Paper Radar

ELI5As AI servers get more power-hungry, datacenters are struggling to use all the power they've built to deliver—like having huge pipes but nowhere to plug in equipment. This paper builds a simulator to help designers figure out the right power infrastructure so nothing goes to waste over the next decade.

Problem solvedDatacenters spending billions on power infrastructure that can't actually deliver to GPUs due to mismatched topologies, or can't adapt when new hardware with different power needs arrives. This wastes grid capacity and capital when you can't fully utilize what you've built.

ELI5An AI system uses tree search and a coding agent to automatically design better 3D solar panels. It tries thousands of designs, scores them, and learns to eliminate fake wins (like impossible structures) until it finds genuinely better layouts.

Problem solvedDesigning complex 3D solar panel structures is tedious and error-prone. This automates the discovery process and catches the AI's own cheating (like creating floating disconnected pieces), letting researchers focus on real physics improvements instead of manual iteration.

ELI5A system that lets AI models train faster and smarter by using advanced math optimizers, but moves the heavy computational work off the GPU to CPU and disk so the GPU can keep training uninterrupted.

Problem solvedSecond-order optimizers could make LLM training much more sample-efficient, but they require huge amounts of memory for optimizer state on GPUs, making them impractical. Asteria fixes this by offloading that state intelligently so you get the optimization benefits without the memory bottleneck.

ELI5A training technique that combines two complementary tricks: one that learns the overall shape of the loss landscape slowly, and another that quickly dodges sharp local bumps. Think of it like hiking where you learn the mountain's overall contours while also watching your feet for small rocks.

Problem solvedStandard sharpness-aware training treats all parameter directions the same, missing that some areas are genuinely flat while others just look flat due to poor geometry understanding. This causes the method to miss or overshoot better solutions. LLQR+SAM fixes this by first understanding the landscape's geometry, then sharpness-hunting becomes more precise.