World Wire

ByteDance Seed and Tsinghua AIR have released CUDA Agent, an agentic reinforcement learning system that trains a large language model to write GPU kernels that beat a compiler. The gap it targets is narrow but stubborn: frontier models already produce correct CUDA, they just produce slow CUDA. On KernelBench, the base model Seed1.6 passes 74.0% of tasks yet outruns torch.compile on only 27.2% of them, at a 0.69× geometric-mean speedup which means its kernels are, on average, slower than what the compiler generates on its own. CUDA Agent closes that gap by putting the model inside a real CUDA development environment with profiling, correctness checks and a permission-locked sandbox, then training it with PPO for 150 steps at a 131,072-token context. The result is a 98.8% pass rate and a 96.8% faster-than- torch.compile rate across the 250-task benchmark, at 2.11× geomean over compile — roughly 40 points ahead of Claude Opus 4.5 and Gemini 3 Pro on the hardest Level-3 split. Is i...

In this tutorial, we develop an end-to-end OCR workflow with docTR and explore how modern document understanding pipelines combine text detection, recognition, geometry, layout analysis, structured extraction, and export. We generate realistic synthetic invoice documents, load images and PDFs through DocumentFile, construct GPU-aware OCR predictors, and benchmark different detection–recognition architecture combinations for speed and accuracy. We then inspect the internal Document hierarchy, visualize confidence-aware bounding boxes, use standalone detection and recognition models, implement two-pass recognition for low-confidence words, tune detection thresholds, and introduce custom pipeline hooks for box filtering and padding. We also handle rotated and skewed documents, experiment with layout detection and KIE, reconstruct reading order and tabular information, extract structured invoice fields, and export results as text, JSON, hOCR, synthesized document images, and searchable ...

DeepSeek released DeepSeek Harness v0.1 in developer preview and published the full source code under the MIT license. The project ships as dsh at deepseek-ai/deepseek-harness . A harness is the layer between a model and the environment it acts in — the tools, files, sandboxes, and control loop that let an agent keep working. DeepSeek frames it as Agent = Model + Harness . Most harnesses hard-code that layer: the agent loop, the tool registry, and the session store are fixed, and extension happens only at whatever hooks the authors exposed. Harness takes the opposite position, stated in the first lines of its README: everything is a plugin. Models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI all sit behind Cordis plugin boundaries, and any of them can be selected, swapped, or extended in configuration without changing Harness source code. That makes this a kit for assembling agent runtimes rather than a fixed coding assistant, and it is why the releas...