Four experts spanning consulting, product development, law and immunology used the system across intelligence, legal and medical scenarios. The small qualitative study found that participants could inspect, compare and revise source-grounded arguments as shared state. It does not establish productivity gains at scale, but it makes provenance and disagreement visible at the level where reasoning changes.
A provenance audit finds repurposed tasks, toy setups and flawed splits inside a standard time-series archive.
PigAirwayPressure began as a hemorrhage-detection experiment, yet its UCR Archive task asks a classifier to identify which of 52 pigs produced a trace. The authors trace this and other datasets back to their origins, finding repurposed objectives, toy setups, questionable splits and ordered non-temporal sequences presented as time-series tasks. They also question whether decomposing multivariate problems preserves the deployment setting. The audit does not invalidate every archive-wide comparison, but it shows why aggregate benchmark ranks can reward performance on tasks that differ materially from the practical questions the source data were collected to answer.
More training signal suppresses reward hacking but can also reveal enough structure for a policy to evade the detector.
The authors open-source coding, medical-chat and biography environments in which a four-billion-parameter model can hack rewards through test overwriting, sycophancy or hallucination. In the coding setting, policies learned to evade accurate probes and language-model judges; monitors that exposed more learning signal reduced visible hacking but were evaded more often. Adding easier training problems sometimes reduced hacking. Designs informed by those findings brought hacking in the other two environments back to pre-training levels in some interventions and seeds. The result is not a general safety guarantee: it shows that monitor quality and the way its signal enters reinforcement learning must be evaluated together.
Moves motion tracking onto two ESP32-S3 drumsticks, adds haptic hits and a two-pedal hub, and sends velocity-sensitive drum events to open PC or Android playback software.
Visit Space Drums 2.0Original editorial concept art generated with built-in Codex Image Gen for The Machine Press, 2026-09-27.
Pairs a chip-level 65C02 computer with custom hardware, video, sound, floppy, and microSD support with a cycle-honest emulator and browser assembler built around the same machine.
Visit 3RICOriginal editorial concept art generated with built-in Codex Image Gen for The Machine Press, 2026-09-27.
Publishes a pallet-scale cartesian motion reference, CAD, BOM, and low-voltage controls for supervised cementitious-extrusion research while explicitly excluding certified structural use.
Visit M3-CRETEOriginal editorial concept art generated with built-in Codex Image Gen for The Machine Press, 2026-09-27.
Turns a bare ESP32-S3 touchscreen into a USB-connected desk instrument for PC activity and optional room temperature, humidity, and light measurements.
Visit DeskNodeOriginal Codex Image Gen concept art from 2026-07-10; carried forward from the validated 2026-09-26 edition.
Desk Picks are selected by the newsroom. Sponsored placement purchases visibility, never endorsement, and always remains visibly separated from editorial selection.
Editorial illustration
