November 2022 Summaries

Recent AI developments include Meta AI’s Galactica, a large language model trained on scientific materials to summarize research, solve mathematical problems, and generate reference-style content, alongside research linking LLM factual accuracy to the frequency of relevant documents in pretraining data. Carnegie Mellon and UC Berkeley introduced a camera-based quadruped robot locomotion system designed to traverse complex terrain such as stairs, curbs, gaps, and stepping stones. Industry updates include Sony’s patent for AI-generated gameplay input captions, Ubisoft and Riot Games’ collaboration on machine learning systems for detecting harmful in-game chat, and a Microsoft–NVIDIA partnership to build an Azure-based AI supercomputer. Open-source and product highlights feature GFPGAN for face-photo restoration and TheGist for AI-generated Slack summaries, while a class-action lawsuit alleges that GitHub Copilot, OpenAI, and Microsoft used copyrighted open-source code without attribution. The update also notes Lightning AI documentation improvements, community code contributions, hiring opportunities, and upcoming research conference deadlines and events.

Muse is an open-source text-to-image generator built on Stable Diffusion that also serves as a reference architecture for deploying diffusion-model applications with Lightning. The post explains how the project combines a React interface, a FastAPI model server, and a Lightning Work component that downloads and loads the model, accepts prompts, and returns base64-encoded images, with options for faster 25-step generation or higher-quality 50-step generation. After the initial deployment reached capacity at roughly 10 concurrent users, the team introduced a FastAPI round-robin load balancer, autoscaling based on request volume, and dynamic batching that queues and aggregates requests before sending them to model servers. These changes reportedly increased simultaneous request capacity from fewer than five to more than 300. Muse can also be integrated into Slack through a slash-command bot that generates and posts images from prompts. The project is presented as a reusable, clonable template demonstrating how Lightning can manage cloud resources, GPUs, storage, parallel workloads, and deployment infrastructure for scalable AI applications.

Recent machine learning developments include AlphaPose for real-time whole-body pose estimation and tracking, Microsoft’s Waveformer for streaming target sound extraction, and Sony’s self-supervised system for transferring mixing styles between music tracks. Engineering announcements feature a GPT-3 and DALL-E-powered text-to-Figma generator, NVIDIA Omniverse’s beta release for creating metaverse applications and synthetic data, and Colossal-AI’s open-source Stable Diffusion tooling, which claims substantial reductions in pretraining and fine-tuning costs. Snowflake and Anaconda also released Snowpark for Python to enable data science and machine learning workflows within Snowflake. Community projects highlighted include tutorials for predicting football expected goals, self-supervised monocular depth estimation, and reducing Transformer model size or inference latency, while Lightning AI announced a redesigned website, NeurIPS career-transition event, and an introductory neural-network notebook. The update also notes upcoming IROS, NeurIPS, PyTorch, and CVPR events, including the November 11, 2022 CVPR paper submission deadline.

Meta AI released an open atlas containing structures for 617 million metagenomic proteins, along with models, research, and an API intended to support research into evolution, medicine, environmental remediation, and clean energy. Other research highlights include a proposed machine-learning algorithm for forecasting properties of unknown quantum processes and a study arguing that large language models do not reliably achieve human-like contextual implicature. Engineering news covers a U.S. Space Force project to test AI-based satellite-failure forecasting, Pinecone’s hybrid vector index for keyword-aware semantic search, and Kahun’s medical-literature-based clinical reasoning engine. Open-source updates include Anaconda’s global data-science survey, DALL·E API access for application developers, and Sigstore’s generally available signing service, while the issue also spotlights synthetic-data generation with NVIDIA Omniverse, community projects built with Lightning, Lightning 1.8 features, and upcoming ML conferences including IROS, NeurIPS, PyTorch Conference, and CVPR 2023.

Lightning 1.8, built through more than 550 commits from 52 contributors, introduces features aimed at improving large-scale model training, Lightning App development, and cluster usability. Its experimental Colossal-AI strategy supports multiple parallelism techniques and can reduce GPU requirements for models such as GPT-3 or enable training larger models with the same hardware, while FSDP enhancements add automatic layer wrapping for models that fit in CPU but not GPU memory. Lightning Apps gain encrypted secrets for securely providing credentials, as well as experimental configurable CLI commands that let users interact with running applications from a terminal. The release also adds customizable batch-size and learning-rate finder callbacks for tuning workflows, enables LightningCLI to be invoked directly from Python, and improves SLURM support through configurable preemption signals and automatic requeuing for array jobs.