Helion says its Polaris prototype became the first privately developed fusion machine to run on deuterium-tritium fuel and the first to reach plasma temperatures of 150 million degrees Celsius. The Washington company reported both firsts on 13 February 2026.
NASA and Katalyst Space Technologies have called off the robotic mission to boost the Neil Gehrels Swift Observatory. Attitude control failures left the rescue spacecraft spinning, and it can no longer capture the telescope.
A benchmark run by startup Discovered Materials had seven frontier models hunt for thermally conductive dielectrics for 3D-stacked chips. They found more than 500 previously unknown materials, but just one came with a synthesis recipe a human expert would attempt.
Helion's Polaris prototype has become the first privately developed fusion machine to demonstrate measurable deuterium-tritium fusion and to reach plasma temperatures of 150 million degrees Celsius, the company said on 13 February 2026.
Three papers show where medical AI models already deliver results and where caution is needed: remote monitoring of heart failure, decoding speech from EEG, and calibrating black-box radiology systems.
Four independent papers from recent weeks show that tables of language model scores look more decisive than the data allows. The problem also touches research done with the help of AI.
Four papers from September expose weak spots in language models. A short, natural-looking piece of context can flip a correct decision, and the gap between what a model encodes and what it actually uses can be very wide.
One paper shows how combining prediction with internal feedback speeds up motor learning by an order of magnitude. The other proposes a way to alter brain representations without physical stimulation.
Initial measurements around data centers show air temperatures rising by 0.7 to 0.9 degrees Celsius within 90 meters to half a kilometer. Simulations predict larger effects at higher power levels.
Three papers show different machine learning strategies in Earth sciences: an unstructured grid for the ocean, generative corrections for threshold scenarios, and a physics-neural model for cyclones.
Two new papers show how AI is entering astronomy: one builds a multi-agent research system for blazars, the other isolates the photon ring from interferometric images.
In a paper signed by co-founder Liang Wenfeng, DeepSeek details DSec: 5,000 new sandboxes per second, 3 million daily, and 380,000 running concurrently.
Fine-tuning Boltz-2's affinity heads on a small set of activity labels improved early enrichment, while limiting rescoring to just ten percent of candidates.
The team behind JEPA-Anything showed that a single predictive architecture can learn the dynamics of cells, molecules, fluids, and planets without being told the laws of physics.
Twenty-five Fields medallists have publicly objected to treating mathematics as a training ground for models. The mathematician Stéphane Mallat adds a reminder: a proof nobody understands is worth very little.
Lagrangian particle tracking in a Mars atmosphere model shows the edge of the Hadley circulation acts as a dynamic wall. Above it, a fast connection links the two poles.
Forschungszentrum Jülich has switched on JION, a gate-based quantum computer that holds ytterbium ions in electromagnetic traps. A few qubits work today; ten should be ready by mid-2027.