The future is arriving badly explained.
The ideas changing the world.
And what they mean for you.
Selected on · Four signals worth your attention.
A coding study trains agents to reject bad fixes.
A newly announced preprint trains coding agents to write tests that distinguish working repairs from deliberately faulty alternatives. On 270 SWE-bench Verified issues excluded from training, the authors report first-attempt resolution rising from 31.9% to 43.0%. The point is not simply to make more attempts, but to improve how an agent checks its work. Those issues had already informed method development, however, so this is not a clean test on wholly unseen problems or proof of dependable production software.
A robot dog adjusts its gait to make less noise.
Researchers describe TACET, a system that changes a quadruped robot's gait and route according to the people and activity around it. Their preprint reports quieter walking on a Unitree Go2 and tests behavior near people who are working, talking, or walking. It treats noise as part of navigating shared space, not just avoiding collisions. The largest reported reduction was measured by a microphone mounted on the robot, not at a listener's ear; selected test scenarios do not establish everyday performance.
PyTorch adds an overlapping compute path for AMD GPUs.
PyTorch's development branch gained a native ROCm implementation that overlaps gathering data from other GPUs with matrix multiplication, rather than waiting for the whole transfer first. The change supports AMD's MI300X and MI355X architectures and remains opt-in. It extends a technique already available on NVIDIA's CUDA path to another hardware stack. This is a code integration, not a stable PyTorch release or a universal speed boost: the author's measurements show that some workload shapes remain slower than the fallback.
Researchers target bad robot habits with local retraining.
A new preprint introduces REDIRECT. Using demonstrations labeled successful or problematic, it locates a troublesome part of a robot task, retrieves a better continuation, and retrains the policy while rehearsing behavior worth keeping. The authors report improvements in simulated manipulation and small physical tests involving cups and towels. The useful idea is selective correction rather than relearning the whole task. The headline '1%' refers to a defined optimization budget in simulation, not a measured 99% saving in total cost or a guarantee that other skills stay intact.
In Charts.
A photo, an assembly manual and tools. The best score among tested models rose from 28% to 80% in Epoch’s furniture test.
Explore the ChartWho Is Gaining Ground in Energy?
Explore the ChartChina Installs More Robots Than the Rest of the World
Explore the ChartLatest
Pick a frontier.
Seven fields. A clear place to begin.
Stay ahead
of the story.
A personal take on the week. By Dorian.
Occasional editions. Free to subscribe.
Unsubscribe anytime.