21 posts
HuggingFace and AWS launched a searchable database showing which popular models run unmodified on Trainium2—plus the exact scripts and pre-compiled graphs to run them yourself.
OpenAI scaled storage for 1B+ ChatGPT users by going full circle: multi-tenant → single-tenant containers → back to multi-tenant. The Azure Cosmos DB tweaks that made it work are fascinating.
Hugging Face turned its dataset hub into a universal RL environment registry. No new infra, no sign-ups—just tags. Harbor, Verifiers, OpenEnv, and NeMo Gym environments all land in one filter.
HF Jobs just turned into the fastest way to stand up a private LLM server. One CLI command gives you an OpenAI-compatible endpoint on H200s, pay-per-second, no Kubernetes required.
HuggingFace shows how to split RL training and inference across separate Jobs with no shared NCCL—just a rank-1 adapter, a FUSE-mounted bucket, and smart routing. Training time drops from 3h27m to 53m.
OpenAI just published how they turned Habitat from a client library into a Python service serving 1B+ users. Python at massive scale? Here's how they made it work.
Dharma-AI built a constraint-aware GPU allocator and benchmarked it against FIFO scheduling. On identical hardware running identical workloads, utilization jumped as much as 33 points. All that changed was order.
Hugging Face Storage Buckets just made the robot data flywheel real. Record demonstrations, train by streaming from the Hub, deploy checkpoints—all without re-downloading gigabytes.
Airlines learned that utilization, not fleet size, decides who wins. Enterprise AI is learning the same lesson—with GPUs. A deep dive into why orchestration is the new frontier.
Ai2 ships a continent-scale geospatial inference platform that processes dozens of terabytes in ~24 hours. The infrastructure choices reveal hard truths about production ML.
NVIDIA, DeepMind, and Disney just open-sourced Newton—a GPU-accelerated physics engine that treats simulation as infrastructure. If you're building physical AI, your bottleneck just shifted.
Amazon and Hugging Face killed the context-switch tax. Deep-link from model discovery straight into a pre-configured Studio environment—no IAM wrangling, no quota hunting, just click and ship.
Photoroom pulls back the curtain on the unglamorous but critical data pipeline behind PRX: JPEG quality experiments, Lance vs MDS tradeoffs, and why pre-training is for breadth, not taste.
Hugging Face and EvalEval just patched the biggest hole in AI benchmarking: scattered, incompatible eval results. Now the same score shows up on model cards *and* links to full reproducibility data.
AWS open-sourced Strands Robots, an SDK that exposes LeRobot's stack as composable agent tools. Record demos in sim, push to Hub, run policies, deploy to hardware—all in one agent.
Google is pouring another $1.5 billion into rural Alabama data centers. The community grants are nice PR, but the real story is where hyperscalers build—and why cheap power beats talent.
Hugging Face, Meta PyTorch, Nvidia, and a dozen others just formed a committee to govern OpenEnv—the protocol layer trying to make agentic RL training actually interoperable.
Google just announced community investments in Missouri targeting workforce development and energy programs. Reading between the lines: they're prepping the ground for data-center expansion.
Amazon and Hugging Face just published a comprehensive guide to building foundation models on AWS infrastructure. It's the playbook we've all been waiting for.
Google just dropped a great explainer on TPUs. Here's what makes their custom silicon tick, why matrix multiplication matters, and how they stack up against GPUs.
Google just announced TPU v8, but instead of one chip, they're shipping two: v8T for training and v8I for inference. Here's why the bifurcation matters for AI's next phase.