PinterestTracking Down Mysterious ML Training Stalls
An in-depth investigation into PyTorch upgrade-induced ML training stalls reveals GPU kernel inefficiencies and a Ray monitoring process as root causes, with solutions leading to significant throughput gains.
MIT AINew software designs eco-friendly clothing that can reassemble into new items
Refashion is a modular software tool developed by MIT and Adobe that enables eco-friendly, reusable clothing designs by allowing users to create adaptable garments that can be visually planned, customized, and reassembled into new fashion items.
MetaScaling LLM Inference: Innovations in Tensor Parallelism, Context Parallelism, and Expert Parallelism
Advanced tensor, context, and expert parallelism techniques at Meta optimize large language model inference by enhancing throughput, reducing latency, and improving resource efficiency for scalable AI applications.
Modular AIModular: Achieving State-of-the-Art Performance on AMD MI355 — in Just 14 Days
Demonstrating how Modular achieved state-of-the-art AI inference performance on AMD Instinct MI355 GPU in just 14 days by leveraging an architecture-agnostic, portable software stack and targeted kernel optimizations.
Snorkel AIScaling trust: rubrics in Snorkel’s quality process
An in-depth look at Snorkel AI's rubric-driven quality process that combines human expertise and LLM evaluators to scale trustworthy, high-quality AI data pipelines efficiently.
MIT AIMethod teaches generative AI models to locate personalized objects
A novel training method enhances vision-language models to accurately locate personalized objects using context-driven video-tracking data and pseudo-naming techniques without compromising general model capabilities.
Meta10X Backbone: How Meta Is Scaling Backbone Connectivity for AI
Meta details its 10X scaling of the Express Backbone network through DC metro architecture, IP platform scaling, and IP/optical integration to meet growing AI workload demands and enable large-scale AI cluster connectivity.
MetaBranching in a Sapling Monorepo
Sapling introduces directory branching to overcome scalability and developer experience challenges in Meta's monorepo by enabling mergeable, linear commit workflows at the directory level.
3 tips to improve your opening moves in chess
Enhance your chess skills by mastering piece development, controlling the center, and protecting your king through strategic castling in the opening phase.
OpenAIPlex Coffee delivers fast service and personal connections with ChatGPT Business
Plex Coffee leverages ChatGPT Business to automate knowledge sharing, streamline onboarding, and enhance operations, enabling faster service and stronger personal connections while scaling their café chain.
MIT AIBlending neuroscience, AI, and music to create mental health innovations
Kimaya Lecamwasam integrates neuroscience, AI, and music to develop innovative, non-pharmacological mental health interventions through emotional resonance and affective computing.
MIT AIRemembering Professor Emerita Jeanne Shapiro Bamberger, a pioneer in music education
Remembering Professor Emerita Jeanne Shapiro Bamberger's pioneering contributions to music education through technology, AI, and computer languages, along with her lasting impact as a mentor and performer.